| license: mit | |
| library_name: pytorch | |
| tags: | |
| - protein-rna-binding | |
| - foundation-models | |
| - frozen-encoders | |
| - esm2 | |
| - rna-fm | |
| # esm-bind (head-only checkpoint) | |
| A ~7.8M-parameter bidirectional cross-attention head over **frozen** ESM-2 650M | |
| and RNA-FM for protein--RNA binding prediction. This repo holds only the | |
| trainable head (fp32, ~31 MB); the two frozen encoders are rebuilt from the Hub | |
| at load time at the exact revisions the head was trained against: | |
| - protein encoder: `facebook/esm2_t33_650M_UR50D` @ `08e4846e537177426273712802403f7ba8261b6c` | |
| - RNA encoder: `multimolecule/rnafm` @ `bcf7cc6e8f4385449c0288b3e1f41f21d68f1be3` | |
| Zero-shot AUROC **0.840** [0.832, 0.848] on RPI7317 (protein-disjoint), obtained | |
| without any supervised training on that benchmark. | |
| ## Loading | |
| See the reproduction instructions and loading code at | |
| https://github.com/ethanarnold/esm-bind (`make bench` reproduces the 0.840 | |
| number end-to-end). | |
| ## File | |
| - `s2_ep04.head_only.pt` -- production S2 epoch-04 head weights + `ESMBindConfig`. | |