Category Archives: Machine Learning
Speaking without vocal folds using a machine-learning-assisted wearable sensing-actuation system – Nature.com
Design of the wearable sensing-actuation system
A thin, flexible, and adhesive wearable sensing-actuation system was attached to the throat surface, as shown in Fig.1a, for speaking without vocal folds. This system comprises two symmetrical components: a sensing component (located at the bottom part of the device) converting the biomechanical muscle activities into high-fidelity electrical signals and an actuation component using the electrical signals to produce sound (located at the upper part of the device), as shown in Fig.1b. Both components consist of a polydimethylsiloxane (PDMS) layer (~200m thick) and a magnetic induction (MI) layer made of serpentine copper coil (with 20 turns and a diameter of ~67m). The serpentine configuration of the coil ensures the flexibility of the device while maintaining its performance, as discussed in Supplementary Note1. The symmetrical design of the device enhances its user-friendliness. The middle layer of the device is the shared magnetomechanical coupling (MC) layer, made of magnetoelastic materials consisting of mixed PDMS and micromagnets. The MC layer, with a thickness of approximately 1mm, is fabricated with a kirigami structure to enhance the devices sensitivity and stretchability (see Fig.S1). The entire system is small, thin (~1.35cm3, with a width and length of ~30mm and a thickness of ~1.5mm), and lightweight (~7.2268g) (see Fig.S2 and Supplementary TableS1).
a Illustration of the wearable sensing and phonation system attached to the throat. b Explosion diagram exhibiting each layer of the device design. c Two modes of muscle movement, expansion induces the elongation in the x- and y-axis, while contraction induces the elongation in the z-axis. Kirigami-structured device response to muscle movement patterns in the x, y (d), and z direction (e): expansion results in x- and y-axis expansion and less deformation in the z-axis, contraction results in less deformation in x and y direction and expansion of the z-axis. f Detailed illustration of the magnetic field change caused by magnetic particles. For one part, the angle change between each single unit of the kirigami structure is represented by . For the other part, the magnetic particle itself undergoes torque caused by the deformation applied onto the polymer (g), thus, generating a change of magnetic flux and, subsequently, current in the coil. The photo of the device in muscle expansion state is shown in h (x-, y-axis), i (z-axis), and in muscle contraction state is shown in j (x-, y-axis), k (z-axis). Scale bars, 1cm.
Multidirectional movement of laryngeal muscles sets the significance of capturing laryngeal muscle movement signals in a three-dimensional manner. Moreover, the learning process of phonation may be heterogeneous across populations: different people may adopt a variety of muscle patterns to achieve identical vocal movements45,53. Such complexity of muscle movement requires the device to be able to capture the deformation of muscles not horizontally or vertically alone, but rather in a three-dimensional way. Figure 1c illustrates the movement of the muscle fiber during two stages, i.e., expansion and contraction. During the expansion phase, the muscle relaxes and elongates in the x- and y-axis. On the other hand, during the contraction phase, the muscle shortens in the x- and y-axis while thickening in the z-axis through the increase in muscle fiber bundle diameters. Figure 1d, e demonstrates the devices response in the x-, y-axis, and z-axis, respectively. During the expansion phase, the kirigami-structured device expands in surface area with slight deformation in the z-axis. Conversely, during the contraction phase, the device opposes deformation in the x- and y-axis and undergoes deformation in the z-axis. Thus, the device captures the muscle movement across all three dimensions by measuring the corresponding deformation, which generates the change of magnetic flux density followed by the induction of an electrical signal in the MI layer. Supplementary Note2 further demonstrates the response of the device to the omnidirectional laryngeal movements and how the kirigami structure ensures the sensing performance.
The key defining characteristic of this system (MC layer) is based on the magnetoelastic effect, which refers to a change in the magnetic flux density of a ferromagnetic material in response to an externally applied mechanical stress, which was discovered in the mid-19th century54. It has been observed in rigid metals and metal alloys such as Fe1xCox54, TbxDy1x Fe2 (Terfenol-D)55, and GaxFe1x (Galfenol)56. Historically, these materials received limited attention within the bioelectronics domain for several reasons: the magnetization variation of magnetic alloys within biomechanical stress ranges is limited; the necessity for an external magnetic field introduces structural intricacies; and a significant mechanical modulus mismatch exists between magnetic alloys and human tissue, differing by six orders of magnitude. However, a breakthrough occurred in 2021 when the pronounced magnetoelastic effect was observed in a soft matter system57. This system exhibited a peak magnetomechanical coupling factor of 7.17108TPa1, representing an enhancement up to fourfold compared to traditional rigid metal alloys, underscoring its potential in soft bioelectronics. Functionally, the MC layer converts the mechanical movement of extrinsic laryngeal muscle into magnetic field variation, and the copper coils transfer the magnetic change into electrical signals based on electromagnetic induction, operating in a self-powered manner. While additional power management circuits are essential for processing and filtering the signals, the initial sensing phase is autonomous and does not rely on an external power supply. After recognition through the machine learning model, the voice signal is output through the actuation system (Fig.1a).
The signal conversion through the giant magnetoelastic effect in soft elastomers can be explained at both the micro and atomic scales. At the microscale, compressive stress applied to the soft polymer composite causes a corresponding shape deformation, leading to magnetic particle-particle interactions (MPPI), including changes in the distance and orientation of the inter-particle connections. The horizontal rotation of each subunit in the kirigami structure (Fig.1d) and vertical bending deformation (Fig.1e) create a micro change of magnetic density. In detail, as shown in Fig.1f, in a subunit of the kirigami structure, deformation-induced angle shift generates a concentration of stress and MPPI in between each single unit of the kirigami structure. At the atomic scale, mechanical stress also induces magnetic dipole-dipole interactions (MDDI), which results in the rotation and movement of magnetic domains within the particles. As shown in Fig.1g, a torque was made on each magnetic nanoparticle, and the shift of angle generates the change in magnetic flux density. The photo of the device design is presented in Fig.1h, i as the x-, y-axis, and z-axis response in the expansion phase; and in Fig.1j, k as in the contraction phase. Fig. 1h,j describes the expansion and contraction in the x-y plane, and Fig. 1i, j describes the corresponding z-axis contraction and expansion.Such structural design also displays a series of appealing features, including high current generation, low inner impedance, and intrinsic waterproofness, which will be presented in the following sections.
Our present work compares previous approaches based on PVDF and graphene for flexible voice monitoring and emitting, as shown in Fig.2a and Supplementary TableS335,36,37,58,59,60,61,62. The device developed in this work has a similar acoustic performance, with a frequency range covering the entire human hearing range. However, it has a much lower driving voltage (1.95V) and a Youngs modulus of 7.83105Pa. As shown in Fig.S3, it exhibits the stress-strain curve and testing photo of the material with and without the kirigami structure, which lowered Youngs modulus from 2.59107Pa to 7.83105Pa. This result ensures a higher comfort level while wearing as the modulus of the device is very close to that of the human skin. Notably, the device we developed has two unique features of stretchability and water resistance, which ensure the detection of horizontal movements, wearing comfort and resistance to respiration. Additionally, the device does not have the issue of temperature rising during use, preventing unexpected low-temperature scalding of users. Subsequently, several standard tests establish the sensing features of the device and its efficacy in outputting voice signals. To enhance the stretchability of the device, a kirigami structure was fabricated onto the MC layer of the device. The unit design of the structure is shown in Fig.S4, and the stretchability with regards to the parameters of the kirigami unit design is exhibited in Fig.S5. Such an approach not only enhances the stretchability of the device to a maximum of 164% with Youngs modulus at the level of 100kPa but also realizes isotropy. Furthermore, the structure enlarges the horizontal deformation of the device under unit pressure, generating a higher current output and enhanced detectable signals of extrinsic muscle contraction and relaxation, as shown in Figs.S6S7. The change in sensitivity brought by the structure on the vertical axis was also tested, and an elevation can be observed, as shown in Figs.S8S9. Moreover, isotropy prevents the device from being disturbed by random and uneven body movements in use. Thus, there are no requirements on wearing orientation which elevates user-friendliness as revealed in Figs.S10S11.
a Performance comparison of different flexible throat sensors in terms of Youngs modulus, stretchability, underwater sound pressure level, temperature rise, driving voltage, and working frequency range. b Pressuresensitivity response of the device at varied degrees of stretching under different amplification levels. (arb. units) referring to arbitrary units. c Response time and signal-to-noise ratio of the device. d Variation of sound pressure level with distance from the device at different amplification levels. e Sound pressure level of the device with resonance point highlighted in human hearing frequency range compared to SPL normal human speaking threshold. f The right shift of the first resonance point towards high frequency with regards to increasing strains. The test of performance is repeated 3 times at each condition. Data are presented as mean valuesSD. g Relationship between Kirigami structure parameters and actuating (first resonance point and sound pressure level)/sensing properties (response time and signal-to-noise ratio). The test of performance is repeated three times at each condition. Data are presented as mean valuesSD. Waveform (h) and spectrum (i) comparison of commercial loudspeaker (Red) and the device (Yellow) sound output at 900Hz and maximum strain (164%).
The stretchable structure of the device was leveraged to examine its sensitivity with respect to deformation degrees, as depicted in Fig.2b. The sensitivity curve demonstrated consistency under varying strains, with a minor change observed under maximum strain (164%). This change could be attributed to the reduction in the MC layers thickness due to deformation, which in turn decreases the magnetic flux density under the same pressure level, resulting in lower current generation. The devices response curve under different frequencies and forces of the shaker was tested, as shown in Fig.S12. We have also validated that the electric output of the device is not due to the triboelectricity in Supplementary Note363. The devices inherent flexibility and stretchability facilitate tight adherence to the throat, yielding a high signal-to-noise ratio (SNR) and swift response time (Fig.2c). In addition to the kirigami structure design parameters, other factors influencing the devices sensitivity, response time, and SNR were also evaluated. Fig.S13 illustrates that an increase in coil turns results in longer response times and lower SNR due to the increased total thickness of the copper coils. This thickness impedes the membranes deformation during vibrations, leading to longer response times and lower signal quality. We have further investigated the increase of thickness with the coil turn ratios in Supplementary TableS2. As the number of coil turns escalates, theres a direct correlation with the likelihood of copper wires stacking. Consequently, a significant number of samples exhibit thicknesses approximating 2 or 3 layers of copper (134 m and 201 m, respectively). This stacking effect amplifies the average coil thickness as the number of turns increases. However, this augmentation isnt strictly linear. For instance, the propensity for overlapping is less pronounced for turn ratios of 20 and 40. In contrast, for turn ratios exceeding 60, a clear trend emerges where the likelihood of overlapping increases with the number of turns. The relationship between the sensing performance and nanomagnetic powder concentrations of the MC Layer is presented in Fig.S14. A semi-linear relationship was observed, with higher magnetic nanoparticle concentration generating a stronger magnetic field and, consequently, higher current output. The influence of varying PDMS ratios in the sensing membrane on the performance of the sensor is delineated in Fig.S15. An increase in the PDMS ratios was found to extend the response times and decrease the SNR while having a negligible effect on the sensitivity curve. The augmentation in PDMS ratios leads to a softer membrane, which is prone to deformation at a slower rate. Consequently, devices with higher PDMS ratios exhibit heightened sensitivity to noise-generating deformations, albeit at a reduced response time. The influence of thickness on sensing performance was tested in Fig.S16, with thicker membranes resulting in quicker response times and a fluctuating SNR. Lastly, the impact of the MC layers thickness was tested in Fig.S17. A thicker MC layer had no influence on response time but reduced SNR. Weve consolidated the results of each optimization factor in Fig.S18, providing a clear overview of the primary variables influencing each performance metric. After considering the sensing performance, weight, and flexibility of the device, the current parameters were determined. The devices durability with these parameters was evaluated in Fig.S19, where the device underwent continuous working for 24,000 cycles with a shaker under a frequency of 5Hz, with no observable degradation in the current generation.
The acoustic performance of the actuation system of the device is examined firstly with a focus on its sound pressure level (SPL) at different distances. The results, presented in Fig.2d, show that larger output magnification led to a higher SPL at all tested positions. Even at a distance of 1 meter, the typical distance during normal conversations, the device provided an SPL of over 40dB, which is above the lower limit of normal speaking SPL (4060dB)64. We also tested the devices SPL at different angles and compared its performance with those of previous works on acoustic devices (Fig.S20, Supplementary TableS3). The devices performance across various frequencies was tested and presented in Fig.2e, which indicates that it could provide sound with SPL louder than normal speaking loudness across the entire human hearing range64. The resonance point in the figure indicates the frequency at which the device has relatively the largest loudness output under the same signal strength as other adjacent frequencies. Further investigation into the SPL regarding frequency under different strains revealed that the first few resonance points tended to have the largest acoustic output across the frequency range (Fig.S21). Since the device under one strain has multiple resonance points that change non-linearly with deformation, investigating the change of every resonance point is complicated. Therefore, we only investigated the first resonance point (FRP) in Fig.2f because of its complexity and our interest in the highest output. According to Fig.2e and Fig.S22, the voice output at each strain was above the normal talking threshold across the whole human hearing range. Figure2f revealed a right shift of FRP of the device as the deformation gets larger, enabling the device to adjust its best output performance under different usage scenarios. Our device can adjust its best output performance by simply changing the deformation degree, thus creating a unique output setting for each individual and realizing user adaptability. More details about the right shift of FRP are shown in Fig.S23.
We also tested the influence of introducing the kirigami design into the device, as presented in Fig.2g. The results show that the parameter of the kirigami design had a negligible impact on the sensing and acoustic performance, further supporting the decision to use this design due to its impact on flexibility (Fig.S5). Additional factors influencing the acoustic performance of the actuation system were evaluated, and the final parameters were determined based on both performance and the devices mass/flexibility. Fig.S24 explores the impact of coil turn ratios on the SPL produced by the device. It was observed that an increase in coil turns led to a decrease in SPL, likely due to the weight of the additional coil impeding membrane vibration and subsequently reducing SPL. The relationship between SPL and the PDMS ratio of the actuator membrane was examined in Fig.S25. As the ratio increased, the membrane softened, leading to a decrease in the generated SPL. The dampening effect of a softer membrane hindered vibration and sound generation, resulting in a semi-linear decrease. Fig.S26 presents the relationship between SPL and magnetic powder concentrations. The devices SPL increased with the addition of higher amounts of magnetic powder in the MC layer, plateauing after a ratio of 4:1. The effect of varying MC layer thickness on SPL is shown in Fig.S27. A sharp increase in the devices SPL was observed as the MC layers thickness increased from 0.5mm to 1mm. However, the increase slowed and eventually plateaued as the MC layer became thicker. Finally, the SPL under different actuator membrane thicknesses was tested in Fig.S28. The devices SPL increased as the PDMS membrane (vibrating membrane) thickness increased from 100 to 200m but decreased when the membrane became thicker. The weight of thicker membranes may dampen the vibration and reduce the loudness produced by the device. Regarding the acoustic output quality of the device, Fig.2h displays the waveform of the commercial loudspeaker and our device at the maximum (164%) strain at the frequency of 1100Hz. The device reproduced the voice signal accurately, even under maximum deformation, with only slight distortion. The distortion was further explained in the spectrogram of Fig.2i, which shows that a noise of around 1400Hz was generated in the output of our device but not strong enough to significantly distort the signal. Output of other strains was tested in Fig.S29, a similar distortion of less extent can be observed with less strain. In the final phase of our study, we evaluated the water resistance of our device. The waveform of the device outputting an identical voice signal segment under water and in air is depicted in Fig.S30. The waveforms are notably similar, with no significant signal distortion observed. A slight loss of the high-frequency component, without major signal attenuation, is evident in the frequency domain (Fig.S31). The device demonstrated consistent performance even after being submerged in water for an accelerated aging test with aduration of 7 days (Fig.S32). The sound pressure level (SPL) in relation to distance underwater is presented in Fig.S33. A correlation was observed between the depth of the device underwater and the sound output, with deeper submersion resulting in lower output. However, the device could produce an output exceeding 60dB when placed 2cm underwater at a distance of 20cm. The SPL of the device in relation to frequency underwater is illustrated in Fig.S34. Despite the attenuation of high-frequency components underwater, the device consistently delivered an SPL above the normal speaking range (60dB) across the entire human hearing range. These results suggest that our device, as a wearable, can effectively withstand conditions of perspiration, damp environments, and rain exposure.
After obtaining the preliminary standard test results, we focused on collecting laryngeal muscle movement signals using our wearable sensing component. The experiment is schematically illustrated in Fig.3a. The analog signal generated by the vibration of the extrinsic laryngeal muscles (Sternothyroid muscle, as shown in Fig.3a) was collected by the sensor and then passed through an amplifier and a low-pass filter exhibited in Fig.3b. The digital signal of the laryngeal muscle movements was output and collected for further analysis. The sensitivity and repeatability of the device were tested in Fig.3c with two successive different throat movements. The device was able to generate distinguishable and unique signals for each different throat movement, indicating its feasibility to detect and analyze different laryngeal movement properties. Furthermore, the device responded consistently to one throat movement, as demonstrated by the participants continuous two throat movements. In addition, larger throat muscle movements, such as coughing or yawning, generated larger peaks, while longer movements, such as swallowing, generated longer signals. We also conducted experiments to test the devices functionality under different conditions. In Fig.3d, we asked the participant to voicelessly pronounce the same word (UCLA) under different conditions, including standing still, walking, running, and jumping. The device was able to discern the unique and repeatable feature syllable wave shape of each word, with only slight differences made by the participants with different pronouncing paces each time. Thus, the wearable device was able to function without being influenced by the users body movements, even during strenuous exercise. Finally, to test the signal quality and accuracy acquired by purely the laryngeal muscle movement, we performed examinations to compare normal speaking and voiceless speaking, as shown in Fig.3e. The five successive signals of participant saying Go Bruins with and without vocal fold vibration were compared in Fig.3f and g, respectively. Both tests generated consistent signals, and the syllables of each word were represented with distinguishable waveforms. Comparing the test results of normal speaking and speaking voicelessly, we observed only a slight loss of maximum amplitude in the signal of speaking voicelessly. This could be explained by the fact that the vibration of vocal folds requires more and stronger muscle movements, thus generating stronger signals. Furthermore, a clear loss of high-frequency components in voiceless signals compared to the signals with vocal fold vibration was observed in Fig.3h, i after the Fourier transform of both signals across frequencies. This finding was consistent with our hypothesis that the high-frequency part of the vibration generated by intrinsic muscles and vocal folds is absent in voiceless signals, leaving a smoother yet distinguishable waveform. Hence, the device was proven to capture recognizable and unique signals with laryngeal muscle movements for further analysis.
a Schematic illustration of the extrinsic muscle and vibration. Created with BioRender.com. b Circuit diagram of the system for collecting extrinsic muscle movement signal. c Sensor output for different throat movementsCoughing, Humming, Nodding, Swallowing, and Yawning. d Device signal output for participant pronouncing UCLA under different body movements. e Sensor output for participant pronouncing Go Bruins! with vocal fold vibration (upper, gray) and voiceless (lower, red). Enlarged waveform of participant pronouncing Go Bruins! with vocal fold vibration (f) and voiceless (g). Amplitude-frequency spectrum of the signal with vocal fold vibration (h) and voiceless (i).
With generated data of laryngeal muscle movement, a machine-learning algorithm was employed to classify the semantical meaning of the signal and select a corresponding voice signal for outputting through the actuation component of the system. A schematic flow chart of the machine-learning algorithm is presented in Fig.4a. The algorithm consists of two steps: training and classifying a set of n sentences for which assisted speaking is required. Firstly, the filtered training data was fed to the algorithm for model training. The electrical signal of each of the n sentences was compacted into an Nth-order matrix for feature extraction with principal component analysis (PCA) (Fig.4b). N is determined by the sampling window, which is the length of the longest sentences signal. PCA is applied to remove redundancy and prepare the signal for classification. Multi-class support vector classification (SVC) was chosen as the classification algorithm with the decision function shape of one vs. rest. For each sentence to be classified, the rest of the n-1 sentences were considered as a whole to generate a binary classification boundary to discriminate the target sentence. A brief illustration of the support vector machine (SVM) process is depicted in Fig.4c. The margin of the linear boundary between two target data groups undergoes a series of optimizing processes and was set to the largest with support vectors. Details of PCA and multi-class SVC are discussed in Methods. After the classifier was trained with pre-fed training data, it was used for classifying newly collected laryngeal muscle movement signals. The real-time data were fed to the classifier, and the class (which sentence) of the signal was output for voice signal selection. Subsequently, the corresponding pre-recorded voice signal was played by the actuation component, realizing assisted speaking.
a Flow chart of the machine-learning-assisted wearable sensing-actuation system. b Illustration depicting the process of data segmentation and principal components analysis (PCA) applied to the muscle movement signal captured by the sensor. Yellow indicates one sentence, and red indicates another one. c Optimizing process of data classification after PCA with support vector machine (SVM) algorithm. d Contour plot of the classification results with SVM, class 1, indicating 100% possibility of the target sentence, dotted lines are the possibility boundaries between the target sentence and the others. e Bar chart exhibiting 7 participants accuracy of both validation set and testing set. f Confusion matrix of the 8th participants validation set with an overall accuracy of 98%. g Confusion matrix of the 8th participants testing set with an overall accuracy of 96.5%. h Demonstration of the machine-learning-assisted wearable sensing-actuation system in assisted speaking. The left panel shows the muscle movement signal captured by the sensor as the participant pronounces the sentence voicelessly, while the right panel shows the corresponding output waveform produced by the systems actuation component. i The SPL and temperature trends over time while the device is worn by participants; no notable temperature increase or SPL decrease was seen for up to 40min. j The devices SPL outputs participant-specific sound signals, both with and without sweat presence. Each participant was asked to repeat testing of N=3 times for both scenarios. Data are presented as mean valuesSD. The p-value between dry and sweaty state is calculated to be 0.818, indicating no significant difference in the devices performance under the two cases. k The devices SPL across various conversation angles while done by the participant. Created with BioRender.com.
A brief demonstration was made with five sentences that we had selected for training the algorithm (S1: Hi Rachel, how you are doing today?, S2: Hope your experiments are going well!, S3: Merry Christmas!, S4: I love you!, S5: I dont trust you.). Each participant repeated each sentence 100 times for data collection. The resulting contour plot in Fig.4d shows an example of the classification result, with the red dots indicating the target sentence and the yellow dots indicating the others. A probability contour was drawn to classify whether a newly input sentence point belonged to the target sentence or not. With the trained classifier, the laryngeal movement signal was recognized for the corresponding sentence that the participant wished to express. To test the robustness and user-adaptability of the algorithm, the device was tested with eight participants, each repeating the sentence 120 times in total, with 100 repeats selected for the training set and 20 separated as the testing set. Of the 100 repeats, 20 were selected as the validation set. Figure4e shows the validation and testing results of seven out of the eight participants, while Fig.4f, g presents a detailed illustration of the confusion matrix of the 8th participant for the validation and testing sets, respectively. Even slightly lower than the validation set, each participants testing set achieved more than 93% accuracy. FigureS35 shows the detailed confusion matrix of both the validation and testing set and the accuracy of every other participant. The overall prediction accuracy of the model was 94.68%, and it worked well with different participants. Each participants voice signal was played by the actuation component, realizing the demonstration in Fig.4h. The left panel shows the muscle movement signal transferred into the correct voice signal, with the waveform shown in the right panel. Further, we extended our analysis to validate the practical usability of the device for vocal output after the selection of the accurate voice signal by the algorithm. As demonstrated in Fig.4i, an evaluation of the SPL and temperature of the device during use by the participant revealed no significant drop in SPL or rise in temperature, even after an extended working period of 40min. This suggests the devices durability in voice output and safe usage. In Fig.4j, we display the SPL of the device as it produces voice signals for seven participants, both with and without sweat. We noted consistent performance by the device across different participants, with no evident signal attenuation despite the presence of perspiration. Finally, Fig.4k illustrates the devices SPL during voice output at various normal conversation angles while worn by the participant. The device demonstrated reliable sound performance across all angles, thereby enabling assisted speaking in multiple real-life scenarios. In conclusion, the device can convert laryngeal muscle movement into voice signals, providing patients with voice disorders with a feasible method to communicate during the recovery process.
DARPA and IBM Secure AI Systems from Hackers – AiThority
The US Department of Defenses (DoD) research and development arm, DARPA, and IBM have been collaborating on several projects related to hostile AI for the past four years. The team from IBM has been working on a project called GARD, which aims to construct defenses that can handle new threats, provide theory to make systems provably robust and create tools to evaluate the defenses of algorithms reliably. The project is led by Principal Investigator (PI) Nathalie Baracaldo and co-PI Mark Purcell. To make ART more applicable to potential use cases encountered by the US military and other organizations creating AI systems, researchers have upgraded it as part of the project.
Read: Data monetization With IBM For Your Financial benefits
With the hope of inspiring other AI experts to collaborate on developing tools to safeguard AI deployments in the actual world, IBM gave ART to the Linux Foundation in 2020. In addition to supporting numerous prominent machine-learning model structures, like TensorFlow and PyTorch, ART also has its own GitHub repository. To continue meeting AI practitioners where they are, IBM has now added the updated toolkit to Hugging Face. When it comes to finding and using AI models, Hugging Face has swiftly risen to the top of the internet. The current geographic model developed with NASA is one of many IBM projects that have been made publicly available on Hugging Face. Models from the AI repository are the intended users of Hugging Faces ART toolset. It demonstrates how to include the toolbox into time, a library utilized to construct Hugging Face models, and provides instances of assaults and defenses for evasion and poisoning threats.
Read:Top 10 Benefits Of AI In The Real Estate Industry
The researchers in this dispersed group would use their standards to assess the efficacy of the defenses they constructed. ART has amassed hundreds of stars on GitHub and was the first to provide a single toolset for many practical assaults. This exemplifies the communitys cooperative spirit as they strive toward a common objective of protecting their AI procedures. Although they have come a long way, machine-learning models are still fragile and open to both targeted attacks and random noise from the real world.
A disorganized and immature adversarial AI community existed before GARD. Digital assaults, such as introducing small disturbances to photos, were the main focus of researchers, although they werent the most pressing issues. Physical attacks, such as covering a stop sign with a sticker to trick an autonomous vehicles AI model, and attacks where training data is poisoned are the main concerns in the real world.
New:10 AI ML In Personal Healthcare Trends To Look Out For In 2024
Researchers and practitioners in the field of artificial intelligence security lacked a central hub for exchanging attack and defense codes before the advent of ART. To accomplish this, ART offers a platform that enables teams to concentrate on more particular tasks. As part of GARD, the group has created resources that blue and red teams can use to evaluate and compare various machine learning models performance in the face of various threats, such as poisoning and evasion. What is included in ART are the practical countermeasures against those attacks. Although the project is coming to a close this spring after four years, it is far from over.
[To share your insights with us as part of editorial or sponsored content, please write tosghosh@martechseries.com]
Excerpt from:
DARPA and IBM Secure AI Systems from Hackers - AiThority
Identifying microRNAs associated with tumor immunotherapy response using an interpretable machine learning model … – Nature.com
ICB response prediction using miRNA expression profiles
We compiled predictive ICB responses using the TIDE model and miRNA expression profiles from 7721 samples across 19 different tumor types within The Cancer Genome Atlas (TCGA) dataset (Table 1). To predict immunotherapy response using miRNA expression profiles, we first developed a random forest classifier to determine CTL levels. The optimal parameters for the random forest classifier were determined through a grid search with tenfold cross-validation (Table 2). Using the identified optimal parameters, we trained random forest classifiers on the designated training data and rigorously assessed the predictive performance on the independent test data. The results showed that the random forest classifier predicted the CTL levels well, with an AUC of 0.9400 (Fig.2A). Furthermore, when evaluating the performance using the F1 score and Balanced AUC indicators, high performance was confirmed, with an F1 score of 0.9849 and a Balanced AUC of 0.7182.
Predicted results for each model learned using miRNA expression profiles. (A) ROCAUC of the random forest classifier that predicts the CTL level. The class True signifies the high group and False signifies the low group. (B) Scatterplot of the random forest regression model for predicting the dysfunction score. The red line indicates the regression line. (C) Scatterplot of the random forest regression model for predicting the exclusion score. The red line indicates the regression line. (D) Scatterplot of the stepwise prediction model predicting ICB response based on the TIDE score. The red line indicates the regression line.
Next, we predicted the dysfunction and exclusion scores based on random forest regression. A grid search with tenfold cross-validation was performed to determine the optimal parameters for random forest regression (Table 2). Employing the optimal parameters, two random forest regression models to predict the dysfunction and exclusion scores were independently learned from the training data. The predictive results with the independent test datasets showed that the MSE of the regression model for predicting the dysfunction and exclusion scores were both 0.0361. The Pearson correlation coefficient (PCC) between the observed and predicted values was also calculated. The PCC for the dysfunction score prediction model was 0.8158 and that for the exclusion model was 0.8704. This indicated a strong positive correlation between the predicted and actual values in both models (Fig.2B,C).
Finally, we predicted the ICB responses based on the TIDE score by combining the two-step machine learning model, constructed a random forest classifier for CTL prediction, and random forest regression models for the dysfunction and exclusion scores. The MSE of the combined stepwise model was 0.0360. Furthermore, the PCC between the observed and predicted values exhibited a strong positive correlation of 0.9270 (Fig.2D).
Thereafter, we used SHAP, an interpretable machine learning approach, to analyze the results of our machine learning models. Using SHAP analysis, we identified informative miRNAs that contributed to the prediction of target values. Figure3 shows the top 20 miRNAs ranked according to their feature importance scores in each model.
Shapley value plot for exhibiting feature importance. (A) SHAP feature importance for the random forest classifier to predict CTL level, (B) summary plot for the random forest classifier when the CTL prediction model predicts the CTL level is high, (C) summary plot for the random forest classifier when the CTL prediction model predicts the CTL level is low, (D) SHAP feature importance for random forest regression to predict dysfunction score, (E) summary plot for random forest regression to predict dysfunction score, (F) SHAP feature importance for random forest regression to predict exclusion score, and (G) summary plot for random forest regression to predict exclusion score. (A, D, F) are plots that arrange features based on the average of the absolute Shapley values, which serve as indicators of feature importance. (B, C, E, G) are summary plots that depict feature importance and feature effects simultaneously. Each point signifies the Shapley value of the feature and instance. The x-axis represents the Shapley value, and the y-axis represents each feature. The color of each point corresponds to the high and low feature values (i.e., miRNA expression values).
For CTL-level prediction based on a random forest classifier, hsa-miR-155 was the most informative feature with the highest Shapley value. In particular, focusing on high and low CTL predictions, the expression of hsa-mir-155 was positively associated with CTL-level prediction (Fig.3B,C). Notably, miR-155 is an essential factor orchestrating the CD8+T cell response in cancer, and its overexpression has been associated with the enhancement of the anti-tumor response22,23. hsa-miR-150, which had the second-highest impact on model predictions, exhibited a similar trend. miR-150 also plays a crucial role in the differentiation and functional regulation of CD8+T cells24. The absence of miR-150 leads to a decline in the killing ability of CD8+T cells24. In addition, hsa-miR-4772, hsa-miR-21, hsa-miR-142, and hsa-miR-10a were also identified with notably high Shapley values.
In the random forest regression model used to predict the dysfunction score, the miRNA with the highest Shapley value was hsa-miR-10b. The Shapley value of hsa-mir-10b was negative when its expression was low, and positive when its expression was high (Fig.3E). This indicated a positive correlation between hsa-miR-10b expression and dysfunction prediction. In contrast, hsa-miR-183 negatively correlated with dysfunction prediction. Both miR-150 and miR-155 showed positive correlations in dysfunction predictions and played an important role in dysfunction mechanisms, as well as in CTL level predictions. Furthermore, miR-151a and miR-210 exhibit negative correlations, similar to those of miR-183.
In the random forest regression model predicting the exclusion score, hsa-miR-10b also showed the largest Shapley value (Fig.3G); however, it exhibited a negative correlation with hsa-miR-10b and the exclusion prediction, in contrast to the dysfunction prediction model. This observation serves as an example of how exclusion prediction, which has a mechanism opposite to that of dysfunction, is negatively correlated with dysfunction prediction. In contrast to the dysfunction results, hsa-miR-150 and hsa-miR-155 demonstrated opposite behaviors in exclusion prediction. Additionally, hsa-miR-10a, which was also identified in the CTL-level prediction, showed a positive correlation with exclusion prediction and played an important role in model prediction. Furthermore, the expression level of miR-194-1 and miR-194-2 is negatively correlated to the exclusion prediction.
Next, we verified whether ICB response could be predicted using a small number of informative miRNAs. We selected miRNAs with an average absolute Shapley value of 0.01 or higher (SHAP 0.01). Using this criterion, three miRNAs were identified in the CTL model, five miRNAs in the dysfunction prediction model, and 12 in the exclusion prediction model (Fig.3A,D,F). Because only a limited number of features were used to construct the models, we employed a simple algorithm to predict immunotherapy response.
To predict the CTL level, we applied logistic regression25 and determined the optimal parameters by conducting a grid search with tenfold cross-validation (Table 2). The model using the three informative miRNAs achieved an F1 score of 0.9805, a balanced accuracy of 0.7249, and an AUC value of 0.9300 (Fig. S1A). This analysis confirmed that a small subset of highly informative miRNAs displayed a similar performance in predicting CTL levels, even when a logistic regression model was utilized.
Subsequently, dysfunction and exclusion scores were predicted using a small number of informative miRNAs based on multiple linear regression. The obtained results showed that the MSE for the dysfunction model using the top miRNA (SHAP 0.01) was 0.0754 and that for the exclusion prediction model was 0.0840. The PCCs between the predicted and actual values were 0.5707 and 0.6638 for the dysfunction and exclusion prediction models, respectively (Figs. S1B,C). From these results, we confirmed that the performance was slightly degraded with a reduced number of features; however, the models still demonstrated comparable performance with only a small number of selected miRNAs.
Finally, to predict the ICB responses based on the TIDE scores, we applied a stepwise machine learning model by combining the logistic regression classifier for the CTL level and the linear regression model for dysfunction and exclusion scores. The MSE of the model that used the most informative miRNA (SHAP 0.01) was 0.0690. We also observed a strong positive correlation with informative miRNAs; the PCC of the top miRNA (SHAP 0.01) was 0.8457 (Fig. S1D).
Similarly, we applied robust criteria for the identification of informative miRNAs and verified whether having fewer miRNAs could result in the accurate prediction of immunotherapy response. We selected miRNAs with an average absolute Shapley value of 0.02 or higher (SHAP 0.02); two miRNAs were identified in the CTL model, four miRNAs in the dysfunction prediction model, and five miRNAs in the exclusion prediction model (Fig.3A,D,F).
For CTL level prediction using logistic regression, the model obtained an F1 score of 0.9800, a balanced accuracy of 0.7459, and an AUC value of 0.91 (Fig. S2A). In addition, the models showed good performance for dysfunction and exclusion score prediction using linear regression. The MSE for dysfunction prediction was 0.0810 and that for exclusion prediction was 0.0984. The PCCs were 0.5220 and 0.5900 for the dysfunction score and exclusion score prediction models, respectively (Fig. S2B,C). Furthermore, for the ICB response prediction based on the TIDE scores using the two-step machine learning model combining logistic regression and linear regression, the MSE was 0.0753 and the PCC was 0.8595 (Fig. S2D). Although the performance was slightly lower than that of the model using all miRNAs for predicting the ICB response, these results suggest that informative miRNAs based on Shapley values still exhibit strong predictive capability, even with a limited number of miRNAs and relatively simple classification and regression models.
To examine the biological roles of the informative miRNAs, we predicted the target genes of the informative miRNAs selected by Shapley values using miRDB and TargetScan. A list of the genes targeted by the top miRNAs from each model is shown in Tables S1 and S2. We investigated the Kyoto Encyclopedia of Genes and Genomes (KEGG) pathways enriched in the target genes. Tables 3, 4, 5 and Tables S3S5 show the results of enrichment analyses using the informative miRNAs (SHAP 0.01) of each model. The top 20 pathways are listed in Tables 3, 4, 5 in ascending order of P-values, and all KEGG pathways satisfying statistical significance (adjusted P value<0.05) are shown in Tables S3S5.
The first-ranked KEGG pathway in the CTL-level prediction model was the TNF signaling pathway (Table 3). The following pathways are involved in the Hepatitis B and IL-17 signaling pathway. Hepatitis B is a significant contributor of hepatocellular carcinoma (HCC)26. Additionally, immune-related pathways such as the Fc epsilon RI signaling pathway and the T cell and B cell receptor signaling pathways were observed at the top. Enrichment analysis also revealed several other cancer-related terms, including Pathways in cancer, PI3K-Akt signaling pathway, Prostate cancer, Renal cell carcinoma, and Pancreatic cancer (Table S3). These results suggest a significant role for these miRNAs and their target genes in cancer and immunotherapy.
Tables 4 and S4 present the informative pathways identified using the dysfunction score prediction model. One of the most significantly enriched pathways was melanogenesis, which produces mutagenic intermediates that induce immunosuppression. The following term represents the Wnt signaling pathway and the ErbB signaling pathway. Moreover, our analysis identified various cancer-related terms, including Hepatocellular carcinoma, Prostate cancer, Breast cancer, and Gastric cancer, as well as Pathways in cancer (Table S4).
Tables 5 and S5 present the pathways identified using the exclusion score prediction model. The first pathway is Proteoglycans in cancer, which plays a significant role in regulating cytokine and chemokine expression on the cell surface. Moreover, various cancer-related pathways and terms, such as MAPK signaling, PI3K-Akt signaling pathway, and Rap1 signaling pathway, Pathways in cancer, Prostate cancer, Renal cell carcinoma, Lung cancer, and Breast Cancer, were also identified, along with immune-related pathways like T cell and B cell receptor pathway and Helper T cell differentiation. Furthermore, the presence of PD-L1 expression and the PD-1 checkpoint pathway in cancer indicate that the genes targeted by miRNAs are directly associated with immunotherapy.
Additionally, it was noted that several pathways related to the brain and neurons were observed, including Axon guidance27, a subfield of neurodevelopment associated with the process of neurons sending axons to reach accurate targets; Neurotrophin signaling pathway28, a protein that supports the survival, development, and function of neurons; Long-term potentiation29, a process that strengthens signal transmission between neurons; as well as Dopaminergic synapse and Cholinergic synapse (Table S5). This could be because the majority of CTL-low (exclusion) samples were involved in the TCGA LGG tumor type (Table S6).
Enrichment analysis results using the top miRNAs (SHAP value 0.02) of each model also identified diverse pathways related to cancer and immunity (Table S7-S9). These findings would provide valuable insights into the molecular mechanisms underlying exclusion and immune response regulation in cancer.
We proceeded to validate the stepwise machine learning model based on a random forest trained on all miRNAs using data from 12 distinct tumor types not included in the previous training and test phases (Fig.1F,I and Table S10). For the random forest classifier predicting CTL levels, we achieved an F1 score of 0.9912 and an AUC value of 0.9400 (Fig. S3A). When predicting the dysfunction and exclusion scores via random forest regression models, the MSE for the dysfunction score prediction model was 0.0478, and that for the exclusion score prediction model was 0.0641. The MSE value of the stepwise machine learning model for predicting the ICB response based on the TIDE score was 0.0475. Moreover, it could be observed that both the predicted value and the actual value showed a positive correlation (PCC=0.8698). (Fig. S3B-D).
Furthermore, we validated the predictive potential of our immunotherapy response prediction model using small subsets comprising informative miRNAs (SHAP 0.01 and SHAP 0.02) by applying the same approaches to the 12 tumor types (Fig.1F,I). The models employing informative miRNAs (SHAP 0.01) to predict CTL levels using logistic regression showed an F1 score of 0.9901 and an AUC of 0.9300 (Fig. S4A). In the dysfunction and exclusion score predictions using linear regression, the MSE were 0.0660 and 0.0677, respectively. Moreover, a positive correlation was observed between the predicted and actual values (PCC=0.2899 and 0.4198, respectively) (Fig. S4B,C). Lastly, the stepwise model used to predict the ICB response based on the TIDE score with informative miRNA (SHAP 0.01) yielded an MSE of 0.0661 and a PCC of 0.8335 (Fig. S4D).
Additionally, the results of models with a smaller number of informative miRNAs and strict criteria (SHAP 0.02) revealed compelling outcomes. CTL-level prediction using the logistic regression classifier model showed an F1 score of 0.9904 and an AUC of 0.9300. (Fig. S5A). The linear regression models to predict the dysfunction and exclusion scores also achieved good performances, with the dysfunction score prediction model showing an MSE of 0.0585 and a PCC of 0.3822 and the exclusion score prediction model displaying an MSE of 0.0797 and a PCC of 0.2816 (Fig. S5B,C). In addition, for the prediction of the ICB response using the combined stepwise machine learning model with SHAP 0.02, the MSE was 0.0594 and the PCC was 0.8538 (Fig. S5D). Notably, the experimental results from the external validation datasets confirmed that not only did our model exhibit robust predictive performance regardless of tumor type, but the informative miRNAs were also useful for tumor immunotherapy response prediction.
We further validated the stepwise machine learning model trained on all miRNAs, using novel external independent data from PCAWG (Pancancer Analysis of Whole Genomes). The parameters of each model were set through grid search with tenfold cross-validation (Table S11). For the random forest classifier predicting CTL levels, we achieved an F1 score of 0.9589 and an AUC value of 0.9226 (Table S12). Regarding the prediction of dysfunction and exclusion scores through a random forest regression model, the MSE for the dysfunction score prediction model was 0.0245, and for the exclusion score prediction model, it was 0.0251 (Table S12). The MSE value of the stepwise machine learning model for predicting ICB response based on the TIDE score was 0.0248 (Table S12).
Furthermore, we identified informative miRNAs using the SHAP analysis in the PCAWG cohort (Fig. S6). In addition, we investigated which miRNAs were informative in each tumor type using SHAP (Table S13). It was noted that the informative miRNAs at TCGA cohorts were also similarly identified even at the PCAWG datasets, even though the direct comparison of the miRNAs is difficult because TCGA represents precursor miRNA expression and the PCAWG provides the mature forms. For instance, miR-150 demonstrated the significance in CTL and Dysfunction models. Furthermore, miR-155 was also assigned at a high ranking.
We also validated the predictability of ICB response prediction models in the PCAWG cohort using the informative miRNAs (SHAP 0.01 and SHAP 0.02) extracted from the TCGA cohort (Table S12). The model employing informative miRNAs (SHAP 0.01) achieved an F1 score of 0.9556 and an AUC of 0.9161 for predicting CTL levels via logistic regression. For dysfunction and exclusion score predictions using linear regression, the MSEs were 0.0371 and 0.0528, respectively. The stepwise model for predicting ICB response based on the TIDE score with informative miRNA (SHAP 0.01) yielded an MSE of 0.0376.
Similarly, the model utilizing informative miRNAs (SHAP 0.02) extracted from the TCGA cohort attained an F1 score of 0.9527 and an AUC of 0.9097 for predicting CTL levels via logistic regression. For dysfunction and exclusion score predictions using linear regression, the MSEs were 0.0364 and 0.0798, respectively. Finally, the stepwise model for predicting ICB response based on the TIDE score with informative miRNA (SHAP 0.02) yielded an MSE of 0.0364. The results with the external datasets from PCAWG further affirmed the effectiveness of the informative miRNAs in predicting ICB responses.
Next, we employed the random forest-based ICB response prediction model on the TCGA cohort, stratified by tumor type, to investigate variations in the efficacy of ICI treatment in each tumor type. The parameters of each model were set through grid search with tenfold cross-validation (Table S11). The MSE values of the combined stepwise models for each tumor type ranged from 0.0093 to 0.0494 (Table S12). Notably, these results closely similar to the predictive performance derived from the entire tumor cohort. Thus, this suggests that the differences in ICI treatment response among various cancer types are minimal.
In addition, we investigated which miRNAs were informative in each tumor type using SHAP (Table S13). Even though there existed some differences in each tumor type, some informative miRNAs such as miR-150 and miR-155 were frequently observed at the highly-ranked miRNAs. This result indicates that these miRNAs are closely related to ICB responses across the tumor types.
Moreover, we also evaluated how well the stepwise model pre-trained using the whole 19 TCGA cohorts predicted the test data (20%) for each tumor type (Table S14). The MSE was ranged from 0.0113 to 0.1824 using total miRNAs. Using the information miRNAs (SHAP 0.01), the MSE was ranged from 0.0166 to 0.5530. Similarly, in the SHAP 0.02 model, the MSE was ranged from 0.0159 to 0.5562. These results showed the informative miRNAs were utilized for the prediction of ICB treatment responses even at a variety of cancer types.
Read more here:
Identifying microRNAs associated with tumor immunotherapy response using an interpretable machine learning model ... - Nature.com
Will we ever be able to accurately predict solubility? | Scientific Data – Nature.com
Kennedy, T. Managing the drug discovery/development interface. Drug Discov. Today 2, 436444 (1997).
Article Google Scholar
Kola, I. & Landis, J. Can the pharmaceutical industry reduce attrition rates? Nat. Rev. Drug Discov. 3, 711716 (2004).
Article CAS PubMed Google Scholar
Millard, J., Alvarez-Nez, F. & Yalkowsky, S. Solubilization by cosolvents. Establishing useful constants for the log-linear model. Int. J. Pharm. 245, 153166 (2002).
Article CAS PubMed Google Scholar
Jouyban, A. & Abolghassemi Fakhree, M. A. Solubility prediction methods for drug/drug like molecules. Recent Pat. Chem. Eng. 1, 220231 (2008).
Article CAS Google Scholar
van de Waterbeemd, H. Improving compound quality through in vitro and in silico physicochemical profiling. Chem. Biodivers. 6, 17601766 (2009).
Article PubMed Google Scholar
Llompart, P. et al Will we ever be able to accurately predict solubility? Recherche Data Gouv https://doi.org/10.57745/CZVZIA (2023)
Wang, J. & Hou, T. Recent advances on aqueous solubility prediction. Comb. Chem. High Throughput Screen. 14, 328338 (2011).
Article CAS PubMed Google Scholar
Elder, D. P., Holm, R. & Diego, H. L. Use of pharmaceutical salts and cocrystals to address the issue of poor solubility. Int. J. Pharm. 453, 88100 (2013). de.
Article CAS PubMed Google Scholar
Saal, C. & Petereit, A. C. Optimizing solubility: Kinetic versus thermodynamic solubility temptations and risks. Eur. J. Pharm. Sci. 47, 589595 (2012).
Article CAS PubMed Google Scholar
Wang, J. et al. Development of reliable aqueous solubility models and their application in druglike analysis. J. Chem. Inf. Model. 47, 13951404 (2007).
Article CAS PubMed Google Scholar
Johnson, S. R. & Zheng, W. Recent progress in the computational prediction of aqueous solubility and absorption. AAPS J. 8, E27E40 (2006).
Article CAS PubMed PubMed Central Google Scholar
Delaney, J. S. Predicting aqueous solubility from structure. Drug Discov. Today 10, 289295 (2005).
Article CAS PubMed Google Scholar
OECD. Test No. 105: Water Solubility. OECD Guidelines for the Testing of Chemicals, Section 1 https://read.oecd-ilibrary.org/environment/test-no-105-water-solubility_9789264069589-en (1995).
Llins, A., Glen, R. C. & Goodman, J. M. Solubility Challenge: Can You Predict Solubilities of 32 Molecules Using a Database of 100 Reliable Measurements? J. Chem. Inf. Model. 48, 12891303 (2008).
Article PubMed Google Scholar
Stuart, M. & Box, K. Chasing Equilibrium: Measuring the Intrinsic Solubility of Weak Acids and Bases. Anal. Chem. 77, 983990 (2005).
Article CAS PubMed Google Scholar
Huuskonen, J., Rantanen, J. & Livingstone, D. Prediction of aqueous solubility for a diverse set of organic compounds based on atom-type electrotopological state indices. Eur. J. Med. Chem. 35, 10811088 (2000).
Article CAS PubMed Google Scholar
Yalkowsky, RM & Dannenfleser, SH. Aquasol database of aqueous solubility. Version 5. https://hero.epa.gov/hero/index.cfm/reference/details/reference_id/5348039 (2009).
Bloch, D. Computer Software Review. Review of PHYSPROP Database (Version 1.0). ACS Publications https://pubs.acs.org/doi/pdf/10.1021/ci00024a602 (2004) https://doi.org/10.1021/ci00024a602.
Dalanay, J. S. ESOL: Estimating Aqueous Solubility Directly from Molecular Structure. J. Chem. Inf. Comput. Sci. 44, 10001005 (2004).
Article Google Scholar
US EPA. EPI Suite. https://www.epa.gov/tsca-screening-tools/epi-suitetm-estimation-program-interface
Wang, J., Hou, T. & Xu, X. Aqueous Solubility Prediction Based on Weighted Atom Type Counts and Solvent Accessible Surface Areas. J. Chem. Inf. Model. 49, 571581 (2009).
Article CAS PubMed Google Scholar
Boobier, S., Hose, D. R. J., Blacker, A. J. & Nguyen, B. N. Machine learning with physicochemical relationships: solubility prediction in organic solvents and water. Nat. Commun. 11, 5753 (2020).
Article CAS PubMed PubMed Central Google Scholar
Tetko, I. V., Tanchuk, V. Y., Kasheva, T. N. & Villa, A. E. P. Estimation of Aqueous Solubility of Chemical Compounds Using E-State Indices. J. Chem. Inf. Comput. Sci. 41, 14881493 (2001).
Article CAS PubMed Google Scholar
Avdeef, A. Prediction of aqueous intrinsic solubility of druglike molecules using Random Forest regression trained with Wiki-pS0 database. ADMET DMPK 8, 29 (2020).
Article PubMed PubMed Central Google Scholar
Sorkun, M. C., Khetan, A. & Er, S. AqSolDB, a curated reference set of aqueous solubility and 2D descriptors for a diverse set of compounds. Sci. Data 6, 143 (2019).
Article PubMed PubMed Central Google Scholar
Sushko, I. et al. Online chemical modeling environment (OCHEM): web platform for data storage, model development and publishing of chemical information. J. Comput. Aided Mol. Des. 25, 533554 (2011).
Article CAS PubMed PubMed Central Google Scholar
Panapitiya, G. et al. Evaluation of Deep Learning Architectures for Aqueous Solubility Prediction. ACS Omega 7, 1569515710 (2022).
Article CAS PubMed PubMed Central Google Scholar
Wiercioch, M. & Kirchmair, J. Dealing with a data-limited regime: Combining transfer learning and transformer attention mechanism to increase aqueous solubility prediction performance. Artif. Intell. Life Sci. 1, 100021 (2021).
CAS Google Scholar
Lowe, C. N. et al. Transparency in Modeling through Careful Application of OECDs QSAR/QSPR Principles via a Curated Water Solubility Data Set. Chem. Res. Toxicol. 36, 465478 (2023).
Article CAS PubMed Google Scholar
Francoeur, P. G. & Koes, D. R. SolTranNet-A Machine Learning Tool for Fast Aqueous Solubility Prediction. J. Chem. Inf. Model. 61, 25302536 (2021).
Article CAS PubMed PubMed Central Google Scholar
Sluga, J., Venko, K., Drgan, V. & Novi, M. QSPR Models for Prediction of Aqueous Solubility: Exploring the Potency of Randi-type Indices. Croat. Chem. Acta 93 (2020).
Meng, J. et al. Boosting the predictive performance with aqueous solubility dataset curation. Sci. Data 9, 71 (2022).
Article CAS PubMed PubMed Central Google Scholar
Lee, S. et al. Novel Solubility Prediction Models: Molecular Fingerprints and Physicochemical Features vs Graph Convolutional Neural Networks. ACS Omega 7, 1226812277 (2022).
Article MathSciNet CAS PubMed PubMed Central Google Scholar
Schrdinger. QikProp. (2015).
United States National Library of Medicine. ChemIDplus advanced. https://pubchem.ncbi.nlm.nih.gov/source/ChemIDplus (2011).
Khne, R., Ebert, R.-U., Kleint, F., Schmidt, G. & Schrmann, G. Group contribution methods to estimate water solubility of organic chemicals. Chemosphere 30, 20612077 (1995).
Article Google Scholar
OECD. eChemPortal: The Global Portal to Information on Chemical Substances, https://www.echemportal.org/echemportal/ (2023).
European Chemicals Agency. ECHA. https://echa.europa.eu/fr/ (2023).
Irmann, F. Eine einfache Korrelation zwischen Wasserlslichkeit und Struktur von Kohlenwasserstoffen und Halogenkohlenwasserstoffen. Chem. Ing. Tech. 37, 789798 (1965).
Article CAS Google Scholar
Hansch, C., Quinlan, J. E. & Lawrence, G. L. Linear free-energy relationship between partition coefficients and the aqueous solubility of organic liquids. J. Org. Chem. 33, 347350 (1968).
Article CAS Google Scholar
Yalkowsky, S. H. & Valvani, S. C. Solubility and partitioning I: Solubility of nonelectrolytes in water. J. Pharm. Sci. 69, 912922 (1980).
Article CAS PubMed Google Scholar
Ran, Y. & Yalkowsky, S. H. Prediction of drug solubility by the general solubility equation (GSE). J. Chem. Inf. Comput. Sci. 41, 354357 (2001).
Article CAS PubMed Google Scholar
Hansen, N. T., Kouskoumvekaki, I., Jrgensen, F. S., Brunak, S. & Jnsdttir, S. . Prediction of pH-Dependent Aqueous Solubility of Druglike Molecules. J. Chem. Inf. Model. 46, 26012609 (2006).
Article CAS PubMed Google Scholar
ChemAxon. Marvin. https://chemaxon.com/products/marvin (2023).
Johnson, S. R., Chen, X.-Q., Murphy, D. & Gudmundsson, O. A Computational Model for the Prediction of Aqueous Solubility That Includes Crystal Packing, Intrinsic Solubility, and Ionization Effects. Mol. Pharm. 4, 513523 (2007).
Article CAS PubMed Google Scholar
Hopfinger, A. J., Esposito, E. X., Llins, A., Glen, R. C. & Goodman, J. M. Findings of the Challenge To Predict Aqueous Solubility. ACS Publications https://pubs.acs.org/doi/pdf/10.1021/ci800436c (2008).
Lusci, A., Pollastri, G. & Baldi, P. Deep architectures and deep learning in chemoinformatics: the prediction of aqueous solubility for drug-like molecules. J. Chem. Inf. Model. 53, 15631575 (2013).
Article CAS PubMed PubMed Central Google Scholar
Huuskonen, J., Livingstone, D. J. & Manallack, D. T. Prediction of drug solubility from molecular structure using a drug-like training set. SAR QSAR Environ. Res. 19, 191212 (2008).
Article CAS PubMed Google Scholar
Zhou, D., Alelyunas, Y. & Liu, R. Scores of Extended Connectivity Fingerprint as Descriptors in QSPR Study of Melting Point and Aqueous Solubility. J. Chem. Inf. Model. 48, 981987 (2008).
Article CAS PubMed Google Scholar
Eri, S., Kalini, M., Popovi, A., Zloh, M. & Kuzmanovski, I. Prediction of aqueous solubility of drug-like molecules using a novel algorithm for automatic adjustment of relative importance of descriptors implemented in counter-propagation artificial neural networks. Int. J. Pharm. 437, 232241 (2012).
Article PubMed Google Scholar
Llinas, A. & Avdeef, A. Solubility Challenge Revisited after Ten Years, with Multilab Shake-Flask Data, Using Tight (SD 0.17log) and Loose (SD 0.62log) Test Sets. J. Chem. Inf. Model. 59, 30363040 (2019).
Article CAS PubMed Google Scholar
Llinas, A., Oprisiu, I. & Avdeef, A. Findings of the Second Challenge to Predict Aqueous Solubility. J. Chem. Inf. Model. 60, 47914803 (2020).
Article CAS PubMed Google Scholar
Hewitt, M. et al. In silico prediction of aqueous solubility: the solubility challenge. J. Chem. Inf. Model. 49, 25722587 (2009).
Article CAS PubMed Google Scholar
Goh, G. B., Hodas, N., Siegel, C. & Vishnu, A. SMILES2vec: Predicting Chemical Properties from Text Representations. Preprint at arXiv:1712.02034 (2018).
Cui, Q. et al. Improved Prediction of Aqueous Solubility of Novel Compounds by Going Deeper With Deep Learning. Front. Oncol. 10 (2020).
Maziarka, . et al. Molecule Attention Transformer. (2020).
Lovri, M. et al. Machine learning in prediction of intrinsic aqueous solubility of drug-like compounds: Generalization, complexity, or predictive ability? J. Chemom. 35, e3349 (2021).
Article Google Scholar
Kohavi, R. & Wolpert, D. H. in International Conference on Machine Learning Bias Plus Variance Decomposition for Zero-One Loss Function (1996).
Follow this link:
Will we ever be able to accurately predict solubility? | Scientific Data - Nature.com
Weekly AiThority Roundup: Biggest ML and Automation Updates – AiThority
This is your AI Weekly Roundup. We are covering the top updates from around the world. The updates will feature state-of-the-art capabilities inartificial intelligence (AI),Machine Learning, Robotic Process Automation, Fintech, and human-system interactions. We cover the role of AI Daily Roundup and its application in various industries and daily lives.
Oracleannounced new generative AI capabilities within the Oracle Fusion Cloud Applications Suitethat will help customers improve decision making and enhance the employee andcustomer experience. The latest AI additions include new generative AI capabilities embedded in existing business workflows across finance, supply chain, HR, sales, marketing, and service, as well as an expansion of the Oracle Guided Journeys extensibility framework to enable customers and partners to incorporate more generative AI capabilities to support their unique industry and competitive needs.
Box, the leading Content Cloud, announced a new integration with Microsoft Azure OpenAI Service to bring its advanced large language models to Box AI. The integration of Azure OpenAI Service enables Box customers to benefit from the most advancedAI modelsin the world, while bringing Box and Microsofts enterprise-grade standards for security, privacy, and compliance to this groundbreaking technology.
Healthcare leaders are continuing to partner withQualtricsto make stressful and emotional care situations easier for their patients. Qualtrics helpshealthcare organizationsand healthcare providers better understand patient needs and respond to patients in the moment. In a recent Qualtrics report, consumers reported 16% of their recent experiences with hospitals and medical centers were very poor.
Royal Philips, a global leader inhealth technology, announced an expanded collaboration withAmazon Web Services (AWS) to address the growing need for secure, scalable digital pathology solutions in the cloud. The collaboration unites Philips leadership and expertise in digitization of pathology to optimize clinical workflows and AWS leadership in scalable, secure cloud solutions.
Salt Security,the leadingAPI securitycompany, released newthreat researchfromSalt Labshighlighting critical security flaws withinChatGPTplugins, highlighting a new risk for enterprises. Plugins provide AI chatbots like ChatGPT access and permissions to perform tasks on behalf of users within third party websites. For example, committing code to GitHub repositories or retrieving data from an organizations Google Drives. These security flaws introduce a new attack vector and could enable bad actors to:
Link:
Weekly AiThority Roundup: Biggest ML and Automation Updates - AiThority
Machine Learning in Fraud Prevention: Exploring how machine learning boosts fraud prevention capabilities – CXOToday.com
By Mr. Jinendra Khobare
Online fraud is a significant issue in India, with various scams such as phishing attacks, identity theft, and counterfeit e-commerce sites. Cybercrime in India has been on the rise, with the country recording over five thousand cases of online identity theft in 2022. Phishing attacks have also seen a surge, with around 83% of IT teams in Indian organisations reporting an increase in phishing emails targeting their employees in 2020. Furthermore, about 38% of consumers have received a counterfeit product from an e-commerce site in the past year.
According to a report, a significant portion of fraudulent transactions occur between 10 PM to 4 AM, with credit card holders over 60 years being the primary victims. From January 2020 to June 2023, 77.4% of cybercrimes were reported in India. The number of cybercrime cases in a city in India rose from 2,888 in 2020 to over 6,000 in 2023.
Machine learning is instrumental in fraud prevention, enabling organisations to detect and prevent suspicious activities in real-time. Traditional fraud prevention methods often struggle to keep up with the evolving tactics of scammers. Machine learning algorithms can quickly analyse vast amounts of data, helping organisations identify patterns and anomalies that may indicate suspicious behaviour. These algorithms learn from past fraud cases, continually enhancing their ability to detect suspicious activities. By integrating machine learning into their fraud prevention strategies, organisations can stay ahead of scams and safeguard their assets effectively.
A key advantage of machine learning in fraud prevention is its ability to detect suspicious activities at an early stage. By analysing historical data and identifying patterns of dubious behaviour, machine learning algorithms can spot suspicious transactions in real-time, enabling organisations to act swiftly and prevent financial losses.
Graph databases, alongside machine learning, have come out as a strong tool in fraud detection. Graph databases record and analyse network interactions at high rates, making them useful for a variety of applications, including fraud detection. They can identify patterns and relationships in big data, reducing the level of complexity so that detection algorithms can effectively discover fraud attempts within a network.
In conclusion, as scammers evolve their tactics, organisations must adapt their fraud prevention strategies to counter these threats effectively. Machine learning and graph databases are powerful weapons in this ongoing battle. With their ability to analyse countless data points rapidly, these technologies can detect suspicious activities accurately, surpassing human capabilities. Its akin to having a team of superhuman fraud detectives working tirelessly around the clock. As quickly as organisations detect and prevent suspicious activities, scammers are equally fast at devising new deception methods.
(The author is Mr. Jinendra Khobare, Solution Architect, Sensfrx, Secure Layer7, and the views expressed in this article are his own)
Accelerating the Development of New Molecules – Promethium to Empower ML Models for Drug Discovery Using … – Morningstar
Accelerating the Development of New Molecules - Promethium to Empower ML Models for Drug Discovery Using NVIDIA Quantum Cloud
PR Newswire
PALO ALTO, Calif., March 18, 2024
PALO ALTO, Calif., March 18, 2024 /PRNewswire/ --QC Ware, makers of Promethium, a quantum-inspired, molecular discovery platform, announced today it is leveragingNVIDIA Quantum Cloud to accelerate drug discovery by providing AI platforms with the ability to generate highly accurate training data on large, complex molecules faster than ever before, aiming to redefine the landscape of in silico molecular simulation. These are critical problems that will be accelerated by future quantum computers, but today are best solved by advanced methods on GPUs. This will enable molecular AI platforms to train better ML models, ultimately helping pharmaceutical companies find quality drug candidates more quickly.
NVIDIA Quantum Cloud is a microservice that lets users for the first time build and test in the cloud new quantum algorithms and applications, including powerful simulators and tools for hybrid quantum-classical programming.
"Promethium is helping steer the future of drug discovery and this latest work with NVIDIA signifies our commitment to advancing pharmaceutical research and underscores the importance of technological innovation in the quest for groundbreaking solutions to some of the most challenging medicinal problems of modern times," said Paul Baines, VP of Product & Engineering, at QC Ware.
ML models are already playing a large role in drug discovery, accelerating the ability to find new and promising candidates for clinical trials over historical methods. With an abundance of these ML models coming into play, one of the main challenges that research leaders face is determining the quality of the models they are using. Typically, the best ML models are trained on the best data. This has created a need for large amounts of extremely high-quality data.
The molecular data that currently exists is limited in scale, scope, and application. To build the best general-purpose ML model, or to customize an ML model for a specific purpose, lots of data must be generated and fast. Promethium is on a journey to generate the highest-quality data faster than it's ever been done. AI platforms for drug discovery are already taking advantage of these benefits. One example is a venture-backed start-up using Promethium to create highly accurate data, very quickly, and at scale.
"For representative systems we currently run today, Promethium is 30 times faster than our current toolsets," says the startup's CTO. "This means that we can generate high-quality training data faster, leading to extremely accurate ML models for researchers to use. The result? Truly transformative ML models that will unlock massive amounts of value for pharmaceutical companies."
Additionally, this type of data has yet to be created at scale for larger molecules. This was primarily due to how computationally intensive it was to create accurate data for molecular modeling. This limitation kept ML models from accurately predicting properties of larger molecules. Promethium's breakthrough software, combined with NVIDIA's state-of-the-art GPUs and cloud scalability, has unlocked the ability to create training data for large, complex molecules that will fuel the advanced capabilities of innovative AI platforms.
"Quantum computing has the potential to transform drug discovery," said Tim Costa, Director of HPC and Quantum at NVIDIA. "Promethium is using NVIDIA Quantum Cloud to rapidly generate training data on large, complex molecules, enabling better machine learning models that will help pharmaceutical companies find drug candidates more quickly."
Leaders in the pharmaceutical industry are already taking advantage of AI as a force multiplier in the drug discovery race. Not only is it helping them hone in on more promising candidates more quickly, but it's also creating a new benchmark for efficiency in R&D spend. Companies that are able to wield this technology effectively will be poised to achieve or better maintain industry dominance in the new age of drug discovery.
About QC Ware
Promethium is a quantum chemistry SaaS platform developed by QC Ware, a quantum and classical computing software and services company focused on delivering enterprise value through cutting-edge computational technology. With specialization in machine learning and chemistry simulation applications, QC Ware develops software for both near-term quantum and state-of-the-art classical computing hardware. QC Ware's team is composed of some of the industry's foremost experts in quantum and classical computing. QC Ware is headquartered in Palo Alto, California, and supports its European customers through its subsidiary in Paris and customers in Asia through its business development office in Tokyo, Japan.
Media Contact:Joseph.stamper@qcware.com 317-525-5882
View original content to download multimedia:https://www.prnewswire.com/news-releases/accelerating-the-development-of-new-molecules---promethium-to-empower-ml-models-for-drug-discovery-using-nvidia-quantum-cloud-302090961.html
SOURCE QC Ware Corp.
Application of machine learning algorithms for accurate determination of bilirubin level on in vitro engineered tissue … – Nature.com
Colour space channel sensitivity analysis
An absorbance spectral scan was performed on freshly prepared bilirubin samples using a microplate reader and compared against reported spectral profiles to ensure the absence of biliverdin. The solutions were shown to strongly absorb light in the blue colour region (Fig.1a), with an absorbance peak at 450nm29. We then sought to confirm the spectral behaviour of our prepared bilirubin samples by verifying the linearity of the optical density (OD) at three wavelengthsred (R, 650nm), green (G, 532nm) and blue (B, 456nm)with varying bilirubin concentrations. These three wavelengths also correspond to the RGB colour filters of the mobile phone digital camera sensor30. As shown in Fig.1b, a strong linear correlation (R2=0.996) and relatively higher sensitivity (slope=0.198) were observed in the blue wavelength, indicating that the blue wavelength has demonstrated the strongest bilirubin signal and it is more sensitive to the changes in the bilirubin concentration. Similarly, the green wavelength also displayed a strong correlation (R2=0.983), but a much lower sensitivity (slope=0.017) as compared to the blue wavelength. The red wavelength had the lowest correlation (R2=0.735) and sensitivity (slope=0.002) among all three wavelengths, suggesting the red channel might not respond sensitively to the changes in the bilirubin concentration.
Spectral characterization of bilirubin solutions with different concentrations. (a) Spectral characterization of bilirubin solution at different days. Red curve represents the freshly made bilirubin. A peak shift towards left has been observed for stored bilirubin due to conversion of bilirubin to biliverdin. (b) Scatter plot of absorbance of bilirubin solution in different colour wavelengths; Statistical differences (P<0.01) in both correlation and sensitivity were observed among all three wavelengths.
In vitro evaluation and characterization of the ML-based bilirubin measurement were conducted using an elastomeric tissue phantom mimicking neonatal skin with varying concentrations of bilirubin. Parametric studies of the various biological and external factors were conducted to evaluate the isolated effect of each biological and external factor through tight control of the in vitro fabrication parameters and environmental conditions. First, Polydimethylsiloxane-Titanium dioxide (PDMS-TiO2) tissue phantom samples with varying thicknesses were evaluated (Fig.2a). Relatively strong correlations between blue channel pixel value and different bilirubin concentrations were observed in all 1mm (R2=0.722, slope=2.775), 2mm (R2=0.921, slope=2.630) and 3mm (R2=0.972, slope=2.033) samples. However, a statistically significant difference (P<0.05) in sensitivity was observed between 1mm samples and 3mm samples, as well as between 2mm samples and 3mm samples, demonstrating that the images from the 3mm samples were less responsive to the changes in the bilirubin concentration.
Parametric study of important biological and external features. (a) Scatter plot of the pixel value of bilirubin concentrations in PMDS-TiO2 tissue phantom samples with different thicknesses; A statistically significant difference (P<0.05 in 1mm, P<0.001 in 2mm and P<0.001 in 3mm) was observed in the sensitivity slope but not in correlation among different thicknesses. (b) Scatter plot of the pixel value of bilirubin concentrations in samples with different light scattering ratios (PDMS to TiO2 ratio); Significant statistical difference was observed in both correlation and sensitivity between 0.01 and 0.02 PDMS-TiO2 ratio (P<0.005), as well as between 0.015 and 0.02 PDMS-TiO2 ratio (P<0.005). (c) Scatter plot of the pixel value of bilirubin concentrations in samples with different WB. Significant statistical difference was observed in both correlation and sensitivity among 2000K (P<0.05), 5000K (P<0.005) and 8000K (P<0.01). (d) Scatter plot of the pixel value of bilirubin concentrations in samples with different ISOs; ISO200 and ISO500 (P<0.05), ISO200 and ISO700 (P<0.05), as well as ISO500 and ISO700 (P<0.05) datasets are statistically different from each other in correlation. At the 0.05 significance level, a significant statistical difference was also observed in sensitivity between the ISO200 and ISO500 datasets (P<0.05), as well as between the ISO200 and ISO700 datasets (P<0.05). (e) Scatter plot of the pixel value of bilirubin concentrations in samples with different illumination tones; A significant statistical difference was observed in correlation among all channels (P<0.05). (f) Scatter plot of the pixel value of bilirubin concentrations in sample images with different light intensities; Different light intensities have demonstrated a considerable statistical significance in both sensitivity and correlation (P<0.05).
Next, images of PDMS-TiO2 tissue phantom samples with different extents of light scattering (varying ratios of PDMS to TiO2 light scattering agent) were used to represent skin containing varying amounts of scattering agents such as collagen and adipose tissue31. As shown in Fig.2b, strong correlations between blue channel pixel values and different bilirubin concentrations were observed in samples with a ratio of 0.01 (R2=0.921, slope=2.630) and samples with a ratio of 0.015 (R2=0.936, slope=2.359) respectively. However, no statistically significant linear regression relationship (R2=0.482, slope=1.065, P>0.05) was observed between blue channel pixel values and different bilirubin concentrations in PDMS-TiO2 tissue phantom images with a TiO2 ratio of 0.02.
Two hardware parameters that are commonly adjusted for image quality control were evaluated. Firstly, PDMS-TiO2 tissue phantom images were captured under different white balance (WB) conditions (Fig.2c). When WB is much lower (2000K) than the actual colour temperatures, weaker correlation and sensitivity (R2=0.774 slope=0.425) were observed, suggesting a strong confounding effect on the bilirubin concentration prediction. However, as the WB approached 5000K, which is close to the actual colour temperature in the image collection environment, a much stronger correlation and sensitivity (R2=0.963, slope=1.940) were observed in the scatter plot. As the WB increased higher (8000K), while the sensitivity remained unchanged (slope=2.097, P>0.05), a weaker correlation (R2=0.848) was observed. Secondly, camera sensor light sensitivity (ISO) was varied to establish its impact on the bilirubin estimation. Despite yielding moderately high sensitivities, it was observed that at low and high ISO values, a relatively weaker correlation was observed (ISO100, R2=0.898, slope=1.944; ISO1000, R2=0.868, slope=2.022) compared to the medium ISO (ISO500, R2=0.921, slope=2.630) in response to the changes in bilirubin concentrations. As expected in Fig.2d, the pixel intensity values increased with increasing ISO values. By extrapolation, it is expected that the pixel value of the diluted bilirubin can be saturated under brighter or darker lighting conditions. This suggests that an appropriate ISO setting needs to be selected to control the image brightness and prevent signal saturation at the lower and higher bilirubin.
The effect of ambient lighting conditions (light intensity and illumination tone) was also investigated (Fig.2e). Statistically significant differences in the correlation and sensitivities were calculated between the low, moderate and high light intensity conditions. It was observed that images with low light intensity generated a weak correlation (R2=0.717) but high sensitivity (slope=3.678). The correlation was observed to increase when the light intensity increased from moderate (R2=0.934) to high intensity (R2=0.942) while the sensitivity slope decreased from 2.531 to 1.793. These results suggest that high but appropriate amount of light intensity would aid the camera pixels in response to collect the colour information from the PDMS-TiO2 tissue phantom images.
Lastly, different illumination tones were supplied and tested using a diffused light source (Fig.2f). A relatively strong linear correlation was observed in all three illumination toneswhite light (R2=0.894, slope=2.106), off-white light (R2=0.860, slope=1.953) and yellow light (R2=0.878, slope=2.000). All three conditions demonstrated satisfactory (~0.9) correlation and sensitivity in response to the changes of bilirubin concentrations in the PDMS-TiO2 tissue phantom images. However, as compared to the pixel value in the images with white light, it is observed that the pixel values are generally lower in images with off-white light and are much lower in images with yellow light. This is expected as the spectral characterization (see Supplementary Fig.1) of illumination tones demonstrated the strongest blue light scattering signal (~450nm) in white light, followed by off-white light and yellow light.
The last factor tested (capture distance away from the subject) did not demonstrate any statistical significance in correlation and sensitivity (P>0.05), suggesting that distance does not affect the image information of the PDMS-TiO2 tissue phantom images (see Supplementary Fig.2).
We have shown that the bilirubin level prediction from images is strongly dependent on the colour representation. This necessitates the correction of the WB to achieve higher colour rendering. We applied different colour constancy algorithms (GW, MSGP, MaxRGB, and CH) and compared the WB-corrected images against the corresponding ground truth images captured with a light temperature of 5600K. The relative performances of the various WB correction methods are shown in Table 1. The different WB correction methods produce considerable differences, and the angular error results showed that the Gray World (GW) method performed the best at 3000K (mean angular error: 0.44) for the PDMS-TiO2 tissue phantom images. We also conducted additional tests to assess the GW accuracy in correcting raw images captured at different colour temperatures (Table 2). The results indicated that the GW method consistently produced a low angular error (<0.5) for all groups, demonstrating the stability and effectiveness of the GW method in correcting the WB variations of tissue phantom images.
Similar to the WB correction, we evaluated several colour spaces for their correlation with the bilirubin concentration using the tissue phantom images taken in a controlled image collection environment. As shown in Fig.3a, like the spectrophotometric characterization presented in Fig.1a, a strong linear correlation and high sensitivity (R2=0.958, slope=1.995) were observed between the blue channel pixel values and bilirubin concentrations in PDMS-TiO2 tissue phantom images. A relatively lower correlation and sensitivity were observed in the green channel pixel value (R2=0.871, slope=0.542), corresponding to a lower accuracy and sensitivity as compared to the blue channel. The pixel value acquired from the red channel did not demonstrate a statistically significant linear relationship (R2=0.014, slope=0.034, P<0.05) with the bilirubin solutions, suggesting an insensitive response to the changes in bilirubin concentration.
Linear regression and evaluation of important feature channels from colour spacesRGB, CMYK, L*a*b*, HSV, YCbCr and LUV. (a) Scatter plot of pixel values of bilirubin concentrations of PDMS-TiO2 tissue phantom samples in the RGB channels respectively; Significant statistical difference in correlation was observed between the B and R channel (P<0.05), as well as between the B and G channel (P<0.005). (b) Scatter plot of pixel values of bilirubin concentrations of samples in the CMY(K) channels respectively; A significant statistical difference in both correlation and sensitivity was observed among all channels (P<0.05). (c) Scatter plot of pixel values of bilirubin concentrations of samples in the CIELAB channels respectively. A statistically significant correlation (P<0.05) was observed in correlation among all channels. Sensitivity also demonstrated a significant statistical difference between the L channel and the a* channel, as well as between the L channel and the b* channel (P<0.05). (d) Scatter plot of pixel values of bilirubin concentrations of samples in the HSV channels respectively. A significant statistical difference in both sensitivity and correlation was observed among all channels (P<0.05). (e) Scatter plot of pixel values of bilirubin concentrations of samples in the YCbCr channels respectively. A significant statistical difference (P<0.05) was observed in both correlation and sensitivity among all different channels. (f) Scatter plot of pixel values of bilirubin concentrations of samples in the LUV channels respectively. A significant statistical difference (P<0.05) was observed in both correlation and sensitivity among all different channels.
Mapping of the RGB values to the CMY(K) colour space was also evaluated (Fig.3b). Among all channels, the pixel values obtained from the Y channel were observed to show the strongest linear correlation (R2=0.986) and the highest sensitivity (slope=1.812) whereas the C channel showed a weaker correlation (R2=0.782) and a less sensitive response (slope=1.086) to the changes in bilirubin concentrations. The M channel showed the weakest correlation (R2=0.122) with bilirubin level.
In the CIELAB (L*a*b*) colour space (Fig.3c), values acquired from the b* channel displayed a strong linear correlation (R2=0.971) and relatively high sensitivity (slope=0.902) to the changes in bilirubin concentrations. The value in the a* channel had a comparably weaker correlation (R2=0.749) and a lower sensitivity (slope=0.203). Notably, the pixel value obtained from the L channel did not show a statistically significant linear relationship (R2=0.3019, slope=0.08, P>0.05) with varying bilirubin concentrations.
Meanwhile, Fig.3d indicated the linear relationship between bilirubin level and the channels in Hue-Saturation-Value (HSV) colour space respectively. In the S channel, a strong linear correlation, which an R2 value is 0.917, was observed between the bilirubin concentration and the S channel value. Sensitivity was also observed to be relatively strong (slope=0.0061) as compared to other channels in the HSV colour space, showing strong capability in response to the change of bilirubin concentrations.
The YCbCr channel colour space demonstrated a relatively linear relationship between the 3 channels and the bilirubin concentration, albeit with an evidently lower sensitivity (Fig.3e). A very strong linear correlation and moderate sensitivity were observed in the Cb channel with an R2 value of 0.970 and a slope of 0.892. Compared to the other channels which demonstrated relatively less strong linear correlation and sensitivity (Y Channel: R2=0.769, slope=0.270; Cr Channel: R2=0.381, slope=0.275), the Cb channel displayed superior sensitivity to changes in bilirubin concentration.
The final colour space mapping evaluated was the CIELUV (L*u*v*) colour space (Fig.3f). As compared to the L (slope=0.080) and u* (slope=0.240) channels, which have either low or statistically insignificant sensitivity (P>0.05) to the changes of bilirubin value, the v* channel does not only have a strong linear correlation (R2=0.969), but also a relatively higher sensitivity (slope=1.308) to changes in bilirubin concentration.
Five machine learning modelsdecision tree (DT), K-nearest neighbour (KNN), random forest (RF), support vector machine (SVM), and LightGBMwere evaluated for their accuracy in performing binary classification of jaundice based on the PDMS-TiO2 tissue phantom image data (Fig.4a). The mean accuracy was 0.672 in DT, 0.737 in KNN, 0.774 in RF, 0.827 in LightGBM and 0.848 in SVM. Statistically significant differences (P<0.05) in accuracy were also observed among the models. The pairwise comparison also showed that among all models tested, SVM performed the best in the bilirubin binary classification task, followed by the LightGBM, RF and KNN, with DT performing the worst. The corresponding receiver-operating-characteristic (ROC) curves and the respective area-under-curve (AUC) scores further validated this observation (Fig.4b). The AUC scores obtained for each model were as follows: 0.74 (DT), 0.82 (KNN), 0.86 (RF), 0.91 (LightGBM), and 0.93 (SVM). The results were consistent with the performance comparison based on the cross-validated accuracy, suggesting that the SVM model has the best predictive capability among all other models tested.
Model performances. (a) Accuracy performance among DT, KNN, RF, SVM and LightGBM models in the classification task; A significant statistical difference in accuracy was observed among models (P<0.05). All models demonstrated a significant statistical difference in pairwise comparison except for the comparison between SVM and LightGBM model (P>0.05). (b) ROC performance of the five different models. The inlet graph shows the AUC performance, which represents the capability of the model to distinguish between the tissue phantom images with normal bilirubin levels and those images with abnormal bilirubin concentrations. The AUC demonstrated a significant statistical difference among all models (P<0.05). (c) R2 value among different models in the regression task with a different number of features as training labels. A significant statistical difference (P<0.05) was observed in R2 between 6 and 17 features among all models. Five models are statistically different from each other except the RF, SVM and LightGBM in pairwise comparison. (d) MSE value among different models in the regression task with a different number of features as training labels. A significant statistical difference (P<0.05) was observed in MSE between 6 and 17 features among all models. Asterisk (*) indicates P<0.05.
In addition, as shown in Fig.4c, d, we also tested the model capability in performing the regression task, evaluating the performance of each regression model (DT, RF, KNN, SVR, and LightGBM) in predicting the exact bilirubin concentration in the PDMS-TiO2 tissue phantoms. When the models were trained with limited features, a relatively large mean square error, MSE, was observed in all models (DT: 16.157; KNN: 19.749; RF: 11.499; SVM: 17.651 and LightGBM: 13.830). A low R2 score (DT: 0.555; KNN: 0.465; RF: 0.684; SVM: 0.524 and LightGBM: 0.624) was observed at the same time. However, the results showed significant improvements (P<0.05) in the model performances (R2 score) when additional features were included for all five models (DT: 0.734; KNN: 0.742; RF: 0.812; SVM: 0.806 and LightGBM: 0.808). The corresponding MSE was also greatly reduced across all models (DT: 9.635; KNN: 9.366; RF: 6.816; SVM: 7.054 and LightGBM: 6.946). Similar to the classification task, LightGBM, SVM and RF models performed the best in predicting bilirubin concentrations. The acquired low MSE value, akin to the square of bilirubin level variance, additionally indicates a minor disparity in relation to the true bilirubin levels. A variance ranging from2.61mg/dl to3.10mg/dl was achieved across all evaluated models, thus illustrating the predictive capacity of machine learning models in determining precise bilirubin levels.
Follow this link:
Application of machine learning algorithms for accurate determination of bilirubin level on in vitro engineered tissue ... - Nature.com
Machine Learning Predicts Lung Cancer Spread to Brain – Technology Networks
Physicians treating patients with early-stage lung cancer face a conundrum: choosing potentially helpful yet toxic therapies such as chemotherapy, radiation or immunotherapy to knock out the cancer and lessen the risk of it spreading to the brain, or waiting to see if lung surgery alone proves sufficient. When up to 70% of such patients do not experience brain metastasis the spread of cancer to the brain the question arises: Who should receive additional aggressive treatments, and who can safely wait?
A new study led by Washington University School of Medicine in St. Louis could help physicians strike the right balance between proactive intervention and cautious monitoring for patients with early-stage lung cancer. The study, published March 4 in The Journal of Pathology, uses an artificial intelligence (AI) method to study patients lung biopsy images and predict whether the cancer will spread to the brain.
Subscribe to Technology Networks daily newsletter, delivering breaking science news straight to your inbox every day.
There are no predictive tools available to help physicians when treating patients with lung cancer, saidRichard J. Cote, MD, the Edward Mallinckrodt Professor and head of theDepartment of Pathology & Immunology. We have risk predictors that tell us which population is more likely to progress to more advanced stages, but we lack the ability to predict individual patient outcomes. Our study is an indication that AI methods may be able to make meaningful predictions that are specific and sensitive enough to impact patient management.
Lung cancer is the leading cause of cancer death in the U.S. and worldwide. Most lung cancers are characterized as non-small cell lung cancers, which are largely, but not exclusively, caused by smoking. For early-stage cancer patients, tumors are confined to the lung, and surgery is recommended as a first line of treatment. Roughly 30% of such patients progress to advanced stages, when the cancer spreads to the lymph nodes and other organs. With the brain often affected first, such patients require additional treatments, including chemotherapy, targeted drug therapy, radiation therapy and/or immunotherapy. However, physicians have no way of knowing whose cancer will progress, so they frequently treat patients with aggressive therapies out of caution.
Cote worked withRamaswamy Govindan, MD, the Anheuser Busch Endowed Chair in Medical Oncology and associate director of the oncology division at Washington University;Mark Watson, MD, PhD, the Margaret Gladys Smith Professor in the Department of Pathology & Immunology; and Changhuei Yang, PhD, a professor of electrical engineering, bioengineering, and medical engineering at the California Institute of Technology, to determine if AI could predict whether cancer will spread to the brain.
In diagnostic testing, a pathologist examines biopsied tissues under a microscope to identify cellular abnormalities that may hint at disease. Advanced technologies such as AI are being explored to replicate what a pathologist sees when making diagnoses but with greater accuracy, Cote explained.
A key question: Can AI detect abnormal features that a pathologist cannot?
The researchers trained a machine-learning algorithm to predict brain metastasis using 118 lung biopsy samples from early-stage non-small cell lung cancer patients. Some of the patients developed brain cancer during a five-year monitoring period, and some did not and were in remission. Then the researchers tested the AI method on its ability to predict brain metastasis, and to identify patients who develop no metastasis, using 40 other patients lung biopsy samples.
The algorithm was able to predict the eventual development of brain cancer with 87% accuracy. In comparison, four pathologists who participated in the study were an average of 57.3% accurate. Importantly, the algorithm was highly accurate in predicting which patients would not develop brain metastasis.
Our results need to be validated in a larger study, but we think there is great potential for AI to make accurate predictions and impact care decisions, said Govindan, who treats lung cancer patients atSiteman Cancer Center, based at Barnes-Jewish Hospital and Washington University School of Medicine. Systemic treatments such as chemotherapy, while effective in killing cancer cells, can also harm healthy cells and are not always the preferred treatment method for all early-stage cancer patients. Identification of patients who are likely to relapse in the brain may help us develop strategies to intercept cancer early in the process of metastasis. We think AI-based predictions could, one day, inform personalized treatments.
The AI system evaluates tumors and healthy cells features, similar to how the human brain allows us to scan facial features for quick recognition of familiar faces. However, what the algorithm sees is unknown; the scientists are working to understand the molecular and cellular features that AI uses for its predictions. This knowledge could lead to the development of novel therapeutics and influence the design of imaging instruments optimized for the collection of data for AI.
This study started as an attempt to find predictive biomarkers, said Yang. But we couldnt find any. Instead, we found that AI has the potential to make predictions about cancer progression using biopsy samples that are already being collected for diagnosis. If we can get to a prediction accuracy that will allow us to use this algorithm clinically and not have to resort to expensive biomarkers, we are talking about significant ramifications in cost-effectiveness.
Reference:Zhou H, Watson M, Bernadt CT, et al. AI guided histopathology predicts brain metastasis in lung cancer patients. J Pathol. 2024:path.6263. doi:10.1002/path.6263
This article has been republished from the following materials. Note: material may have been edited for length and content. For further information, please contact the cited source.
See more here:
Machine Learning Predicts Lung Cancer Spread to Brain - Technology Networks
Revolutionising Textile Manufacturing: AI, Machine Learning and Sustainability – Innovation in Textiles
ITMA, the worlds largest international textile and garment technology exhibition, has launched a new series of webinars on its ITMAconnect online sourcing platform and knowledge hub. Called the Innovator Xchange Virtual Edition, the series features five webinars to be held between 25 January and 22 April 2024 exclusively for ITMAconnect users.
Each webinar focuses on the latest industry developments, drawing reference from the Innovator Xchange discussions held during ITMA 2023 in Milan. Through panel discussions, industry experts will share their insights and engage participants with questions and answers.
The third webinar in the series is:
Revolutionising Textile Manufacturing: AI, Machine Learning and Sustainability
Date & Time: 21 March 2024, from 10.00am - 10.45am CET.
Explore the future of textile manufacturing in this groundbreaking webinar on revolutionising the textile industry through AI, Machine Learning and Sustainability. Hear from the leading experts in the field how the fusion of cutting-edge technology and environmental responsibility would transform the world of textiles.
This webinar will inspire, educate and empower you to lead the charge towards a more sustainable, smarter textile industry.
Panel Discussion Highlights:
- Examine the role of collaborative robots within the textile machinery sector and the revolutionary potential of AI in optimising textile sorting for recyclability.
- Learn how AI and Machine Learning is reshaping every aspect of textile production, from design to delivery, driving not only efficiency and productivity, but also paving the way for a more sustainable future.
- Discover how thought leaders and industry pioneers adopt the core principles of sustainability to drive innovation and profitability in textile manufacturing
Moderator
Dr. Andre West, Associate Professor and Director of The Zeis Textiles Extension (ZTE), North Carolina State University College of Textiles, United States
Panel Members
- Mr Florian Pohlmeyer
Head of Digitalisation
ITA Institut fr Textiltechnik der RWTH Aachen University
- Prof Calvin Wong
CEO & Centre Director
Laboratory for Artificial Intelligence in Design (AiDLab)
- Mr Sverker Evefall
Senior Application Manager
ACG EyeTech (A part of ACG Group)
Find out more about the webinar
Register for the webinar now!
Check out other webinars in the series
Read the original:
Revolutionising Textile Manufacturing: AI, Machine Learning and Sustainability - Innovation in Textiles