Toward a Mechanistic Atlas of Protein Sequences (MAPS)
Interpreting ESM-C embeddings to explain missense variant pathogenicity.
A missense variant of a hemoglobin protein — a single amino-acid substitution highlighted within the fold.
Introduction
As genome sequencing becomes increasingly accessible, it has become a valuable source of clinical
information. A sequencing run returns the positions where a patient's genome diverges from the
reference. Most differences are harmless — a normal part of human variation — but occasionally one
is largely responsible for a patient's illness.
Sorting the harmful from the harmless is the problem of variant effect prediction (VEP). It
is hard because many variants are rare, and the same change can be catastrophic in one context and
benign in another. It also reaches beyond disease risk to molecular consequences — whether a change
disrupts a splice site, perturbs a regulatory signal, or alters how a protein folds. When these
questions can be answered, a lab can assign the variant a clinical significance, helping clinicians
reach a diagnosis, anticipate a condition's course, choose a treatment, or, just as valuably, rule a
suspect variant out.
When the evidence is too thin, the variant is instead classified as one of uncertain
significance (VUS). Labs deposit these calls — pathogenic, likely pathogenic, benign, likely
benign, or uncertain — with their supporting evidence in ClinVar, the public archive of
observed variants and their interpreted significance, and the source of every clinical label in this
study. But ClinVar grows by accretion, and the uncertain pile grows fastest: a large fraction of
missense entries are VUS, real and observed in patients but too thinly supported to classify. Each
is a patient and a clinician left without an answer, sometimes for years. A good computational
predictor promises to triage that backlog — flagging the VUS most likely to be pathogenic for
follow-up, and reassuring those most likely benign.
In prior work, we introduced EVEE (Evo 2 Covariance Probing), showing that information
learned inside a DNA foundation model, Evo 2, is a powerful tool for VEP on ClinVar. Covariance
probing, a technique we developed, reads pathogenicity from Evo 2 embeddings across variants of both
known and uncertain significance, and yields a mechanistic hypothesis by showing how a variant
disrupts a panel of probes for functional genomic information. However, a DNA model cannot fully
capture the protein-structural consequences of a coding change. For the large class of variants that
act by altering a protein, there is a further layer of "why": is the mutated residue buried in the
core or exposed on the surface, does it line an active site or a binding pocket, is it a cysteine
locked into a structural disulfide or a residue anchoring a transmembrane helix? These are the terms
a structural biologist reaches for, and they live most naturally in a model that has learned
proteins rather than DNA.
We therefore extend covariance probing to ESM-C (Evolutionary Scale Modeling Cambrian), a
bidirectional transformer trained on billions of protein sequences to predict amino acids from their
surrounding context. Because evolution preserves only mutations compatible with a protein's
structure and function, this objective forces ESM-C's representations to encode structural protein
information. It is this latent knowledge that we probe for variant effects.
ESM-C covariance probing is competitive with state-of-the-art approaches,
outperforming AlphaMissense and EVEE, and its functional information adds mechanistic
explanations on top of EVEE.
To explore these, we introduce the Mechanistic Atlas of Protein Sequences (MAPS), a viewer
of the variant-effect mechanisms ESM-C uncovers. Grounding VEP in a protein language model
complements genome-level approaches like EVEE. In future work, combining mechanistic insight across
the layers of the central dogma — starting with DNA and protein — is a promising path to fully
understanding variant effects.
Predicting pathogenicity from ESM-C embeddings using covariance pooling
To predict pathogenicity, we train probes over covariance-pooled embeddings from a frozen
ESM-C model. Covariance pooling aggregates a model's per-residue embeddings into a single
representation that captures how embedding features co-occur, rather than simply averaging them. The
change in this representation from wild-type to mutant is what we probe for pathogenicity. We read
probes off layer 78, which a per-layer sweep identifies as the depth where the pathogenicity
signal peaks.
For training, we used ~50,000 ClinVar variants confidently labelled pathogenic or benign,
and evaluated on homology-aware splits so that train and test never share a protein family.
We found that ESM-C covariance pooling performed best (test AUROC 0.944) compared to other
baselines, edging out AlphaMissense (0.943) and Evo 2 covariance (0.940).
Held-out AUROC on a homology-disjoint test set. ESM-C @ L78 (covariance) ranks at the
top of 17 methods, ahead of AlphaMissense and Evo 2 (covariance). Whiskers show confidence
intervals.
Mechanisms of pathogenicity can be explained by interpreting ESM-C embeddings
We've found that ESM-C embeddings carry a strong pathogenicity signal. Can we also extract the
biological properties associated with pathogenicity? To test this, we train linear probes for
several per-residue annotations spanning structural, topological, functional, and evolutionary
features, along with stability and InterPro domains.
To see how these properties are disrupted by a variant, we embed both the wild-type and mutant
sequences and probe the difference. We collect the shift in each decoded property to form a
"disruption profile" for each variant — the interpretation of what a mutation disrupts, sitting
alongside the single pathogenicity number that predicts whether it matters.
Case studies demonstrate mechanisms that genomics models can't find
Finally, we return to the original motivation of our work: to test whether variant prediction from
protein models does provide additional mechanistic information beyond genomic models. In several
cases, ESM-C probing provided accurate information on variant mechanisms not captured by Evo 2
probing.
Explore these variants and their disruption profiles interactively in the
Cards, Map, and per-variant views of this atlas.
Discussion
In this work, we introduce a mechanistic atlas of functional and structural disruptions that
missense variants can plausibly cause in proteins. Where Evo 2 probing (EVEE) characterizes variant
effects at the DNA level, MAPS serves as a complementary view that explains them at the protein
level. As our case studies show, ESM-C often captures information beyond the reach of a genome
language model.
Several directions could build on this work. Our interpretations are bounded by the annotations
available to probe; future work could improve extraction of known annotations or develop
feature-discovery methods that surface new ones. Combining genomic and protein language models,
along with protein folding models, is another promising avenue, and likely necessary to fully
characterize variant effects.
Related work
Several prior works have explored how protein language models can score how mutations affect
function and disease risk. Typically, this is done by reading a variant's effect from sequence
likelihoods in a zero-shot manner (Meier et al., 2021; Brandes et al., 2023). Related approaches
model evolutionary distributions directly (Frazer et al., 2021) or fold in structural context, as in
AlphaMissense (Cheng et al., 2023). Additional signals from sequence prediction models with
structure- and energy-based methods can also improve performance in variant effect predictions (Guo
et al., 2026). Here, we extend this line of work by applying our covariance-pooling method on the
latest model in the ESM family.
Another related thread of research focuses on mechanistically explaining what biological
information is encoded in protein language models. Many of these methods use sparse autoencoders to
decompose dense protein-sequence embeddings into interpretable features that map onto known biology —
amino-acid identity, secondary structure, tertiary contacts, folds, and functional sites (Simon and
Zou, 2025; Gujral et al., 2025; Silberg et al., 2025; Candido et al., 2026). Our work differs in that
we are interested in interpreting features to characterize clinical disruption.
Our work is most related to that of Fang et al., 2026, where the authors use SAE features for
zero-shot variant effect prediction and mechanistic explanations. A key difference is that our
pathogenicity predictors are separate from our annotation predictors. In addition, SAEs are
unsupervised feature discoverers that use auto-interpreted labels, while we take a supervised
approach with known, ground-truth annotations. Future work comparing or combining these two
approaches may provide a more cohesive mechanistic explanation for pathogenicity.
References
Brandes, Nadav, et al. "Genome-wide prediction of disease variant effects with a deep protein
language model." Nature Genetics 55.9 (2023): 1512–1522.
Meier, Joshua, et al. "Language models enable zero-shot prediction of the effects of mutations
on protein function." Advances in Neural Information Processing Systems 34 (2021):
29287–29303.
Frazer, Jonathan, et al. "Disease variant prediction with deep generative models of evolutionary
data." Nature 599.7883 (2021): 91–95.
Cheng, Jun, et al. "Accurate proteome-wide missense variant effect prediction with
AlphaMissense." Science 381.6664 (2023): eadg7492.
Wei, Angela, et al. "Investigating the sources of variable impact of pathogenic variants in
monogenic metabolic conditions." Nature Communications 16.1 (2025): 5223.
Guo, Ruihan, et al. "MutAtlas: A PDB-Wide Energy-Guided Atlas of Protein Mutation Effects."
Forty-third International Conference on Machine Learning.
Candido, Salvatore, et al. "Language modeling materializes a world model of protein biology."
bioRxiv (2026): 2026-06.
Simon, Elana, and James Zou. "InterPLM: discovering interpretable features in protein language
models via sparse autoencoders." Nature Methods 22.10 (2025): 2107–2117.
Silberg, Jake, Elana Simon, and James Zou. "Towards functional annotation with latent protein
language model features." bioRxiv (2025): 2025-10.
Gujral, Onkar, et al. "Sparse autoencoders uncover biologically interpretable features in
protein language model representations." Proceedings of the National Academy of Sciences
122.34 (2025): e2506316122.
All labels and features are assembled over 2,014 human proteins with AlphaFold structures and rich UniProt annotation. The headline numbers:
Count
What it is
38,960
Pathogenic / Benign missense variants used for supervised training
62,727
Clinically-annotated ClinVar variants in total (incl. VUS)
~416k
Variants scored proteome-wide by the trained probe
2,014
Human proteins with AlphaFold structure + UniProt annotation
Clinical labels — ClinVar. ~38,960 Pathogenic/Benign missense variants for supervised training; 62,727 clinically-annotated variants in total (including VUS), and prediction scores extended to ~416,000 variants across the proteome.
Temporal snapshots. ClinVar frozen at 2021-06, 2023-06, 2024-06, with the former-VUS that were later reclassified used as a leakage-free prospective test set (2,897 / 2,420 / 1,682 resolved variants respectively).
Functional annotation. UniProt active/binding/metal sites, disulfides, transmembrane spans, signal peptides, PTMs, domains — the label sources for the annotation-probe battery.
External baselines. AlphaMissense, EVEE, ESM-1v / ESM2 zero-shot likelihoods, and MSA-based conservation, all evaluated on the identical splits.
Methods & probe architectures
Representations
We extract frozen per-residue embeddings from ESM-C 6B layer 78 (2,560-d per residue). Layer 78 is not arbitrary: in a per-layer sweep, linear pathogenicity separability rises through the network and peaks in the deep layers around 78, so that is the layer we read every probe off. Each variant is summarised by its mutated-site embedding, and we also test mean and covariance ("cov-pool") pooling over the whole sequence. The WT→mutant delta embedding feeds the disruption analysis.
Probes are deliberately simple: StandardScaler → LogisticRegression(class_weight="balanced") for classification and ridge regression for continuous targets. Simplicity is the discipline — a linear read-out can only succeed if the information is already linearly present in the embedding.
Two probe families
We train two architecturally distinct kinds of probe, which answer two different questions.
A · Variant pathogenicity probes
These answer the clinical question directly: is this variant pathogenic? The strongest variant is covariance pooling (logcov) — instead of reading a single residue, we summarise the whole-sequence WT→mutant change as a covariance ("how the embedding's internal correlations shift"), and probe that. The single-site supervised probe we ship in the viewer is P_PB (the score colouring the atlas). Cross-validated AUROC reaches ~0.94. Because it pools across the sequence, it is a predictor, not an explanation.
B · Per-residue annotation probes
These don't ask "is it pathogenic" — they decode an interpretable property of a position, so we can later explain why a variant matters. There are two sub-types:
[1] Wild-type probes. These read a property off a single protein sequence as it is: "Is residue 47 in a helix?" "Is position 112 buried?" "Is this site evolutionarily conserved?" The input is the embedding of the wild-type protein. These build the static, per-residue annotation tracks.
[2] Disruption / variant probes. These ask what changes when you mutate a residue. We embed the wild-type protein, embed the mutant, and probe the difference vector — the shift in each decoded property (burial, secondary structure, conservation, contacts, site membership…). This decomposition is what turns one pathogenicity number into a channel-by-channel mechanistic profile.
Why covariance pooling, and why ESMC
How you pool the per-residue embedding matters as much as which model you use. Covariance pooling beats simple mean pooling, and ESM-C beats the comparable ESM2 at every pooling scheme — the gap widens as the pooling gets cruder, which is exactly the signature of a richer representation.
Held-out pathogenicity AUROC for ESM-C 600M vs ESM2-650M under three pooling schemes. Covariance pooling is best for both models, and ESM-C leads everywhere — most decisively under the cruder mean / full-field-mean pooling.
Honest evaluation
Homology-aware splits. MMseqs2 id30 clustering, protein-grouped and UniRef50-cluster cross-validation, so train and test never share a protein family (no homology inflation).
Baseline comparisons for every score. Every pathogenicity prediction is shown against its baselines (AlphaMissense, EVEE, ESM2, zero-shot likelihoods, MSA, structure-only), and every annotation probe is scored by its lift over the best model-free baseline rather than raw accuracy — the honest measure of what the model adds. Throughout, results are shown as horizontal-histogram comparisons with 95% bootstrap CIs.
Leakage-free temporal test. The strongest validation (see Temporal Analyses): retrain on what was known at an old cutoff, then predict the then-unknowns.
Mechanistic explanations with GPT-5.5
The disruption probes produce a per-variant channel profile (which properties shift, and by how much). On its own that is a bar chart; to make it readable we feed the profile to GPT-5.5, which synthesises a short, grounded mechanistic explanation of why the variant is disruptive — strictly constrained to the channels actually present in the data. This is wired live into the viewer; see Mechanistic profiles & GPT-5.5 below.
Pathogenicity prediction
The ESMC supervised probe is a strong variant-effect predictor — on the general benchmark it tops every method, including the specialised AlphaMissense, and on the leak-free temporal test it holds up.
Data & method
We train on the ~38,960 ClinVar variants confidently labelled Pathogenic or Benign, extract their ESMC L78 representations, and fit the simple linear probes described above — single-site P_PB and whole-sequence covariance pooling. That is a small training set by deep-learning standards, which makes the result below notable: a frozen model plus a linear read-out, on ~39k labels, competes with purpose-built supervised variant-effect models. Validation is gene-holdout cross-validation here, and the stricter leakage-free temporal protocol in the next section.
Figure 1 · General pathogenicity benchmark
Gene-holdout AUROC on the full ClinVar pathogenic/benign set, on a homology-disjoint test set (95% bootstrap CIs). ESM-C covariance pooling at layer 78 is the top method (0.944), edging AlphaMissense (0.943) and Evo 2 covariance (0.940); the previous-generation ESM2, the zero-shot likelihoods, MSA and structure-only baselines trail clearly.
Held-out pathogenicity AUROC for every method on the homology-disjoint test set, ranked (dots = point estimate, whiskers = 95% bootstrap CI). ESM-C @ L78 covariance leads, with AlphaMissense, Evo2 covariance, and ESM-C mean-pool within a CI of it; pooling matters (covariance > mean > full-mean within each model), ESM-C beats ESM2 at matched pooling, and the structure-only baseline sits at the bottom (~0.81). The same numbers are tabulated below.
Method
AUROC (95% CI)
ESM-C cov
0.9441 [0.9403–0.9479]
AlphaMissense
0.9429 [0.9390–0.9468]
Evo 2 cov
0.9398 [0.9357–0.9439]
ESM-C mean
0.9330 [0.9286–0.9374]
Evo 2 mean
0.9313 [0.9270–0.9356]
ESM-C zero-shot
0.9299 [0.9253–0.9345]
ESM-C P_PB
0.9266 [0.9219–0.9313]
ESM2 cov
0.9218 [0.9171–0.9265]
MSA + struct
0.9092 [0.9042–0.9142]
ESM-C full-mean
0.8958 [0.8903–0.9013]
Evo 2 full-mean
0.8952 [0.8899–0.9005]
ESM2 P_PB
0.8802 [0.8742–0.8862]
ESM2 zero-shot
0.8776 [0.8701–0.8851]
ESM2 mean
0.8692 [0.8630–0.8754]
MSA log-odds
0.8543 [0.8476–0.8610]
ESM2 full-mean
0.8255 [0.8183–0.8327]
Structure-only
0.8086 [0.8012–0.8160]
Colour key (in the live figure): ESM-C · AlphaMissense · Evo 2 · ESM2 / MSA / structure.
Held-out pathogenicity AUROC by model (covariance probe; 95% bootstrap CIs). Scaling up within the ESM-C family — 600M → 6B — lifts the probe past ESM2 and to the front of the pack, level with the specialised AlphaMissense and ahead of Evo2-7B.
Figure 2 · Which layer carries the signal?
Pathogenicity separability is not uniform across the network. Probing the representation at each depth shows linear P/B separability climbing through ESM-C's 80-layer stack, plateauing through the middle, and then rising to a clear peak in the deep layers around layer 76–78 (AUROC ≈ 0.93, well above the L60 reference) — which is why every probe in this study reads off L78.
Held-out pathogenicity AUROC of the ESM-C probe at every layer index (0–80). Signal climbs early, plateaus through the trunk, and peaks in the last handful of layers — layer 78 is the best read-out point. Dashed line marks the L60 subsample reference (0.854).
The same pattern holds when ESM-C and ESM2 are placed on a common axis of relative depth: both gain most of their pathogenicity signal in their deepest layers, and ESM-C edges ahead of ESM2 at the very top of the stack.
ESM2-650M vs ESM-C 600M, held-out AUROC against relative depth (layer ÷ n_layers). Both peak at the deepest layers; ESM-C finishes higher, consistent with the model-comparison and pooling results above.
The full leak-free, prospectively-validated version of this benchmark — retraining on what was known at each past ClinVar cutoff and predicting the then-unknowns — is in the Temporal analyses section.
Experimental validation (DMS)
ClinVar labels are expert clinical opinions. Deep mutational scanning (DMS) is a different kind of evidence entirely — a direct wet-lab fitness readout for thousands of mutations in a single protein. If the probe reads real biology, it should track DMS fitness on proteins it never trained on.
On leak-free, non-human enzymes (entirely absent from ClinVar training), the ESM-C covariance probe anti-correlates strongly with measured fitness — direct evidence it reads genuine loss-of-function, not human-annotation artefacts. A negative ρ is the correct sign: higher predicted pathogenicity should mean lower experimental fitness.
DMS assay
n
Spearman ρ
AUC vs LOF
BLAT_ECOLX (β-lactamase)
4,783
−0.751
0.896
AMIE_PSEAE (amidase)
6,227
−0.613
0.826
DYR_ECOLI (DHFR)
2,916
−0.523
0.882
CBS_HUMAN (in-distribution ref.)
7,217
−0.347
0.702
The three held-out non-human assays are the honest test, and the probe does well on all of them; CBS_HUMAN is shown only as an in-distribution reference. The same pattern is visible whether we use the bilinear covariance probe or the log-covariance probe.
Covariance-pooling pathogenicity probe vs DMS fitness. Left: |Spearman| against measured fitness; right: AUC separating loss-of-function from tolerated mutations. Leak-free non-human assays (BLAT, DYR, AMIE) plus the leaky in-distribution CBS reference, for the cov_bilinear and logcov probes.
Per-residue annotation probes
Beyond one pathogenicity number, can the embedding tell us why a variant is disruptive? We train a battery of per-residue annotation probes — one per biological property — and ask, for each, whether the model genuinely adds signal, then confront the severe class imbalance these biological labels carry.
How we decide what is worth probing: lift over baseline
A high raw AUROC is not enough — many "annotations" are recoverable from amino-acid identity or simple chemistry alone. The goal of a lift-over-baseline analysis is to compute the probe's score minus the best model-free baseline trained on the same split. We compare against two baselines:
[1] One-hot amino acid. A classifier that sees only which of the 20 residues is present. If a property is mostly determined by amino-acid identity, this baseline already captures it — and the probe deserves no credit for re-deriving it.
[2] Physico-chemistry only. Hand-computed sequence chemistry (Kyte–Doolittle hydropathy, charge, complexity, position, FoldIndex). This asks whether the annotation falls out of basic chemistry rather than the learned representation.
Lift — not raw score — is the honest number, and it cleanly separates learned structural channels from identity-shortcut ones. The figure below shows AUROC-lift and AUPRC-lift for every head on one common axis (positive = the embedding beats the best baseline):
ESMC-6B L78 probe battery — AUPRC-lift (orange) and AUROC-lift (blue) for each binary head, plus R²-lift / accuracy-lift (gold) for regression and multiclass heads, all measured as lift over the best amino-acid / physico-chemistry baseline. 21 of 22 binary heads keep positive AUPRC-lift; hatched bars are coverage-fragile (n⁺ < 30). The channels at the top (glycosylation sequon, zinc finger, DNA binding, pLDDT, conservation, accessibility, contact geometry) are the ones the embedding genuinely earns.
The same numbers, tabulated, with the per-probe verdict:
ESMC-6B L78, all-residue training, homology-aware CV. Binary heads: AUROC / AUPRC; regression heads (italic): R² / Spearman. Lift = probe − best of {one-hot AA, physchem}, on AUPRC (binary) or Spearman (regression).
Probe
AUROC / R²
AUPRC / ρ
lift
verdict
pLDDT (R_PLDDT_WT)
0.871
0.840
+0.650
learned ✓
Intrinsic disorder (P_DISORDER)
0.981
0.930
+0.621
learned ✓
Conservation (R_CONSERVATION) — new
0.639
0.789
+0.575
learned ✓
Contact number (R_CONTACT_NUMBER) — new
0.807
0.902
+0.573
learned ✓
Solvent accessibility (R_RSASA_WT)
0.816
0.905
+0.566
learned ✓
Buried / exposed (P_RSASA_BIN_WT)
0.961
0.929
+0.431
learned ✓
Secondary structure (P_SS3_WT)
0.823
0.783
+0.414
learned ✓
Signal peptide (P_SIGNAL_PEP)
0.995
0.983
+0.834
learned ✓
Transmembrane (P_TRANSMEM)
0.982
0.809
+0.711
learned ✓
Active site (P_ACTIVE_SITE)
0.982
0.778
+0.743
learned ✓ (was shortcut on ESM2)
Binding site (P_BINDING_SITE)
0.945
0.682
+0.555
learned ✓
In domain (P_IN_DOMAIN)
0.862
0.611
+0.415
learned ✓
Disulfide (P_DISULFIDE, within-Cys)
0.924
0.459
+0.383
learned ✓
Phosphorylation (P_PHOSPHO)
0.960
0.586
+0.288
identity-leaning
Ubiquitination (P_UBIQUITIN)
0.958
0.345
+0.230
identity-leaning
The channels that earn the embedding — fold confidence, disorder, conservation, packing geometry, accessibility, secondary structure, transmembrane/signal topology, active & binding sites — lift +0.4 to +0.85 over identity, exactly the channels a mechanistic disruption profile should lean on. Two notes: (1) on ESMC, active site is now genuinely learned (+0.74), where on ESM2 it was an identity shortcut; (2) the residue-restricted PTM heads (phospho, ubiquitin) post high raw AUROC but small lift — they are useful as rare, specific flags, not as mechanism. The full ~40-head battery is tabulated in docs/maps_report.md. A caveat we hold honestly: projecting the WT→mutant delta onto these named channels is lossy (it separates P/B at AUROC ≈0.65 versus ≈0.91 for the raw delta probe), so the channel profile is an explanation tool, not a competing classifier.
Class imbalance across annotation features
Biological site labels are extremely rare — many features have well under 1% positives, which is why AUPRC (not accuracy) is the honest metric and why several heads are "positive-starved". The viewer widget shows the positive prevalence of every binary annotation probe (sort by prevalence or by embedding lift; toggle a log scale to see the rarest features). Bar = % positive residues; colour = embedding AUPRC lift over baseline (darker = more genuinely learned).
The full probe catalog — and an honest status for each
Below is the complete scope of what we have probed (or scoped to probe) so far, grouped by what kind of biology each channel captures. The status column is the honest verdict:
✓ built & learned — positive lift over the best baseline; worth leaning on in a mechanism profile.
◐ built but identity-leaning / rare — useful as a flag, but its raw score partly reflects amino-acid identity or it is coverage-fragile.
○ planned — designed, not yet trained. The most promising of these, conservation-coevolution and ΔΔG stability, are the clearest levers for future work.
Structure
Probe
What it measures
Label source
Status
Relative solvent accessibility
How exposed vs. buried a residue is — continuous (RSASA) and a buried/exposed binary.
DSSP ACC ÷ per-AA max ASA (Tien 2013).
✓ learned
Secondary structure (SS3/SS8)
Helix / strand / coil class per residue.
DSSP on PDB (SIFTS-mapped) or AlphaFold.
✓ learned
pLDDT (fold confidence)
AlphaFold's per-residue confidence; low pLDDT often flags disorder/flexibility.
AlphaFold DB B-factor column.
✓ learned
Intrinsic disorder
Whether a residue lies in a region with no stable folded structure.
DisProt / MobiDB (gold) + IUPred3 (silver).
✓ learned
Contact number
Count of residues packed within 8 Å — local density / burial.
Cβ–Cβ contacts from PDB/AlphaFold.
✓ learned
Half-sphere exposure / long-range contacts / contact order
Directional exposure and how sequence-distant a residue's contacts are.
Computed from structure coordinates.
✓ learned
Residue depth / packing
Distance to the molecular surface — a finer burial measure than the binary.
How strongly a position co-varies with others (3D contact / epistasis).
EVcouplings / CCMpred couplings (APC).
○ planned
Stability & pathogenicity
Probe
What it measures
Label source
Status
ΔΔG sensitivity
How much, on average, mutating this position destabilises the fold.
Tsuboyama 2023 mega-scale folding stability.
○ planned
Pathogenicity (P_PB)
Pathogenic vs benign — the headline predictor.
ClinVar P/LP vs B/LB (38,960; ≥1★).
✓ headline
Evaluation throughout: binary heads report AUROC + AUPRC as lift over the best of {one-hot AA, physico-chemistry}; regression heads report R²/Spearman; all on homology-aware (MMseqs2 id30) splits, with external sanity checks where available (CB513 / NetSurfP for structure, ConSurf-DB for conservation, ProteinGym DMS for pathogenicity). The most promising ○ planned channels — conservation/coevolution and ΔΔG stability — are exactly the axes EVEE and stability predictors exploit, and are the clearest levers to push the mechanistic profile from explanation toward prediction.
Mechanistic profiles & GPT-5.5
The pathogenicity score says whether; the annotation probes say why. Together they turn one opaque number into a per-variant disruption profile — a readable account of which structural and functional channels a mutation breaks.
From delta vector to disruption profile
For any variant, we embed the wild-type and the mutant, take the difference, and ask each disruption probe how much its decoded property shifts: does the residue become more exposed? does the local secondary structure change? does a binding-site or disulfide signal weaken? The result is a channel-by-channel bar chart — the disruption profile you can open for any variant in the Map and Tracks views. We hold one caveat honestly: projecting the WT→mutant delta onto named channels is lossy (it separates P/B at AUROC ≈ 0.65 vs ≈ 0.91 for the raw delta probe), so the profile is an explanation tool, not a competing classifier — and channels that partly read amino-acid identity (active-site, metal) are flagged as such.
GPT-5.5 synthesis
A bar chart of fourteen channels is still work to read. So we hand the structured profile — the substitution, the P_PB score, and the per-channel shifts — to GPT-5.5, which writes a short, grounded mechanistic explanation: what the call is, which genuine channels drive it, and the relevant side-chain chemistry, strictly constrained to the channels present in the data. The model is instructed never to invent a channel, to lead with conservation and structure when they move, and to say plainly when the profile is near-silent and the call rests on the holistic embedding rather than a decomposable mechanism.
Appendix
Coming soon.
Mechanism synthesis
AI synthesis from the probe profile · not a clinical determination
