iMut-CDR-GC

iMut-CDR-GC models

Evaluated models for the iMut-CDR-GC affinity-maturation-inspired antibody-design workflow. The workflow starts from a candidate antibody and combines local CDR mutation with computational selection.

Each folder contains the original evaluated model.pt, the required small inference assets, a model-specific usage guide, and a SHA-256 release manifest. Use the model-specific guide to download only the required folder. These are custom PyTorch tensor state dictionaries, not Transformers AutoModel archives. The full fine-tuned ESM2 encoder is included in JM/JM-Epi; separate base ESM2 weight downloads are unnecessary.

Code and loading

The portable code is pinned to 23dc11908d6e09b6e40feca2da4c503e9228d5c8. See the generation, antigen-graph, and scoring documentation for input contracts and command-line entrypoints.

For immutable downloads, specify the release commit in snapshot_download's revision argument. Verify the file hashes against the manifests before loading. Use torch.load(..., weights_only=True) through the documented loaders.

Scope and interpretation

Sequence probabilities and interaction-classifier outputs are computational signals, not experimentally measured affinity, specificity, or safety. Implementation-equivalence checks and target-specific limitations are described in each model card. Antibody candidates require appropriate independent validation.

The separate P5 likelihood model is not included in this release. The workflow's sequence-based scorer is available from the official DeepCDR-ESM2 repository; its weights are not mirrored here.

The original ESM license is retained alongside the two ESM-based models. Public hosting does not assign a new license to project weights or third-party material; retain the applicable component terms.