AmplusFold — Notes

I'm building AmplusFold, a model that predicts how well small molecules bind to proteins relevant to pesticide discovery — insect, plant, and fungal targets.

This repo has a few notes I've written along the way. The code, data, and results live in a private repo. What you'll find here is the thinking: what the field looks like, what makes wet-lab data trustworthy, how to choose what to test next, and the common ways these projects go wrong.

Early models score around Pearson R = 0.35 on a held-out set of pesticide-relevant measurements, under a strict split where no protein in the test set was seen during training. That's enough to be useful for narrowing down candidate molecules before sending them to a lab, but not enough to rely on alone. The next step is getting real wet-lab data on the actual targets we care about.

 public databases → clean + standardize → Postgres → pre-trained
 (BindingDB, (canonical SMILES, (one row embeddings
 ChEMBL, unit conversion, per (ESM2 for proteins,
 PubChem) dedupe) measurement) ChemBERTa for molecules)
 │
 ▼
 train a model to
 predict binding
 affinity

A few notes on the choices:

Why pre-trained embeddings. ESM2 was trained on tens of millions of protein sequences; ChemBERTa was trained on tens of millions of molecules. The public binding-affinity data is too small to learn good representations from scratch, so it pays to start from theirs and only train the part that predicts affinity.

Three model variants, in order of complexity.

1. Random forest on simple chemistry and protein features. The cheapest model that should work. If a fancier model can't beat this, something is wrong with the fancier model, not the data.
2. A small neural network on top of the ESM2 and ChemBERTa embeddings. Tests whether learned representations help over hand-engineered ones.
3. Same as #2, but with an explicit "compound feature ×ばつ protein feature" interaction term. Tests whether the model benefits from being shown where to look for compound-protein interactions, instead of having to discover that pattern on its own.

The point of the progression isn't to crown a single best model — it's to figure out what kind of capability actually helps on this problem, so the next round of effort goes to the right place.

• docs/01_domain_primer_dti.md — what drug-target interaction modeling is, and what the standard binding measurements (Kd, Ki, IC50) actually mean
• docs/04_assay_quality.md — how wet-lab measurements work, and why their noise sets a hard ceiling on any model trained on them
• docs/05_active_learning.md — how to pick which molecules to test next when lab time is limited
• docs/06_failure_modes.md — ten ways these projects commonly fail, most frequent first
• docs/07_dti_method_landscape.md — a survey of how people have built these models, from simple to complex
• docs/09_foundation_models_explained.md — what ESM2 and ChemBERTa actually do, in plain language

Python, PostgreSQL, RDKit, PyTorch, ESM2, ChemBERTa.