Plain language.

Terms used throughout the project overview, protein explorer, and methods.

AlphaMissense

Google DeepMind's AI that predicts whether each possible single-letter protein mutation is likely disease-causing. It supplies the prediction used in this comparison.

AlphaFold

DeepMind's AI that predicts a protein's 3D shape. We use its structures and per-residue confidence to examine the structural context, not to predict disease.

pLDDT

AlphaFold's confidence at each residue, from 0 to 100. High (≥90, blue) means the predicted shape is more trustworthy there. Low (<50, orange) suggests a flexible or disordered region.

Missense variant

A mutation that swaps one amino acid for another. For example, R175H means arginine changes to histidine at position 175. This is the kind AlphaMissense scores.

Pathogenic / Benign

Disease-causing versus harmless. These labels form the clinical classification used in the comparison.

ClinVar

A public database of clinically interpreted variants. It is our primary clinical comparison and was also used to calibrate AlphaMissense, so independent evidence matters.

Review stars (★)

ClinVar's confidence in a classification. We keep ≥2★ (multiple submitters, no conflicts) and up; 3★ = expert panel, 4★ = practice guideline.

VUS

Variant of Uncertain Significance. ClinVar has not reached a clear classification, so the variant is excluded from accuracy scoring.

CFTR2

An expert-curated cystic-fibrosis variant database (89,052 patients). It supplies the benign CFTR variants ClinVar lacks.

DMS (deep mutational scanning)

A laboratory experiment that measures the effects of thousands of variants directly. It supplies evidence independent of clinical records and AlphaMissense calibration.

Circularity

Because AlphaMissense was calibrated on ClinVar, grading it against ClinVar partly grades it on its own answer key. Independent experiments break this.

Confusion matrix

The 2×2 of right/wrong: True Positive (correctly pathogenic), True Negative (correctly benign), False Positive (benign called pathogenic), False Negative (pathogenic missed).

Balanced accuracy

Accuracy that weights the two classes equally. Reported as the headline because ClinVar is mostly pathogenic, so plain accuracy flatters the model.

MCC

Matthews Correlation Coefficient, a score from −1 to +1 that measures agreement even when the classes are imbalanced.

ROC-AUC

How well the AI's continuous score ranks pathogenic above benign, regardless of threshold. 1.0 = perfect, 0.5 = coin flip.

Mantel-Haenszel OR

An odds ratio computed within each class and then combined. It controls for the tendency of benign variants to sit in low-confidence regions.