pathlaw 0.1.0

Supervised Pathway Projection + Symbolic Regression for Interpretable Drug Response Prediction

pathlaw

Supervised Pathway Projection + Symbolic Regression for Interpretable Drug Response Prediction

pathlaw compresses a deep drug response model's latent representation into biologically named pathway scores, then fits a symbolic regression head — producing a single, globally readable formula:

lnIC50 ≈ 2.51 × EML(x, y) − 2.09
  x (resistance) = +1.56·KRAS  +1.12·DNA_REPAIR  +1.04·G2M  + ···
  y (sensitivity) = +0.87·APOPTOSIS  +0.70·PI3K_AKT  + ···

No post-hoc aggregation. No direction analysis. No manual interpretation.
The formula is the explanation.

Install

pip install pathlaw

Quick start

from pathlaw import PathLaw

# X: intermediate representations from your model  (N, n_input)
# y: target values, e.g. lnIC50                    (N,)
model = PathLaw(n_input=1536, n_pathways=8, depth=3)
model.fit(X, y, epochs=300)

print(model.formula())   # human-readable formula
print(model.explain())   # clinical sentences from coefficients

How it works

Your model's latent (1536-dim)
        ↓
Pathway Projection  Linear(1536 → 8)   ← trainable, named after Hallmark gene sets
        ↓
EML Symbolic Head   depth-3 tree       ← learns lnIC50 ≈ EML(x, y)
        ↓
Formula + clinical sentences           ← 0 analyst steps

The 8 default pathways (HALLMARK_PATHWAYS): APOPTOSIS, G2M_CHECKPOINT, DNA_REPAIR, PI3K_AKT_MTOR, KRAS_SIGNALING, ESTROGEN_RESPONSE, HYPOXIA, E2F_TARGETS

You can supply any pathway names matching your domain.

API

Method	Description
PathLaw(n_input, n_pathways, depth, pathway_names)	Create model
model.fit(X, y, epochs, ...)	Train on intermediate reps + targets
model.predict(X)	Predict lnIC50
model.pathway_scores(X)	Get named pathway activations
model.formula()	Human-readable formula string
model.explain(threshold)	Clinical sentences from coefficients
model.formula_dict()	Full formula as dict (includes LaTeX)
model.evaluate(X, y)	Pearson r, Spearman ρ, RMSE
model.save(path) / PathLaw.load(path)	Persistence

Citation

If you use pathlaw in your research, please cite:

@article{hong2026pathlaw,
  author  = {Hong, Seokchol},
  title   = {Supervised Pathway Projection with Symbolic Regression
             Enables Interpretable Drug Response Prediction},
  year    = {2026},
  url     = {https://github.com/seokcholhong/pathlaw}
}

License

MIT