README.md

---
license: apache-2.0
tags:
  - pharmacogenomics
  - CYP2C9
  - variant-classification
  - MAVE
  - AlphaMissense
---

# CYP2C9 Variant Function Classifier (research artifact — honest negative)

Experimental classifier for CYP2C9 variant functional classification
(no_function / decreased_function / normal_function), built by
[Anukriti AI](https://anukritiai.com).

> **Read this first.** This repository is published as a transparent
> **negative result**. The v2 model fixes the circularity of v1 but still
> fails the held-out clinical test (1/6 correct). It is **not** a clinical
> predictor and must not be used for dosing decisions. It is shared so the
> finding — that single-assay MAVE labels do not generalize to CPIC clinical
> phenotype for CYP2C9 — is reproducible.

## Versions

- **v1** — MAVE-threshold scaffold. 8,050 training rows. 5-fold CV accuracy
  0.996 (XGB) but **circular**: the `click_score` / `vamp_score` features the
  labels were thresholded from drive ~77% of feature importance. Leave-anchors-out:
  4/4 CPIC anchors misclassified without 500× upweighting. A MAVE-threshold
  reproducer, not a clinical predictor.
- **v2** — non-circular. `click_score` / `vamp_score` removed; AlphaMissense
  (genomic-coordinate-corrected) + CADD added. Trained on the 2,514-row
  SNV-reachable subset. 5-fold CV AUC ~0.88 (XGB 0.886) — believable, not
  hollow. **Held-out clinical test: 1/6 = 17%** (only `*11` predicted
  correctly).

## The finding

Removing the circular features fixed the inflated CV score, but the model still
fails clinically because it was trained on **MAVE-threshold labels**, and MAVE
assay function ≠ CPIC clinical function for CYP2C9. **The bottleneck is the
label definition — not feature quality or model architecture.**

AlphaMissense is discriminative where available (monotonic class separation:
normal 0.21 → decreased 0.44 → no_function 0.65 mean), but covers only **31.3%**
of this codon-saturation MAVE library because 67.5% of variants require
multi-nucleotide AA changes that AlphaMissense cannot score by design. Coverage
is the blocker, not feature quality.

## Ground truth / sources

MaveDB (Click-seq + VAMP-seq CYP2C9 libraries), CPIC CYP2C9 allele-function
table, PharmVar, Ensembl VEP / AlphaMissense, CADD.

## Citation

Part of the Anukriti AI platform validation effort.
Project-level preprint: https://doi.org/10.5281/zenodo.20727790
(This DOI covers the broader Anukriti validation study, not a CYP2C9-specific
artifact.)