AI for Drug Discovery

Problem

Testing a compound in the lab is slow and expensive. If a model can rank which compounds are worth testing first, the same budget covers more ground. The target here was bioactivity prediction for colon-cancer compounds.

This was my first research internship, at GGSIPU in May and June 2024. I led the machine learning work for a team of five.

Approach

A molecule is a graph, not a table row, so it needs turning into numbers first. RDKit fingerprints do that by encoding which substructures are present as a fixed-length vector.

On those features I trained two model families, a Random Forest and an artificial neural network, and compared them rather than assuming the neural network would win. On fingerprint features with a dataset this size, that assumption is often wrong.

Results

The work ran end to end: featurisation, training, comparison, and a recommendation the rest of the team could act on.

It was also the first time a model of mine had to answer to something outside itself. A loss curve going down is a claim about the optimiser. Whether a compound is worth a lab’s time is a claim about chemistry.

Limits, and what I would do next

Fingerprint features throw away 3D structure, which matters for how a molecule actually binds.

The split is the part I would redo. A random split lets near-identical compounds appear in both training and test, which inflates the score. A scaffold split, which keeps whole structural families on one side, gives a much less flattering and much more honest number.