When an electrophysiologist records a stretch-activated current in a cell, they see an effect without knowing its cause: the membrane expresses hundreds of ion channel genes, many of which pass the same ions under the same triggers, so identifying the one pore-forming protein behind an anonymous current requires years of knockout, rescue, and mutagenesis work. In parallel, a very different literature has learned to read the preferences of language models directly off their token probabilities: minimal-pair scoring, where a model's likelihood is compared on two texts differing in a single word, exposes systematic bias before any text is ever generated. This proposal transfers that diagnostic machinery to protein language models (pLMs) — transformers trained on amino-acid sequences instead of words — and points it at mechanosensitive ion channels. The claim is that the same mathematical pathology the text-bias community diagnosed and cured, frequency confounds and sampling skew, is corrupting pLM likelihoods exactly where channel identification needs them most, and that a bias-corrected minimal-pair score can tell an electrophysiologist which mutations will selectively kill gating before anyone touches a pipette.
Three lines of bridging work already treat pLM likelihood as a fitness proxy: the "preference" study shows zero-shot mutation scoring works only in a Goldilocks band of wild-type sequence likelihood and degrades through phylogenetic overfitting, with thresholded evo-tuning lifting ESM-2 650M's mean Spearman correlation on ProteinGym from 0.439 to 0.462 Protein Language Model Fitness is a Matter of Preference; Likelihood-Fitness Bridging (LFB) averages mutation log-likelihoods across homologs to wash out species-specific drift, lifting ESM-2 15B's ProteinGym DMS Spearman from 0.398 to 0.436 and its clinical ROC-AUC from 0.895 to 0.938 From Likelihood to Fitness; and percentile recalibration at matched mutational depth corrects the similarity bias that makes raw scores collapse for distant variants Removing bias in sequence models of protein fitness. All three were developed and validated on ProteinGym's 217 substitution assays, dominated by soluble proteins ProteinGym, and none of them distinguishes a mutation that specifically breaks force-gating from one that merely destabilizes folding — the exact distinction channel identification hinges on.
Core Claim
Raw pLM likelihood is systematically biased against mechanosensitive gating residues, in a way structurally identical to the demographic-frequency confound in text models, and a minimal-pair score adapted from the text-bias toolkit — plus the existing recalibration and homolog-bridging corrections — can (i) flag which residues in a channel are gating- or pore-critical, and (ii) separate gating-specific mutations from globally destabilizing ones. This turns a diagnostic tool built to catch social stereotypes into a mutational-dissection planner for channels.
The structural analogy is precise. CrowS-Pairs showed that masked text models prefer sentences for reasons of token frequency rather than meaning, and fixed this by scoring the context conditioned on the demographic term via pseudo-log-likelihood, with an unbiased model expected to prefer stereotype and anti-stereotype pairs equally (50%), though in practice ALBERT preferred the stereotype in 67.0% of 1,508 pairs and religion pairs reached 75.2% CrowS-Pairs. In proteins, the gating residues of the auditory mechanotransduction channel TMC1 are glycine hinges (G411, G518, G539) and a pore-lining aspartate (D528), identified by structural ensembles, evolutionary coupling across ~5,380 TMC sequences, and patch-clamp recordings in double-knockout hair cells: G411 controls force sensitivity, G518/G539 set open probability, and D528 lines the conduction pore Mechanical gating of the auditory transduction channel TMC1. Glycine is among the most frequent residues in any training corpus, so a raw residue log-odds confounds "this position tolerates mutation" with "this amino acid is common," while the phylogenetic sampling skew documented by the bridging papers plays the role of the uncurated web corpus. The remedy transfers too.
The Diagnostic
For a channel sequence $x$ with wild-type residue $r^{\mathrm{wt}}$ at a position of interest, define two masked-model scores over the mutant residue $r^{\mathrm{mut}}$. The residue-marginal score $S_{\mathrm{res}} = \log P(r \mid x_{\setminus i})$ is the standard zero-shot quantity — how likely does this amino acid look at this site. The context-coherence score is the CrowS-Pairs transfer:
$$S_{\mathrm{ctx}}(r) = \sum_{j \in \text{context}(i)} \log P\big(x_j \,\big|\, x_{\setminus j},\, x_i = r\big)$$
that is, fix position $i$ to either the wild-type or mutant residue and ask how well the rest of the channel still reads as a viable protein given it. Because every summed term is the same true residue $x_j$, residue-frequency confounds cancel, exactly as the demographic-word frequency confound cancels in text minimal pairs. The pair $(\Delta S_{\mathrm{res}}, \Delta S_{\mathrm{ctx}})$, each a wild-type-minus-mutant difference, then decomposes a mutation's effect: a gating-specific variant should show a large residue-marginal shift with context coherence largely intact (the channel still folds; only the gate changes), while a folding-destructive variant collapses context coherence. The single-inference pseudo-log-likelihood approximation, calibrated to exhaustive masking at Spearman $\rho = 0.923$ Protein Language Model Fitness is a Matter of Preference, makes the $O(L)$ masking cost affordable, and the depth-matched percentile background and LFB homolog averaging from Removing bias in sequence models of protein fitness and From Likelihood to Fitness are applied on top, for the first time to multi-pass transmembrane pore domains.
Research Program
The program has three deliverables. First, a curated benchmark, MechanoPair, of electrophysiologically characterized minimal pairs in mechanosensitive and mechanically regulated channels, seeded from the file's documented mutants: the TMC1 panel with its qualitative effect classes (force-slope at G411, open-probability at G518/G539, conductance at D528 Mechanical gating of TMC1), and expanded by literature curation of point-mutant patch-clamp characterizations for the other channel families in scope — mBest1, the gene shown to encode the long-anonymous astrocytic calcium-activated anion current Park et al. 2009, TMEM87A/GolpHCat, the voltage-dependent cation channel behind Golgi counter-ion flux GolpHCat, and the RyR2 complex whose FKBP12.6 dissociation drives pathological calcium release in pulmonary hypertension Rieske iron-sulfur protein and RyR2 remodeling. Second, the decomposed minimal-pair scoring pipeline with the bias-correction stack, released with the benchmark. Third, a retrospective identification case study: replaying documented identification episodes (mBest1 for the astrocyte anion current, TMEM87A for GolpHCat) to test whether the decomposed score uniquely flags the true pore's gating residues among redundant bystander channels — the L-type, TRP, and store-operated channels that made vascular smooth-muscle calcium dissection so hard Hamada et al. 1994.
Proposed Experiments
Experiment 1 is the confound study. Build frequency-matched minimal pairs in which the wild-type residue's amino-acid frequency is varied independently of its functional role, and test whether raw residue log-odds tracks frequency rather than documented gating function — a direct replication, in sequence space, of the CrowS-Pairs pathology — and whether $S_{\mathrm{ctx}}$ removes it. Success means the raw score's preference rate on functional-versus-gating-dead pairs correlates with residue frequency (the pathology), while the context-coherence score does not.
Experiment 2 evaluates on MechanoPair. Because the file's TMC1 characterization is qualitative (no reported result found on Lacuna for numerical force-activation slopes or conductance values of the TMC1 mutant panel), the primary metrics are categorical: preference accuracy on functional-versus-loss-of-function pairs, and ROC-AUC for flagging gating/pore residues versus neutral ones, with the effect classes (slope, open probability, conductance) as separate strata. Where curated studies report graded effects, we add Spearman correlation.
Experiment 3 is the regression check on ProteinGym's 217 substitution assays, using the file's anchors: our corrections must not degrade ESM-2 15B below the LFB-corrected Spearman of 0.436 reported in From Likelihood to Fitness, and the 650M model should remain near its evo-tuned 0.462 from Protein Language Model Fitness is a Matter of Preference.
Experiment 4 is the specificity test, the proposal's decisive scientific question: on the TMC1 panel, does $(\Delta S_{\mathrm{res}}, \Delta S_{\mathrm{ctx}})$ separate D528-like pore mutations (residue-marginal shift, context intact) from G411-like gating-hinge mutations and from folding-disruptive controls, where raw likelihood and LFB — which average without conditioning on functional assay type — should not?
Baselines
The naive reference point is a coin flip on each minimal pair and, for ranking, raw wild-type log-odds from ESM-2 15B, which on ProteinGym DMS achieves a Spearman correlation of 0.398 From Likelihood to Fitness. The strongest existing competitor is LFB: homolog-averaged mutation log-likelihoods, reported at Spearman 0.436 on ProteinGym DMS and ROC-AUC 0.938 on the clinical benchmark for ESM-2 15B From Likelihood to Fitness. We will run LFB on MechanoPair ourselves with the same model, the same UniRef50 homolog retrieval, and the same compute budget, so the comparison isolates the minimal-pair decomposition rather than model scale or tuning. For gating-residue flagging, the field's incumbent is the evolutionary coupling analysis of Mechanical gating of TMC1, which identified TM4/TM6 from co-variation across ~5,380 sequences; we rerun it on the same alignments, noting that no reported result found on Lacuna for its numeric flagging accuracy. The key ablation is our own score stripped of its ingredients: residue-marginal only, context-coherence only, and each with and without depth-matched percentile recalibration, so that any gain over LFB is attributable to the decomposition and not to the borrowed corrections. Beating LFB means a materially higher preference accuracy and ROC-AUC on MechanoPair's gating strata, and a demonstrable separation of gating-specific from destabilizing mutations that LFB's single averaged score cannot produce, while staying within its ProteinGym performance margin.
Why not just…?
Why not just use structural models and evolutionary couplings, as the TMC1 study did? That approach did solve TMC1, but at the cost of a full structural-ensemble, coupling, and double-knockout patch-clamp campaign, and it relied on binary co-variation rather than a continuous per-mutation score; its own authors note the predictions lack membrane bilayer physics and cryo-EM verification of open and closed states Mechanical gating of TMC1. Our score is a cheap computational prescreen that complements it, and local alignment methods like EVMutation suffer even worse similarity bias than global language models Removing bias in sequence models of protein fitness. Still, on well-aligned channels the coupling baseline may win on flagging; Experiment 2 runs it head-to-head precisely to find out.
Why not just train a supervised model? Supervised ProteinNPT reaches a Spearman correlation of 0.613 on ProteinGym when labels exist ProteinGym, but patch-clamp electrophysiology is labor-intensive and low-throughput Mechanical gating of TMC1, and no labeled dataset of mechanosensitive gating mutants exists at the scale deep supervision requires; the ProteinGym analysis itself notes that supervised models overfit noisy, assay-specific labels and can trail zero-shot likelihoods in uncharacterized regimes. If MechanoPair's curation does yield a few hundred graded mutants, we will fit a small supervised head as an additional honest baseline.
Why not just use the largest raw model, or LFB as-is? This is the real threat. Raw ESM-2 15B likelihood plateaus or degrades with scale, and larger models overfit phylogenetic artifacts on high-likelihood sequences Protein Language Model Fitness is a Matter of Preference; LFB fixes much of this but conditions on nothing functional, leaving the stability-versus-gating ambiguity its own authors flag as open From Likelihood to Fitness, and it fails where homolog depth is thin — orphan channels are exactly the identification frontier. Experiment 4 is designed to settle this: if LFB's averaged score separates gating-specific from destabilizing mutations on the TMC1 panel as well as our decomposition, the proposal's central claim fails and we will report it.
Why not just call the analogy a category error — proteins have no stereotypes? The objection would hold if the transfer were metaphorical, but it is formal: both domains use masked transformers trained by cross-entropy on uncurated corpora, the demographic-word frequency confound of CrowS-Pairs maps onto residue frequency and phylogenetic sampling skew, and the pseudo-log-likelihood conditioning trick is the same computation. Experiment 1 is the falsification test: if the frequency confound does not reproduce in sequence space, the analogy is empty and the proposal should die there.
Decisive Checks
Three outcomes decide the proposal. First, Experiment 1 must show the frequency confound is real and removable; without it the text-bias framing is decoration. Second, Experiment 4 must show the decomposition separates gating-specific from destabilizing mutations where LFB cannot — this is the one result no existing method delivers. Third, the retrospective identification replays must place the true pore (mBest1, TMEM87A) at or near the top of the candidate ranking by gating-residue coherence, a small proof of concept given only two fully documented episodes, but the first evidence that likelihood bias correction can shrink the knockout-and-rescue search that channel identification has demanded for decades.
References
- Protein Language Model Fitness is a Matter of Preference
- From Likelihood to Fitness: Improving Variant Effect Prediction in Protein and Genome Language Models
- Removing bias in sequence models of protein fitness
- ProteinGym: Large-Scale Benchmarks for Protein Fitness Prediction and Design
- CrowS-Pairs: A Challenge Dataset for Measuring Social Biases in Masked Language Models
- Mechanical gating of the auditory transduction channel TMC1 involves the fourth and sixth transmembrane helices
- Bestrophin-1 Encodes for the Ca2+-Activated Anion Channel in Hippocampal Astrocytes
- GolpHCat (TMEM87A), a unique voltage-dependent cation channel in Golgi apparatus, contributes to Golgi-pH maintenance and hippocampus-dependent memory
- Rieske iron-sulfur protein induces FKBP12.6/RyR2 complex remodeling and subsequent pulmonary hypertension through NF-κB/cyclin D1 pathway
- Effect of extracellualr calcium removal on intracellular free calcium concentration in cultured vascular smooth muscle cells from two different strains of genetically hypertensive rats