Description

Missense variants change a single amino acid in a protein and are a common source of variants of uncertain significance (VUS): about 90% of missense variants in ClinVar are VUS. The MPC score ("Missense deleteriousness Prediction by Constraint") is a machine-learning score that flags missense variants likely to be deleterious by combining three lines of evidence: (i) regional missense constraint (how depleted the surrounding sub-genic region is of rare missense variation in the general population), (ii) the biochemical severity of the specific amino-acid substitution as captured by PolyPhen-2, and (iii) cross-species conservation (phyloP). The model is trained to separate pathogenic from benign missense variation under strong heterozygous selection; higher scores indicate greater predicted deleteriousness. The authors report that MPC ≥ 2.5 is strongly enriched for de novo variants in individuals with severe developmental disorders relative to their unaffected siblings, with MPC 2–2.5 showing intermediate enrichment and MPC < 2 little enrichment.

This track shows MPC v4.1.1, computed by the Broad Institute gnomAD team from the gnomAD v4.1.1 release of 730,947 exomes aligned to GRCh38. Scores are provided for every possible single-nucleotide missense variant in 17,841 MANE Select or canonical protein-coding transcripts that passed gnomAD QC, as well as for an additional 1,534 transcripts that failed QC (the authors note that scores may be less accurate in the latter). Our MPC track on hg19 is an older release of the MPC score, calculated on gnomAD v2 (125,748 exomes) rather than gnomAD v4 (730,947 exomes).

Display Conventions and Configuration

Two views of the same underlying data are available:

• gnomAD MPC — four bigWig subtracks, one for each possible alternate allele (A, C, G, T). At a given genomic position, the subtrack for a given alternate allele shows the MPC score for the substitution from the reference base to that alternate base. Positions where the alternate base equals the reference base, or where no MPC score is available (for example, outside coding regions), are drawn as 0. Scores range from 0 to 6, where 6 is the ceiling assigned when a variant is more severe than every benign variant in the MPC training set (the maximum real computed value is just over 5); the track's default view limit is 0–3 because the vast majority of variants score below 3. Mouseover shows the score. When a variant is annotated against more than one transcript, the four bigWigs show the maximum (most deleterious) MPC across transcripts.
• gnomAD MPC overlaps — a bigBed track restricted to the small subset of variants that are scored against more than one transcript (about 250,000, or 0.4% of the ~70 million scored variants). Each row carries the full list of Ensembl transcripts scoring that variant together with their individual MPC scores on the details page. Single-transcript variants are not included in this track because the bigWig view already fully represents them. A range filter on mpcMax is available from the track configuration page. Items are colored by max MPC (gray < 1, orange 1–3, red ≥ 3). Use this view to inspect transcripts whose MPC scores disagree for the same underlying variant.

Across the 250,000 multi-transcript variants, per-transcript MPC scores typically agree within 0.5 units; only a few percent differ by more than 0.5. The bigBed view is the authoritative source for the full transcript-level detail in those cases; the bigWig view collapses transcripts by showing the maximum MPC at each position.

Methods

Regional missense constraint (MCR). For each of 17,841 QC-passing MANE Select or canonical coding transcripts, the authors tallied the observed rare missense variants (allele count > 0, allele frequency < 0.1%, %AN ≥ 20, QC PASS) in gnomAD v4.1.1 against the expected count under a position- and coverage-adjusted mutational model. A recursive likelihood-ratio test (Poisson model, p-value threshold 0.001, minimum 16 expected missense variants per sub-region) identifies change-points at which the transcript-wide observed/expected (OE) ratio deviates significantly; each resulting segment is a missense constraint region (MCR). 36% of transcripts (6,361/17,841) harbor two or more MCRs. MCR missense OE was calibrated against ClinVar P/LP vs. B/LB missense variants following ClinGen recommendations for the ACMG/AMP guidelines: OE ≤ 0.36 meets moderate evidence for pathogenicity, OE ≤ 0.59 meets supporting evidence for pathogenicity, OE > 0.97 and OE > 1.23 meet supporting and moderate evidence for benignity, respectively.

MPC score. MPC is an XGBoost gradient-boosted-tree classifier that takes as input (1) MCR missense OE, (2) gene-level constraint, (3) a per-substitution amino-acid severity feature, (4) the PolyPhen-2 pathogenicity score, and (5) phyloP conservation. Training: 20,931 "pathogenic" variants (high-quality ClinVar P/LP in 2,987 haploinsufficient genes with pHaplo ≥ 0.86 or in 359 non-LoF DD genes from Gene2Phenotype) vs. 93,638 "benign" variants (high-quality ClinVar B/LB or gnomAD variants with AF > 0.1% in the same gene set). The model is applied to all 70,313,598 possible exome-wide missense variants in the Ensembl VEP table. For a variant i, MPC is d_i = log₁₀(M / m_i), where M is the number of benign training variants and m_i is the number of those with a fitted pathogenicity probability lower than variant i's; when m_i is 0 the score is capped at 6. Higher scores indicate greater predicted deleteriousness. The authors caution that MPC is best suited to modelling strong fitness effects (as expected given its training set) and that naively taking the maximum of MPC and AlphaMissense decreases case/control discrimination for de novo variants relative to either score alone. Code for calculating the MPC scores and MCRs is available at the broadinstitute/regional_missense_constraint GitHub repository.

At UCSC. The precomputed MPC score table was downloaded from the gnomAD Broad public bucket at gs://gcp-public-data--gnomad/papers/2026-rmc/gnomad_v4.1.1_mpc.tsv.bgz, companion to the Hail-table release gnomad_v4.1.1_mpc.ht in the same directory. The input TSV contains one row per (locus, alleles, transcript) combination, for 70M rows. Two Python scripts in src/hg/makeDb/scripts/gnomadMpc convert to bigWig/bigBed formats. Build commands are documented in the hg38/gnomadMpc.txt makeDoc file.

Data Access

The raw data can be explored interactively with the Table Browser or the Data Integrator. For automated access, this track is available via our API. The underlying bigWig and bigBed files are at our download server as a.bw, c.bw, g.bw, t.bw, and mpcOverlaps.bb. Individual positions or whole chromosomes can be extracted with bigWigToBedGraph / bigWigToWig (for the bigWigs) or bigBedToBed (for the bigBed), for example:

bigWigToBedGraph -chrom=chr1 -start=100000 -end=100500 \
    http://hgdownload.soe.ucsc.edu/gbdb/hg38/gnomAD/mpc/a.bw stdout

The original MPC table and the accompanying missense constraint regions can be downloaded from the gnomAD downloads page.

Credits

Thanks to the gnomAD production team and the Samocha and MacArthur laboratories for generating and releasing the MPC scores.

References

Wang L, Chao KR, Panchal R, Liao C, Abderrazzaq H, Ye R, Schultz P, Compitello J, Grant RH, Kosmicki JA, Weisburd B, Phu W, Wilson MW, Laricchia KM, Goodrich JK, Goldstein D, Goldstein JI, Vittal C, Poterba T, Baxter S, Watts NA, Solomonson M, gnomAD consortium, Tiao G, Rehm HL, Neale BM, Talkowski ME, MacArthur DG, O'Donnell-Luria A, Karczewski KJ, Radivojac P, Daly MJ, Samocha KE. The landscape of regional missense mutational intolerance quantified from 730,947 exomes. bioRxiv April 23, 2026; doi: 10.1101/2024.04.11.588920.