ToxVision: 1,000-molecule measured-assay validation
==================================================
Exactly 1,000 distinct stereo-free fragment parent structures, 10,623 observed
molecule-assay labels: 598 NIH final evaluation, 285 NIH leaderboard test,
117 recorded held-out test compounds. External-only results (883) are separate.
These are retrospective experimental assay labels, not clinical safety outcomes.
Frozen GCN/GINE/GATv2 weights were not retrained. Two fingerprint baselines
were trained only on the recorded training partition. Missing labels are masked.

This archive preserves the executed scripts, raw source inputs, split manifests,
checkpoint weights, predictions, reference labels, per-task/cohort metrics,
bootstrap summaries and deployed-service checks. The PDF explains limitations.

Reproduction: install the web project's dependencies plus scikit-learn, scipy,
joblib, matplotlib, reportlab, requests and pymupdf. The bundled script copies
use the archive's repository folder; the evaluated application remains the
Desktop clone. Run from the extracted archive root:
  python work/redesign/validate_1000.py
  python work/redesign/charts_1000.py
The API check requires the application running at http://127.0.0.1:7860:
  python work/redesign/test_deployed.py
CPU timings depend on hardware and load. The report's recorded timings are
local sequential requests, not a load-test or production throughput promise.

Primary data: https://tripod.nih.gov/tox21/challenge/data.jsp
Project: https://github.com/eliot-99/Graph-ML-Enabled-Molecular-Design-Assistant-using-Graph-Neural-Networks
