Cloud Hosting

Deep learning approaches have potential to substantially reduce the astronomical costs and long timescales involved in drug discovery. KarmaDock proposes a deep learning workflow for ligand docking that shows improved performance against both benchmark cases and in a real-world virtual screening experiment.

Drug discovery is a long and arduous process that is staggeringly expensive the average estimated time needed to take a new drug from discovery to launch is 1012 years1, at a high cost of ~US$2.2 billion per drug2, which is a major problem considering that this process is also plagued by low hit rates. Computer-aided drug discovery (CADD) can substantially aid this process3, including both by predicting how a range of drug-like ligands would bind to a given drug-target (virtual screening) using docking algorithms, as well as predicting the corresponding binding free energies of the docking predicted poses, which are a measure of the strength with which the ligand binds to its target. However, despite significant progress in this area, challenges remain, including (1) the quality of the binding poses predicted, which is crucial for rational drug discovery, a process that is complicated by the presence of error, non-linearity, and randomness4; (2) the precision and accuracy of the predicted binding free energies for those poses there can be, for instance, significant variation in the pose ranking for the same ligand/target combination between docking approaches and (3) the speed of the approach, which is particularly an issue in the face of increasing library sizes. That is, computational approaches need to be both efficient enough to be able to perform ultra-large docking on libraries that can reach billions of compounds5, without significantly compromising the quality of the binding pose and free energy predictions. Such huge libraries are out of the scope of conventional CADD approaches, but are an ideal target for deep-learning (DL) approaches5, which typically perform better than traditional shallow machine learning techniques (or even deep learning approaches with expert descriptors) when processing large data sets6. However, even DL approaches face challenges optimizing both accuracy and computational speed, due to the inherent complexity of the problem, as well as the degree of seeming randomness involved4. Writing in Nature Computational Science Xujun Zhang and colleagues7 propose KarmaDock, a DL approach for ligand docking, showing both improved speed and accuracy compared to benchmark data sets, as well as performing well in a real-world virtual screening project where it was used to discover experimentally validated active inhibitors of LTK, which is a target for the treatment of non-small-cell lung cancer8.

Read the rest here:
Progress in using deep learning to treat cancer - Nature.com