sinhakrishnendu/BABAPPAlign: BABAPPAlign v1.0.4

Stable release with externalized learned scorer and fully reproducible model loading

This release finalizes the BABAPPAlign architecture by decoupling learned scoring parameters from the source distribution while preserving deterministic alignment behavior and benchmark performance.

Highlights

Externalized learned scorer The trained BABAPPAScore parameters (babappascore.pt) are no longer bundled with the source code and are instead distributed via GitHub Releases. The scorer is downloaded automatically on first use and cached locally using an XDG-compliant directory.

Reproducible, lazy model loading Learned scoring weights are retrieved deterministically from the tagged release and reused across runs, enabling reproducibility without inflating package size or violating conda-forge policies.

No changes to alignment logic or benchmarks The progressive alignment core, profile–sequence dynamic programming, and learned residue–residue scoring remain unchanged. All previously validated adversarial sanity tests and BAliBASE benchmarks produce equivalent results.

conda-forge compliant packaging Source distribution now contains only code and metadata, with all binary model artifacts excluded. This release is suitable for submission to conda-forge and other curated package repositories.

Notes

BABAPPAlign requires the learned scorer to function; alignment without BABAPPAScore is intentionally unsupported.

On first execution, users will see a one-time download message for the scorer weights. Subsequent runs operate fully offline.

The ESM2 model is used strictly in feature-extraction mode; warnings related to unused pooler parameters can be safely ignored.

Assets

babappascore.pt — trained BABAPPAScore model parameters (required)