The process of finding a new drug candidate against a chosen target related to a particular disease usually involves hit identification, hit-to-lead, and lead optimization. During this process, numerous chemicals (and peptides) are assessed to various drug-like characteristics such as high affinity towards the target, low cytotoxicity, high water solubility, lipophilicity, etc. Especially, we cannot overemphasize the importance of rigorous measure of binding affinity and toxicity, because uncertainty in the properties can cause ~57% and ~30% of drug failures from phase 1 or 2, respectively. To satisfy accurate evaluation on these quantities, Aimble builds a deep-learning-based high-performance platform, consisting of Aimble’s own database and three self-developed solutions such as high-throughput docking solution, binding prediction solution, and toxicity prediction solution. Aimble platform will continue to be developed in order to satisfy unmet needs in drug discovery. In parallel, the high-quality database, generated from quantum calculation and molecular dynamics simulation, will continue to be accumulated in order to improve accuracy of prediction of Aimble’s deep-learning approach.
High-throughput Docking Solution
Aimble’s high-throughput docking solution is an automated computing tool to predict bound conformations and binding free energies for large libraries of chemicals to the target as well as to sort the chemicals by their free energies in order to successfully identify validated hit series. Difficulty in this approach usually comes from quite a bit of computational time required to test with huge amount of chemicals. Aimble’s docking solution overcomes the difficulty with efficient usage of distributed computer system. In other words, for a given target, several preprocessed chemical compounds are simultaneously arranged for docking calculations and results are automatically saved for analysis.
Binding Affinity Prediction Solution
For the accurate prediction of binding free energies between ligand molecules and target proteins, we train our machine learning (ML) models with the data on the atomic structure of ligand-protein complexes and corresponding binding affinities. Furthermore, we incorporate physics-based domain knowledge into our ML model, which hence can offer a more reliable prediction.
Toxicity Prediction Solution
Over 30% of new drug candidates fail at (pre-) clinical phase due to their toxic effects. We have developed in-house toxicity prediction technology by combining cumulative data-preprocessing skills and multi-task-trained AI machines. Specifically, a chemical compound is virtually fragmented into smaller parts and they are more elaborately encoded in the form of molecular descriptors which are then feeded to our neural network. A toxicophore can be identified via cooperative roles of these descriptors with the help of multi-criteria (hence more reasonable) decision of our neural machine.