Overview / Usage

Active topics are 1) predicting drug binding sites in target proteins, 2) developing ML-based scoring functions to identify bioactive molecules in virtual screening, and 3) using generative adversarial network methods to discover bioactive molecules with novel chemotypes. We are focusing on drugs for cancer, immunology, and malaria.

We recently placed 4th in the Drug Design Data Resource Grand Challenge (https://drugdesigndata.org/) to predict the binding activity of a set of molecules to the Alzheimer's Disease drug target, BACE1. Our strategy used a neural network trained on chemical data from closely related BACE1 inhibitors.

Methodology / Approach

We are focusing on several different approaches in our projects.

1. Training ML-based scoring functions on target-specific data for docking and virtual screening. We treat docking (predicting the correct binding pose) as a separate problem from predicting binding affinity. Scoring functions must be fast to compute to be useful for virtual screening.

2. Training ML-based scoring functions for fragment screening. Fragments tend to be smaller and more soluble than traditional drug-like molecules. We are developing scoring functions specific for their chemical properties.

3. Using structure conservation data, docking, and fragment screening to predict alternative drug binding sites in protein targets. This approach can be useful to obtain better target specificity, synergistic effects with other drug treatment, and identification of drug molecules with novel chemical scaffolds.

4. Developing GANs to generate novel drug-like molecules. By including new cost functions, one can in principle train GANs to focus or avoid certain chemical scaffolds, propose molecules with lower toxicity, and explore new chemical space.

Technologies Used

Intel python, numpy, scipy, scikit-learn, pytorch, tensorflow, keras, intel math kernel library, intel compilers, intel CPUs

Other links

• http://holeungng.com