Projects in AI method development for hit-identification algorithm

Hit-ID
To improve the existing deep-learning based regression model, So I started by curating some new benchmarks to evaluate the model to assess their failure modes: I curated, two new benchmarks- allosteric inhibitors and activity cliffs. From this I showed that the model’s performance degraded for novel binding pockets as well as on the pair of molecules having non-linear structure- activity relationship. To make upgrades to the model, I firstly re-wrote and re-organized the code base to make it more modular, amenable to test and debug and deployable to the cloud infrastructure.

Property Prediction
Besides the hit-id effort, I’ve developed an algorithm that predicts various properties of the drug using transfer learning techniques. For this I trained graph-based transformer pretrains, language based transformer, and simple neural network applied to morgan fingerprint.