Methodology

Pipeline

Stage 1: Variant annotation

• Source: NCBI ClinVar
• Input: a coding variant, given as an rsID (e.g. rs1065852) or a protein change
• Output: gene symbol, protein consequence, clinical significance, and review status

Stage 2: Target and structure mapping

• Source: UniProt, PDB, and SIFTS for residue-level cross-referencing
• Output: the canonical protein sequence, the residue the variant hits, and any experimental structure available for that protein

Stage 3: Structure prediction

• Source: ESMFold (Meta), run locally
• Output: a predicted 3D structure with per-residue pLDDT confidence, cached per gene and variant
• These are computational models, not experimental structures.

Stage 4: Stability channel

ThermoMPNN estimates the change in folding free energy (ΔΔG) the variant causes on the folded structure. Measured against real abundance data, this channel tracks how much protein is actually made. A large fraction of loss-of-function is the protein not being produced rather than a broken active site, and this is the channel that captures it.

Stage 5: Sequence-likelihood channel

A masked-marginal log-likelihood ratio from ESM-2 scores how surprising the variant residue is to a protein language model. This channel carries the activity signal, and, unlike the stability channel, it transfers between genes rather than being fit to one. The two channels answer different questions: is the protein there, and does the active site still work.

Predictor

The channels above, plus ligand descriptors, feed a regularized linear model trained on measured variant activity from the literature. The target is not just a ranking but a calibrated value: the fitted slope against measured activity sits near 1, which is what makes the output map onto a dose scale rather than only ordering variants. The training corpus is measured kinetic data, curated paper by paper. No synthetic or generated data is used to support any claim.

What this does not do

• It does not predict substrate specificity: the way one variant hits one drug and spares another. On the current features, that signal measures null across ten model families. The specificity is real in the data, it is not recoverable from whole-protein sequence embeddings plus ligand descriptors. The negative result is characterized, down to the panel experiment that would resolve it.
• No molecular docking or physics-based free-energy calculation.
• ESMFold structures are predictions, not experimental structures.
• Outputs have not been experimentally validated and are not for clinical use.

Data sources