Harnessing molecular docking and machine learning to target proteins via small molecule compounds for cancer therapy

Predicting whether a particular protein and drug interact is an important but largely unresolved goal. We present a rapid and computationally inexpensive approach to determine whether proteins and a compound bind each other. Herein, we train a machine learning algorithm that distinguishes between docking results from true protein-drug pairs and those from pairs that do not interact. The features used for training include structural and biophysical features of specific poses as well as features derived from the distribution of docking results across all proposed binding modes for a given protein-drug pair. It is worth noting that we do not perform explicit energy calculations at any stage of our calculations. Using this approach, we identified true protein-drug interactions from a pool of 488 true and 194,770 putative false complexes with an accuracy of 0.6 (i.e., 60% of the predicted interactions were correct) and a recognition of 0.2. This is >500 times better than the random value and >30 times better than the accuracy that would be achieved using only the docking score of the best pose. By applying this method to a large dataset of proteins and FDA-approved drugs, we were able to identify potential protein-drug interactions. Finally, we experimentally confirmed this by identifying a small molecule to bind and degrade the Wiskott Aldrich Syndrome Protein (WASp) known to play a vital role in hematological malignancies.