How to use these predictions

Evidence levels

Every variant gets an evidence level: Benign strong, moderate or supporting, Indeterminate, or Pathogenic supporting, moderate or strong. Missense variants also get the matching ACMG/AMP code, such as PP3_Moderate or BP4_Supporting.

The levels were set with the method ClinGen recommends for prediction tools (Pejaver et al. 2022). For each score we estimate how much more often it turns up among pathogenic variants than among benign ones, a likelihood ratio, assuming that 4.4% of rare variants are pathogenic. Each cut-off sits at the cautious end of a 95% confidence interval, so a variant only reaches a level when the data clearly support it.

Missense variants and other variant types were calibrated separately, because the same score means different things for each. The With AlphaMissense version has its own calibration for the missense variants that AlphaMissense covers. These are the score cut-offs:

No variant type reached Pathogenic very strong. Missense variants reach moderate pathogenic evidence at most: the strong level needs more labelled missense variants than the roughly 2,000 we could test against.

How the levels held up on new genes

We set the levels using half of the genes and checked them on the other half, then swapped the halves. For the labelled variants in the test genes, the table shows how many at each level were pathogenic and the likelihood ratio that implies.

Releases 1.0.x used fixed odds thresholds (Tavtigian et al. 2018) without the confidence interval. In the same test, their top level, Pathogenic very strong, only performed at the strong level. Each variant's earlier level is kept in the evidence_v1 column of the download files.

The raw score

The raw score runs from 0 to 1. It comes from a model trained on a set of variants that was 45% pathogenic. Most variants in a gene panel are benign, so a cut-off of 0.5 calls far too many of them pathogenic.

Use the evidence level when classifying a variant. The raw score is on the gene pages and in the files for reference and for reporting.

Two versions of the data

The data come in two versions. They differ in whether AlphaMissense was used, and in how you may reuse them. Use the switch at the top of a gene page to move between them.

Our model only (files named native)

Every variant, scored by our model alone, with no third-party predictions. Licence CC BY 4.0: you can use it for anything, including commercial work, as long as you credit it.

With AlphaMissense (files named routed)

For the 177,464 missense variants that AlphaMissense covers, our model is combined with AlphaMissense. All other variants keep our model's score. This version is more accurate on missense variants. Because AlphaMissense is licensed CC BY-NC-SA 4.0, this version carries the same licence: non-commercial use only, and anything you build from it must be shared under the same terms.

If you reuse the data, keep the two versions in separate files so the licences stay clear.

Flags

Some rows in the gene tables carry a flag under the variant name.

Seen in training

The model was trained on this variant's ClinVar classification, so its prediction is not an independent test. Leave these 147,162 variants out of any accuracy estimate. You can hide them with the checkbox above each table.

Class-level score

We could not place this missense change on the UniProt protein sequence, usually because it changes more than one DNA base or because ClinVar numbers it on a different transcript. It was scored by a simpler model that only knows the variant is missense. Its evidence is weak, and it has no activity or mechanism value. This applies to 2,852 variants.

AlphaMissense or No AlphaMissense

Shown in the With AlphaMissense version. It tells you whether AlphaMissense covers the variant. Where it does not, the score is our model's alone. In the download file, ignore the v2_plus column for these rows.

Layers in-sample

This protein has a deep mutational scan that was used to train the activity and mechanism predictions, so those two values are not independent for it.

Protein map

The line is the protein chain, from the first amino acid on the left to the last on the right. Where UniProt marks a helix that crosses the membrane, the chain crosses the yellow band. Loops above the band are outside the cell, and loops below it are in the cytoplasm.

The colour at each position is the average prediction for ClinVar missense variants at and near that position. Red means variants there tend to be damaging. Blue means they tend to be tolerated. Grey means there are too few variants nearby to say. Only variants reported in ClinVar are counted, so a grey stretch is not necessarily unimportant.

The 3D view shows the AlphaFold model of the protein, coloured the same way but averaged over residues that are close together in space. Dots on the map and spheres in 3D are individual ClinVar variants, coloured by evidence level.

Click any position on the map to see which part of the protein it is. The helix or loop it belongs to is shaded gold on the map and drawn in gold on the 3D structure, with the rest of the protein in grey, and the view turns to face it. The table then shows only the variants at that position. Clicking a region in the list beside the 3D view does the same for that region. Use “Clear selection” to go back, and “Reset view” to recentre the structure.

Drag across the map to zoom into a stretch of the chain. Double-click it to zoom back out.

Activity and mechanism

Predicted activity estimates how much of its normal transport or enzyme activity the protein keeps with a given missense change, from 0 (none) to 1 (normal). It was trained on deep mutational scans of 14 proteins. On proteins it had not seen, it agreed with measured activity at a Spearman correlation of 0.50, so treat it as a rough estimate.

Mechanism is exploratory. It is given only for variants with moderate or strong pathogenic evidence. It estimates whether the protein reaches its place in the cell but fails to work (activity-specific, values near 1), or never gets there (expression-driven, values near 0). Across proteins it separates the two with an AUC of 0.78. But on SLC19A2 and SLC19A3, two independent models gave almost unrelated answers (Spearman correlation 0.009), so a single prediction on a gene without lab data is unreliable. It can help you decide which variants to test in the lab. It should not be used to classify a variant. A few values fall just outside 0 to 1; the site shows them as 0 or 1.

Using this with ACMG/AMP

• For a missense variant, the acmg_code column gives the code to apply: PP3_Supporting, PP3_Moderate, BP4_Supporting, BP4_Moderate or BP4_Strong. Indeterminate variants have no code.
• Codes are not given for frameshift, nonsense, splice and other variant types. Those are judged with other criteria, such as PVS1 or the ClinGen splicing recommendations, and adding PP3 on top would count the same evidence twice. Their evidence level is still shown for reference.
• Use one prediction tool for PP3 and BP4. If you use these predictions, do not also apply AlphaMissense, REVEL or another predictor to the same variant. The With AlphaMissense version already includes AlphaMissense.
• Pejaver et al. (2022) recommend that PP3 and PM1 together count for no more than strong evidence, because both draw on conservation and protein structure.
• When you report a call, note the release number, which version of the data you used, and the raw score.

Accuracy and coverage

We tested the predictions on 7,039 labelled variants that were not used in training, across all 13 variant types.

Knowing only whether a variant is a frameshift, a synonymous change and so on already gives an AUC of 0.9588, because variant type says a lot about pathogenicity. That lookup table is the baseline to beat.

On the 1,998 missense variants in that set that AlphaMissense can also score:

On missense variants, our model on its own is less accurate than AlphaMissense. Combined with AlphaMissense, it is more accurate than either (DeLong test, p = 0.004).

Which variants AlphaMissense covers

AlphaMissense only scores missense variants. Here it covers 177,464 of them. The other variants, about half of all ClinVar entries in these genes, have no AlphaMissense score; this resource scores all of them.

Hatched part: covered by AlphaMissense

Limitations

• ClinVar classifications are partly based on computational predictions, so some circularity in the accuracy figures cannot be ruled out. The results within a single variant type are the safer reading.
• The evidence levels are cautious by design. Many did better than their label on held-out genes, and indeterminate variants still lean toward pathogenic; see Evidence levels.
• Missense variants reach moderate pathogenic evidence at most, because the calibration set holds only about 2,000 labelled missense variants.
• 2,852 missense variants have only a class-level score.
• Only variants in ClinVar are scored. Other possible changes at the same position are not.
• Mechanism predictions are exploratory.
• These are research predictions. They are not a clinical classification.

Columns in the files

Searching

The search box accepts a gene symbol (SLC19A2), part of a protein name (thiamine), an rsID (rs113993960), a ClinVar variation ID (5957), or an HGVS name, written either as NM_006996.3:c.515G>A or as SLC19A2 p.Gly172Asp.