RDKit + Molecule Library Optimization - Improve an existing algo

Improve an Existing Search Algorithm for Molecule Selection (ML/Python) We have a working drug-discovery pipeline with a reference search algorithm that we want improved. What the pipeline does: Every day we are given: - 1 target protein (we want molecules that bind it) - 2 antitarget proteins (molecules must avoid binding these) Our code has 15 minutes to search a large library of chemical building blocks (millions of combinations), assemble 100 valid drug-like molecules, and output them. Each molecule is scored by an internal AI protein-ligand model: molecule score = target binding − 0.9 × antitarget binding The final score is the average over the 100 molecules. There are strict validity rules: minimum heavy atoms, limited rotatable bonds, no duplicates, and a required level of chemical diversity. What we already have: - A strong reference algorithm (genetic/evolutionary search) that we want to beat - An exact local evaluation harness (same scoring, same 15-minute budget) on a GPU machine - A/B testing setup so any improvement can be proven with controlled experiments Your job: - Understand the existing algorithm and reproduce its baseline scores - Design and implement improvements so your algorithm outscores the reference by a small but consistent margin (3–5%) across many different protein targets ( 4 protein targets at least ) - Prove every improvement with controlled A/B validation on our harness Ideal candidate: strong Python + RDKit experience, understands evolutionary algorithms or Bayesian optimization, knows how to squeeze performance under a hard time budget, and validates ideas with rigorous experiments rather than guessing. Deliverables: improved search algorithm + reproducible validation showing it beats the baseline on ≥8 internal test targets.