Skip to content

v0.3.9 — robust loader + AD-Init seeding + explicit scope warning

Choose a tag to compare

View all tags
@krisss0mecom krisss0mecom released this 04 May 17:03
· 7 commits to main since this release
v0.3.9
f6362d6

Patch release. No model weights changed. Pretrained checkpoints (ad_transfer_v2_clean_best.pt, ad_init_full_1477_best.pt) are re-used from the v0.3.0 release assets — _RELEASE_TAG stays at v0.3.0 and the cache layout is unchanged.

Bug fixes

  • Robust checkpoint loader (weights.py). Strict load_state_dict plus tolerant handling of trainer checkpoint layouts (model_state / state_dict / model with model_config / config / args). Constructor kwargs are filtered against the model class signature, so unknown fields no longer silently corrupt config. Mismatched state dicts now raise instead of silently random-initialising part of the model.
  • AD-Init seeding (api.py). predict_torsion_ensemble now defaults to sequence-conditioned initial torsions from ad_init_full_1477 (the AD-Init prior) instead of a uniform random [-π, π] seed. Falls back to the legacy uniform seed if the AD-Init checkpoint cannot be loaded, with a stderr note. Pass init_model_name=None to opt out.
  • N=98 mdCATH sequence shift (data.py). Added align_sequence_to_residues(). Previously, when a sequences manifest stored the full domain sequence, the loader sliced it as seq[:n_residues] — for N=98 mdCATH this shifted the sequence by one residue and appended X. Now indexed via residue_indices whenever available.

UX

  • Banner now shows an explicit scope NOTE: AlphaDynamics is best validated for short peptides (4-15 aa). Sequences longer than 20 aa are outside the calibrated scope.
  • Soft UserWarning emitted from predict_torsion_ensemble when the input sequence is longer than 20 aa. Prediction still runs — this is a calibration disclaimer, not a hard guard.
  • Interactive CLI's length warning lowered from >200 to >20.

Honest scope

The phase-flow propagator couples oscillators within a residue but does not exchange information across residues. Empirically: excellent on 4–8 aa peptides (2.39× better Ramachandran JSD than Microsoft Timewarp on the 4AA benchmark), strong on 9–15 aa, usable but weaker on 16–25 aa, and aggregate Ramachandran density on chains >30 aa tends toward an "average amino-acid" pattern. For chains over ~20 aa, prefer per-residue Ramachandran panels over the aggregate plot.

Install

pip install --upgrade alphadynamics==0.3.9

Weights download lazily on first use from the v0.3.0 release.

Co-authored with Claude Opus 4.7 (1M context).

Assets 2
Loading