Property-specific neural oracles that predict five biophysical developability assays for therapeutic IgGs from paired VH/VL sequences, combining frozen protein-language-model encoders (ESM-Cambrian, ProtT5) with configurable attention decoders.

Code: https://github.com/SimonCrouzet/CrossAbSense Dataset: GDPa1 (242 IgGs, Ginkgo Bioworks)

Each property folder (<PROPERTY>_<config-checksum>/) contains: final.ckpt (model trained on all data — used by predict.py), fold0-4.ckpt (5-fold CV checkpoints), config.yaml, and property.txt.

These weights were trained from the published configs but in an environment without BioPhi (OASis humanness) and ScaLoP available. Those two antibody-feature sources were substituted with sentinel values during training, so the feature inputs differ slightly from the paper runs. This mainly affects AC-SINS (~0.05 below paper); the other four properties match or exceed Table 1. A future v1.0 will retrain the feature-using properties with BioPhi/ScaLoP restored. Pin revision="v0.9" if you need exactly these weights.

By default only final.ckpt (+ small metadata) is downloaded; the 5 CV fold checkpoints are fetched only when you ask for them (--folds, or predict.py --use-cv/--fold).

Or let predict.py fetch on demand:

Apache-2.0, matching the CrossAbSense repository.