bioseq-models weights

Pretrained weights consumed by the bioseq-models Python package. The package downloads only the files required by the model selected by the user.

The repository contains weights for Optimus 5-Prime, DeepSTARR, DeepSTARR-Mouse, DeepSTARR-Embryo, Malinois, DREAM LegNet, Cetnar stability, PARM K562, Agarwal lentiMPRA, UTR-CODE, and the multi-species CDS stability model with CodonBERT.

DeepSTARR-Mouse includes separate activity and accessibility models for heart, limb, midbrain, forebrain, hindbrain, and neural tube. Each package resource uses all six official checkpoints from three cross-validation folds and two replicates; requesting one resource downloads only its six checkpoint files. DeepSTARR-Embryo includes six accessibility resources and five activity resources. Each resource uses all author-released checkpoints with finite parameters; only the selected resource is downloaded.

Most checkpoints are project-trained reproductions and are not represented as author-released weights. See the package's MODEL_PROVENANCE.md for details.

License

The CodonBERT files under cds_stability/codonbert/ are restricted to academic research and non-commercial use. See THIRD_PARTY_NOTICES.md and the linked upstream artifact license. The complete repository is distributed under these restrictive terms so that a single unambiguous license applies to the weight collection.