What is MissenseDB?

MissenseDB is a proteome-wide database of human missense variant effect predictions integrating three state-of-the-art computational predictors — AlphaMissense (AM), VESM, and VESPA — across 19,436 UniProtKB/Swiss-Prot human protein entries. Per-variant scores are retained on each predictor's native scale and converted to a common calibrated 0–1 scale (higher = more pathogenic; see Empirical Score Calibration), and are mapped onto experimental and AlphaFold-predicted protein structures for 3D visualisation.

For each position in every protein, MissenseDB precomputes a residue mutational-sensitivity score for each predictor — the mean of its calibrated scores across all 19 possible amino acid substitutions — and the mean of these scores across the three predictors, providing a position-level summary of mutational intolerance independent of any specific observed variant. Calibrated scores are obtained by empirical binning of each predictor's raw score against HUMSAVAR LP/P and LB/B labels (see Empirical Score Calibration). For backward compatibility, position-level scores on a global per-predictor min-max normalised scale (VESM log-likelihood ratios inverted, values clipped to [0, 1]) are also stored. Structural annotations (DSSP secondary structure, ligand proximity), domain assignments (Pfam), and clinical variant data (HUMSAVAR, dbSNP, ClinVar) are integrated to enable both per-protein inspection and proteome-wide programmatic discovery.

The three predictors are:

• AlphaMissense (AM) — Cheng et al., Science 2023. A structure-informed pathogenicity classifier built on AlphaFold2, fine-tuned on population frequency and functional data. Scores represent the probability of pathogenicity.
• VESM — A sequence-only protein language model trained by co-distillation across the ESM family, compressing collective evolutionary knowledge into a single 3-billion parameter model. Raw scores are log-likelihood ratios; higher (more positive) values indicate more tolerated substitutions. On the calibrated 0–1 scale the orientation is reversed so that 1 = most pathogenic, consistent with AM and VESPA.
• VESPA — Marquet et al., Human Genetics 2022. T5 protein language model embedding-based predictor combining conservation and substitution signals. Scores represent predicted pathogenicity probability.

Comparison with Related Resources

No comparable resource combines three complete variant-effect predictor landscapes — AlphaMissense, VESM and VESPA — a position-level mutational intolerance map, synchronised multi-structure views, and composable proteome-wide querying in a single persistent precomputed database. ProtVar, dbNSFP and CATVariant each address complementary use cases; the table below summarises how MissenseDB differs.

✅ fully supported  ·  ⚠️ partial or different implementation  ·  ❌ absent.

Data Sources

Update policy

MissenseDB is updated annually. Each release incorporates accumulated updates to PDB, AlphaFold, UniProt Swiss-Prot, HUMSAVAR, SIFTS, Pfam, dbSNP and ClinVar over the preceding year. The current release versions are listed in the Data Sources table above. A changelog will be maintained on this page between releases.

Variant Processing Pipeline

MissenseDB integrates HUMSAVAR release 2026_01 (28-Jan-2026). HUMSAVAR was chosen as the clinical annotation backbone because it provides expert-curated ACMG/AMP classifications tied directly to UniProt canonical sequences, offering stable protein-coordinate anchoring and consistent variant nomenclature across releases. For HUMSAVAR variants that carry an rsID, the corresponding dbSNP record is fetched to enrich each variant with population allele frequencies and per-submission ClinVar significance calls. Not all HUMSAVAR variants appear in the queryable database — the following pipeline explains each filtering step and the resulting counts.

Using the Web Interface

Searching for a protein

Enter a UniProtKB accession (e.g. P04637) in the search box on the home page. Autocomplete suggestions appear after two characters — only proteins present in MissenseDB are shown. Press Enter or click Visualise to load the results page.

Results page layout

The results page is divided into three areas:

• Structure viewer (left) — Mol* 3D viewer rendering the AlphaFold or experimental PDB structure coloured by the calibrated residue score (0–1, higher = more pathogenic; B-factor channel). Switch between AM, VESM, and VESPA colouring using the predictor buttons. Click any structure in the Available Structures table to load it.
• Residue detail panels (right) — the top-right panel shows all 19 substitution scores at the clicked residue position for the selected predictor, sorted by raw score. Each raw score is followed, in brackets, by its calibrated pathogenicity probability (0.5 = decision boundary), so the three predictors can be compared on a common scale. The position summary strip below shows the calibrated AM, VESM, VESPA and mean scores at that position (the mean of the 19 calibrated substitution scores), the UniProt position, the mapped structure residue number and the pLDDT. This is only activated after a single position column is selected from the feature viewer in the second row.
• Sequence feature viewer (bottom) — an RCSB Saguaro 1D track viewer showing the full protein sequence with overlaid annotation tracks. Click any position in the sequence track to select it — this synchronises the 3D viewer highlight, the residue detail table, and the position summary.

Sequence feature tracks

Downloading score-annotated structures

For each protein, calibrated residue scores are mapped into the B-factor column of the corresponding structure files, producing three score-annotated PDB files — one per predictor (AM, VESM, VESPA). Calibrated scores (0–1) are written to the B-factor column multiplied by 100, so values lie between 0 and 100. These can be downloaded directly from the results page using the Download PDB button beneath the structure viewer. Downloaded files can be opened in any molecular visualisation software (PyMOL, ChimeraX, VMD) and coloured by B-factor to reproduce the MissenseDB score mapping. The structure identifier, predictor and calibrated are encoded in the filename.

dbSNP variant panel

For proteins with HUMSAVAR-annotated variants, a dbSNP Variants section appears below the sequence viewer. The full-width table lists all annotated substitutions with their ACMG/AMP classification, rsID, AM/VESM/VESPA/mean scores, and disease association. Click any row to expand the detail panel showing population allele frequencies across multiple cohorts (gnomAD, TOPMed, 1000 Genomes, TOMMO and others) and all ClinVar submissions with significance classifications.

Genomic coordinate lookup

The Genomic Lookup page maps GRCh38 nucleotide positions to UniProt residues and retrieves predictor annotations for each mapped position. Enter one coordinate per line in any of the accepted formats (chr17:7674220, 17:7674220, chr17 7674220) — up to 5,000 per submission. Tick Include calibrated predictor scores to add the calibrated AM, VESM, VESPA and mean residue scores (each the mean of the 19 calibrated substitution scores at that residue), together with pLDDT, DSSP secondary structure, and ligand proximity, to each mapped residue. Results are displayed in a filterable table with status indicators (mapped / unmapped / invalid) and can be downloaded as TSV. Click any mapped UniProt ID in the results table to open the full MissenseDB results page for that protein. Programmatic access is available via the POST /api/db/gencord endpoint described in the API Reference and the non-programmer guide above.

VUS Catalogue

The VUS Catalogue page provides predictor annotations for all HUMSAVAR variants classified as uncertain significance (US). Each variant is independently evaluated by AlphaMissense (AM), VESM-3B, and VESPA using its own calibrated substitution score, with a shared threshold of  ≥ 0.50 across all three predictors after empirical binning calibration (see Empirical Score Calibration). Score colours reflect the binary P/B call at this threshold rather than generic numerical ranges: red indicates a high-effect (P) call and green a low-effect (B) call. The consensus column records the pattern of calls across the three predictors: 3P indicates three high-effect calls; 2P · 1B and 1P · 2B indicate predictor disagreement; and 3B indicates three low-effect calls. Consensus filter buttons allow users to focus on concordant or discordant subsets, and the complete active view can be downloaded as TSV for further analysis. These predictor calls are computational annotations and should not be interpreted as clinical reclassification of HUMSAVAR variants.

Guided Examples

Use case 1 — Clinical variant interpretation (TP53, P04637)

This example walks through interpreting a known pathogenic variant in TP53 — one of the most clinically studied tumour suppressor genes — using the MissenseDB web interface. TP53 R248W is a well-characterised Li-Fraumeni syndrome variant (rs28934578) used here to illustrate the full workflow.

1 Navigate to the home page and type P04637 in the search box. Select the autocomplete suggestion and press Search.

2 The results page loads with the AlphaFold structure coloured by the calibrated AM residue score. The DNA-binding domain (residues ~100–290) appears in red — reflecting its high mutational intolerance — while the N-terminal transactivation domain is more tolerant.

3 In the Sequence Feature Viewer, locate the purple dbSNP track. Dense purple marks indicate positions with annotated clinical variants. Click position 248 in the sequence track. The 3D viewer highlights arginine 248, and the Residue Details panel shows all 19 substitution scores — R248W and R248Q both score ≥ 0.99 under AM (raw score), with calibrated probabilities of about 0.95 shown in brackets, consistent with their known pathogenicity.

4 Scroll to the dbSNP Variants table below the sequence viewer. Find position 248, substitution R→W, rsID rs28934578. The ACMG category shows LP/P. Click the row to expand the detail panel.

5 The detail panel shows population frequencies — gnomAD AF ≈ 0 confirms this variant is absent from population cohorts, consistent with its pathogenic classification. ClinVar submissions confirm pathogenic classifications across multiple submitters for Li-Fraumeni syndrome.

Figure 1. MissenseDB results page for TP53 (P04637) with position 248 selected. The structure viewer (left) is coloured by the calibrated AlphaMissense residue score. Residue Details (top right) shows all 19 substitution scores at R248 (raw AM scores all ≥0.996), each followed by its calibrated probability in brackets. Position summary reports calibrated AM=0.946, VESM=0.931, VESPA=0.933, Mean=0.937, with pLDDT 97. Sequence feature tracks show the Pfam P53 domain (blue), dense dbSNP annotations (purple), DSSP, and the calibrated AM, VESM, VESPA, Average and pLDDT tracks; the highlighted column marks position 248 within the uniformly high-scoring DNA-binding domain.

Figure 2. dbSNP Variants table and detail panel for P04637 around position 248. The table shows annotated substitutions with AM, VESM and VESPA raw scores, each followed by its calibrated probability in brackets, ACMG badges (LP/P in red, US in orange), and disease associations. The detail panel for rs121912651 shows population frequencies across gnomAD, ExAC and dbGaP, and 70 ClinVar submissions all classified pathogenic for Li-Fraumeni syndrome and hereditary cancer-predisposing syndrome.

Retrieving and constructing custom datasets through the API — a guide for non-programmers

Every panel on the MissenseDB results page is backed by a REST API endpoint. The same data the interface displays can be retrieved directly using curl from a terminal, or with requests in Python. Basic queries can be reproduced by copying the provided commands directly, while TSV output can be opened in standard spreadsheet and statistical software without additional parsing. Beyond the results page, the scan endpoint enables proteome-wide filtered dataset construction that is not available through the GUI at all.

The base URL for all API calls is https://biosig.lab.uq.edu.au/missense_db/. The examples below use TP53 (UniProt accession P04637) throughout. Database query endpoints support tab-separated output where indicated by appending ?format=tsv, suitable for opening directly in Excel, R, or other statistical software. The structure mapping and PDB download endpoints return JSON and PDB format respectively and do not support TSV output.

Sequence feature tracks (AM / VESM / VESPA / mean / pLDDT)

The area tracks shown in the sequence feature viewer — one per predictor plus the mean and pLDDT — correspond to the am_cal_load, vesm_cal_load, vespa_cal_load, mean_cal_load, and plddt columns in DB1 (the min-max *_load columns are returned alongside them). One row is returned per residue position. The same response also includes the structural context columns (dssp_ss, proximal_ligand, in_pfam, pfam_name) shown in the annotation tracks, and the has_pathogenic and worst_category flags shown in the dbSNP track.

Available structures and structure residue mapping

The mapping endpoint returns, for each UniProt residue position, the corresponding author residue number in that structure — this is what the GUI uses to synchronise the sequence track with the 3D viewer. The auth field gives the structure residue number; dssp and prox give the per-structure DSSP assignment and ligand proximity at that residue. Score-annotated PDB files (Panel 6 below) use these same author residue numbers, so the mapping response is the bridge between UniProt positions and structure coordinates.

Residue Details and Position Summary

Clicking a position in the GUI populates the Residue Details table (all 19 substitution scores for the selected predictor, each with its calibrated probability in brackets) and the Position Summary strip (calibrated per-predictor averages, pLDDT, structure residue number). Both come from the single-position endpoint. The am_scores, vesm_scores, and vespa_scores columns contain all 19 raw substitution scores in ALT:score format, pipe-separated, and the am_cal_scores, vesm_cal_scores and vespa_cal_scores columns contain the corresponding calibrated scores in the same format. The am_cal_load, vesm_cal_load, vespa_cal_load, mean_cal_load, and plddt fields populate the Position Summary strip.

HUMSAVAR Variants table

The variants table lists all HUMSAVAR-annotated substitutions for the protein with predictor scores, ACMG/AMP classification, population frequencies, and ClinVar data. Key columns: variant_category (LP/P, LB/B, US), am_load / vesm_load / vespa_load (position-level residue mutational-sensitivity scores, min-max normalised) with their calibrated counterparts am_cal_load / vesm_cal_load / vespa_cal_load / mean_cal_load, gnomad_af / topmed_af / thousandg_af (population frequencies), and clinvar_json (full per-submission ClinVar records). The bracketed calibrated probabilities shown alongside raw scores in the GUI are computed from the empirical binning calibration described in the Empirical Score Calibration section. The API returns position-level calibrated scores in the *_cal_load columns; per-substitution calibrated scores are available from the single-position endpoint in the *_cal_scores columns.

Score-annotated PDB download

For each structure, MissenseDB maps predictor scores into the B-factor column, producing one PDB file per predictor (AM, VESM, VESPA). The URL pattern is /missense_db/<uniprot_id>/<pdb_stem>/<predictor>. By default the B-factor column holds the calibrated residue score multiplied by 100 (0–100); the file is served as <pdb_stem>_calibrated.pdb. Appending ?scale=minmax returns the corresponding file in which the B-factor column holds the min-max normalised residue score instead. Downloaded files can be opened in PyMOL, ChimeraX, or VMD and coloured by B-factor to reproduce the MissenseDB score mapping. Residue numbers in the file correspond to the auth values returned by the mapping endpoint above.

Genomic coordinate lookup

Maps GRCh38 nucleotide positions to UniProt residues. Accepts up to 5,000 coordinates per request. Setting enrich: true adds the calibrated DB1 residue scores (*_cal_load), the min-max normalised scores (*_load, retained for backward compatibility), pLDDT, DSSP, and ligand proximity for each mapped residue. The browser page (/missense_db/gencord) displays and exports the calibrated scores only. Append ?format=tsv for tabular output.

The strand field uses P for plus strand and M for minus strand.

Proteome-wide filtered dataset construction (scan endpoint)

This functionality goes beyond the results page GUI and is the most powerful query mode in MissenseDB. The scan endpoint filters all 10.7 million positions across the proteome simultaneously by any combination of score threshold, secondary structure class, Pfam domain membership, ligand proximity, and clinical variant presence. Results are returned as paginated TSV for immediate download and analysis. The example below retrieves all 335,158 positions that are highly sensitive (mean calibrated score ≥ 0.8), inside a Pfam domain, and within 5 Å of a bound ligand — a dataset relevant to structure-guided variant prioritisation that is not retrievable from any per-protein or per-variant interface. Use offset=500, offset=1000 etc. to retrieve subsequent pages; available filter parameters are listed in the API Reference section.

API Reference

All DB-backed endpoints are available under /api/db/. Responses are JSON by default. Append ?format=tsv to any endpoint for tab-separated output suitable for command-line pipelines. The API is open with no authentication required.

Database Schemas

DB1 — Position Summary

One row per (UniProt ID, position). ~10.7 million rows covering all 19,436 proteins.

DB2 — HUMSAVAR + dbSNP

One row per annotated variant (UniProt ID, position, alt) with a matched dbSNP record. ~64,800 rows.

Licence

Contact:

Ashar J. Malik
ashar.malik@uq.edu.au
BioSig Lab · The University of Queensland