YAML Metadata Warning:empty or missing yaml metadata in repo card
Check out the documentation for more information.
DPP-4 pIC50 LTN β DEPLOYMENT REFIT (checkpointed variant)
DEPLOYMENT REFIT β trained on ALL labeled data including the regulatory set (absorbed 2026-11-14 by Hossain's decision). Benchmark metrics do NOT apply to this version; it has not been and must not be evaluated on the absorbed sets. The benchmark version remains Delower/dpp4-pic50-ltn.
What this is
The tuned b256 LTN recipe (lr 1e-4, hidden [768, 384, 192], p=2, batch
256, fingerprint-4608 features) refit on ALL labeled data: 10,085
curated rows + the 11 FDA gliptins with experimental pIC50 values from
the regulatory master table = 10,096 rows (gosogliptin excluded β
no experimental IC50). Hard gate: PASS β all 11 regulatory compounds
verified present in the actual training file by full InChIKey
(gate_verification.json, re-run for this variant).
The model is a regressor (raw pIC50); the derived classification
(potent β₯ 7.0, moderate β₯ 5.0, weak < 5.0) is reported alongside the
regression output by infer.py, same convention as the benchmark repo.
Honest disclosure of the variant
Two prior bounded attempts at this refit diverged: attempt 1 (seed 42) produced a NaN loss at epoch 1; attempt 2 (seed 1234) trained stably for 18 epochs (in-training monitor RMSE improving to 0.3377), then produced NaN at epoch 19 β and its snapshots were held in memory, so they were lost.
This run (Hossain-approved variant, 2026-11-14) differs from attempt 2 only in snapshotting: the weights were written to disk after every epoch with finite loss and monitor RMSE. Recipe, training file, monitor indices, loop mechanics (including the patience-12 early stop) and the seed (1234) are identical.
This run did not reproduce the epoch-19 NaN β TensorFlow CPU execution is not bit-deterministic, so the trajectory need not match attempt 2 exactly. It trained to a patience early stop at epoch 87, with a snapshot saved at every one of the 87 epochs.
Shipped weights
- Shipped snapshot: epoch 75, the best finite snapshot by in-training monitor RMSE = 0.1951.
- The monitor is the benchmark's original stratified validation split (seed 42), which for this all-data refit is a subset of the training data. Its RMSE is an in-training model-selection signal only β it is not a benchmark metric and must never be reported as one.
- Ship bar (fixed in advance): monitor RMSE β€ 0.60 β (0.1951); snapshot reloads cleanly into a fresh model β; smoke predictions on 3 training compounds finite and sane β (7.07, 6.86, 5.17).
- This refit has never been scored on the FDA/absorbed set or any held-out benchmark, and no benchmark metric may be claimed for it.
Files
ltn_model.npzβ shipped weights (w0..w7, same format as the benchmark repo).infer.pyβ regression + derived classification from a SMILES (identical mechanics to the benchmark repo's infer.py).ltn_config.jsonβ configuration, incl. selected epoch and the in-training monitor value (labelled as such).epoch_trace.csvβ per-epoch trace: epoch, loss, monitor RMSE, snapshot saved (y/n) for all 87 epochs.gate_verification.json/training_manifest.csv/absorbed_compounds.csvβ hard-gate verification and the absorbed regulatory compounds.training_summary.jsonβ full run record.
Benchmark repo (untouched): Delower/dpp4-pic50-ltn.