ralle-600m (v6.1, full-data)

Per-residue transmembrane topology predictor: frozen ESMC-600M embeddings + CNN + light-attention + grammar-CRF head (5 classes: i/o/H/B/S).

Trained on ralle dataset v6.1 (91,806 train records; release data-v6-full, https://huggingface.co/datasets/clear-bio/ralle-dataset), early-stopped at epoch 11/50 on an NVIDIA L4.

DeepTMHMM bench (n=3,571, split-mode masking, bootstrap 1000x): Q5 0.9608 [0.956-0.966], topology 0.8373 [0.825-0.849], SOV_all 0.939.

Known limitations: beta-barrel class collapses at this data scale (B recall 0.025); signal-peptide vs TM confusion 369 goldSP->TM / 21 goldTM->SP. See https://github.com/clear-bio/ralle for code and docs.