NABpred

The input sequence is validated, then embedded by the ESM2-650M transformer, producing a per-residue representation of dimension L × 1280. Masked mean and max pooling along the residue axis yields a fixed 2560-dim vector capturing both average and salient local features.

This vector is scored by an ensemble of five small fully-connected classifiers (Arm F; each 2560 → 128 → 1 with ReLU and dropout; training seeds 42, 1, 2, 3 and 4), trained to discriminate LLPS-associated nucleic-acid binding proteins from non-LLPS NABs. Each model's output logit is converted to a probability via sigmoid, and the five probabilities are averaged into the final score, with a default threshold of τ = 0.50. Reliability estimates are empirical, derived from held-out performance on the Hou cross-species benchmark (taxonomy-stratified AUROC), not a learned confidence score. The organism category is set once per request (a single category applies to every sequence in a batch). The per-residue disorder profile is a sequence-composition estimate based on FoldIndex (Prilusky et al., 2005), combining Kyte–Doolittle hydropathy and net charge over a 51-residue window and sign-inverted so that positive values indicate disorder-leaning regions; it is not a dedicated disorder predictor. The reference percentile places each score within the score distribution of Arm F's own validated holdout of 470 proteins.

POST /nabpred/predict accepts either { "sequence": "..." } for one sequence or { "sequences": [...] } for a batch (max 16 per request — chunk client-side for larger sets).