First Ascent by
Nathan Tran
The Molecular Glue Problem
Medicinal Chemistry / Targeted Protein Degradation
The Proposed Route
A route through the emerging landscape of molecular glue degraders (MGDs) — small molecules that hijack the cell's protein disposal machinery. The climber navigates ~100 compounds from their own research, comparing Ridge regression on Morgan fingerprints to Chemprop (a message-passing neural network), predicting degradation efficiency (pDC50, Dmax) against both the target protein (ZBTB11) and an off-target (IKZF1). The route asks: can ML help design selective degraders?
🧗 The Crux
N≈100 is small for Chemprop — the MPNN may not have enough data to outperform simpler fingerprint methods. Noisy or incomplete pDC50 values could poison the model. The dual-target setup (ZBTB11 vs IKZF1) is the scientific hook but doubles the modeling complexity.
⚠️ Pre-Climb Checklist
✅ Dataset is from own research — real, relevant, and available. ✅ Chemprop v2 with Optuna hyperparameter tuning is well-specified. ⚠️ With ~100 compounds, expect Chemprop to struggle — that's a valid finding. ⚠️ AlphaFold3 stretch goal (ternary complex modeling) is ambitious — complete core pipeline first.
Guidance
- The selectivity question (ZBTB11 vs IKZF1) is the real story — lead with it
- If Ridge beats Chemprop, that's interesting and publishable for small-data regimes
- AF3 interface scores as features is a clever idea — but validate on a few examples first
Source proposal: Tran_Nathan_finalProject_proposal.docx
CHEM 169/269 · Applied AI & Machine Learning for Biochemistry