esm-bind: How much protein–RNA binding signal is already in frozen ESM-2 + RNA-FM representations?

How much protein–RNA binding signal is already accessible from frozen pre-trained encoders, and how small can the task head be? esm-bind is a ~7.8M-parameter bidirectional cross-attention head over frozen ESM-2 650M and RNA-FM, trained on 100K ENCODE eCLIP positives with a three-class blended negative sampler. Without fine-tuning to either external benchmark, it reaches zero-shot AUROC 0.840 on RPI7317 protein-disjoint, a standard protein–RNA pair benchmark with held-out proteins. On TheNovel — a 2,364-pair NPInter5-derived holdout from the ZHMolGraph 2025 supervised graph baseline, constructed node-disjoint to our NPInter2 training set so no train protein or RNA appears at evaluation — our 5-fold ensemble scores the same pairs and matches the published numbers (AUROC 0.793 [0.774, 0.810] vs 0.798; AUPRC 0.815 [0.796, 0.834] vs 0.820, both reference points inside our 95% CIs), despite training on a different dataset (NPInter2 vs NPInter5). Five capacity-augmenting ablations — encoder LoRA, RNA-encoder swap, extended training, fixed-weight graph context, and learnable graph attention — all fail to lift in-bench RPI7317 multi-seed AUROC outside a 0.812 ± 0.005 band. A zero-vector input ablation isolates a protein-class "RBP-vs-non-RBP" shortcut that inflates within-distribution headlines but does not transfer. We read these as evidence that frozen protein and RNA language-model representations already carry the bulk of pair-specific binding signal accessible at this scale. Head-only weights and a script that reproduces the 0.840 RPI7317 AUROC exactly are released anonymously at https://anonymous.4open.science/r/esm-bind-anon-3FCC/.