About / Methods
TriCyp is the deposition for an ESM2-based three-state cysteine classifier — disulfide-bonded, metal-binding, or free thiol — applied across roughly 700,000 ECOD F70 representative domains. Section anchors below match the manuscript Methods subheadings.
For every cysteine in a domain, the pipeline runs a fine-tuned ESM2 classifier that emits a probability over three states — free thiol, disulfide, metal-binding. Probabilities are thresholded into a three-state call; per-cysteine outputs are aggregated to per-domain and per-H-group summaries that back the dashboard, the family browser, and the H-group browser. PDB-source domains additionally carry structural ground truth from geometric Sγ–Sγ scanning, PDB SSBOND records, and PDB metal LINK records, which the benchmark page uses to score the classifier.
The classifier is trained on PDB-source cysteines with structural ground truth: disulfide labels from PDB SSBOND records cross-checked by Sγ–Sγ geometric scanning at 2.5 Å, metal-binding labels from PDB LINK / SITE records where a cysteine sulfur coordinates a metal ion or metal-bearing cofactor, and free-thiol labels for cysteines not captured by either evidence stream. Held-out validation and test sets are drawn at the F-group level so the splits do not leak structural family between training and evaluation.
The published 3-state classifier is an ensemble of five ESM2 checkpoints (best_modelA.pth … best_modelE.pth), each fine-tuned from the same ESM2 base with a per-cysteine softmax head over the three classes. Inference averages per-class probabilities across the ensemble. The full source — cys3state — is available alongside the model weights; see Downloads.
Operating thresholds were tuned on the held-out validation set to hit a fixed per-task precision target. The published 3-state call on TriCyp uses P(Met) ≥ 0.972 for metal-binding and P(Dis) ≥ 0.742 for disulfide; otherwise a cysteine is called free thiol. Raw per-class probabilities for every cysteine remain in the per-cysteine TSV so users can re-threshold for their own workflows.
The full ROC + PR + threshold-tuning panels and the metal-type stratified strip live on the Benchmark page (paper Fig 2 + Fig S1). On the held-out v2 (zinc-rebalanced) benchmark, all three metal-binding tools score in the same band on the metals they were trained for (Zn / Ca / Mg / Mn, AUROC 0.994–0.996); ESM2-3state's residual advantage shows up specifically on iron coordination, where iron-stratum AUROC reaches 0.993 for ESM2-3state versus 0.917 (LMetalSite) and 0.877 (GPSite). This is a training-coverage difference, not an architectural one — the specialist tools were not designed to predict iron-coordinating cysteines.
cys3state repository — see the link on the Downloads page.paper/figure_data/ exports.TriCyp data is released under CC-BY-4.0: free reuse with attribution to the manuscript. The predictor source code (cys3state) carries its own existing license — check the repository for terms before redistribution.
Please cite Classification of cysteine fates in structure predictions using a protein language model (Yuan, Durham, Cong, Schaeffer, 2026). preprint DOI pending.
@article{yuan_tricyp_2026,
title = {Classification of cysteine fates in structure predictions using a protein language model},
author = {Yuan and Durham and Cong and Schaeffer},
year = {2026},
journal = {bioRxiv},
note = {TriCyp companion site: https://tricyp.swmed.edu},
}TY - JOUR TI - Classification of cysteine fates in structure predictions using a protein language model AU - Yuan AU - Durham AU - Cong AU - Schaeffer, Dustin PY - 2026 JO - bioRxiv ER -