PXR Induction Challenge

Repository for my OpenADMET PXR Blind Challenge work (April 1 - July 1, 2026).

• Challenge page: https://huggingface.co/spaces/openadmet/pxr-challenge
• Track 1 public model report: https://github.com/N283T/openadmet-pxr-model-report
• Track 1 Phase 1 research log: issue #100
• Track 1 Phase 2 research log: issue #208

The issues are intentionally part of the public record: they contain the day-by-day experiment notes, wrong turns, leaderboard reactions, null results, and decision logs that do not belong in a polished README. If you want the battle record rather than the cleaned-up code map, start with #100 and #208. Some issue comments are informal notes in Japanese.

Final Result

Track 1 Activity: predict pEC50 for 513 blinded compounds.

What Is Here

This repository keeps the reproducible Track 1 code, schemas, lightweight documentation, public leaderboard snapshots, and a curated lightweight data bundle. Large embeddings, Boltz outputs, model weights, private reports, and most generated submissions remain outside git.

data/                         curated source/derived tables; large artifacts ignored
db/                           database schemas, loaders, feature builders
docs/                         documentation, literature notes, leaderboard snapshots
docs/leaderboards/activity/   timestamped Track 1 leaderboard snapshots
track1_activity/src/          shared Track 1 loading, features, splits, metrics
track1_activity/scripts/      training, ensembling, calibration, diagnostics
track1_activity/boltz2/       Boltz-2 feature-generation pipeline for Track 1
track1_activity/submissions/  ignored Track 1 CSV submissions
structures/                   ignored Boltz-2 / structure runtime artifacts

Track 1 Notes

The final Track 1 system was an ensemble workflow built around deterministic UMAP-split validation, feature families from 2D descriptors and foundation-model embeddings, Boltz-2-derived features, Caruana-style ensemble selection, and submission preflight checks against trusted anchors.

Useful context:

• Issue #100 is the Phase 1/live-leaderboard chronology.
• Issue #208 is the Phase 2 answer-check and finalization log.
• docs/track1_explain/ keeps compact public explanation and audit notes.
• docs/leaderboards/activity/ keeps the public leaderboard snapshots used for retrospective checks.
• AGENTS.md is the durable operating guide for coding agents in this repo.

Environment

This project uses pixi and a local PostgreSQL + RDKit database.

Useful commands:

pixi install
pixi run db-start
pixi run db-stop
pixi run db-psql

pixi run ruff format <file>
pixi run ruff check <file>

Database details used in local scripts:

• PostgreSQL 18 + RDKit cartridge
• port: 5433
• socket: /tmp

GPU-heavy work was developed on an RTX 5080 with WSL2 CUDA override CONDA_OVERRIDE_CUDA=13.1.

Data

Challenge data comes from openadmet/pxr-challenge-train-test. Fresh clone setup is intentionally script-driven through the database and data loading scripts under db/.

The small Git-tracked subset is documented in data/README.md. It includes the official activity datasets, RDKit descriptors, compact model predictions, and presentation-oriented analysis tables. Embeddings and structure artifacts remain local-only.

Local submission clients can contain personal account state, so they are ignored by git. Recreate them locally rather than committing credentials or API state.