Why Score Calibration Matters
A missense annotation can come with several computational scores, each intended to summarize evidence about a possible protein effect. The numbers can look comparable: a score of 0.8 from one tool and 0.8 from another are both between zero and one. But the tools do not necessarily measure the same thing, and the numbers do not share a common probability scale. Interpreting them requires knowing what each score ranks, which direction indicates a stronger prediction, and what a threshold was designed to classify.
These predictions provide evidence about a possible coding effect; they do not establish that a variant causes disease, explains a genetic association, or is causal at a locus. Their meaning also depends on the variant being represented correctly and on which transcript and protein change are scored. As in the earlier material on transcript consequences and coding-variant ambiguity, one genomic variant may have different consequences across transcripts.
What the Common Scores Represent
The scores most often encountered for missense substitutions have different construction and interpretation. Before comparing values, record the tool and version, the score field, the reference genome and variant representation, and the transcript or protein change supplied to the tool. A tool-specific threshold is meaningful only in that setting.
| Score | Scale and direction | What a commonly used interpretation means |
|---|---|---|
| SIFT | 0 to 1; lower scores indicate a prediction of a less tolerated substitution. | A score at or below 0.05 is commonly called deleterious; above 0.05 is commonly called tolerated. This is a prediction label, not a probability of disease. |
| PolyPhen-2 | 0 to 1; higher scores indicate stronger prediction of a damaging protein effect. | Labels such as benign, possibly damaging, and probably damaging depend on the selected model and its thresholds. Do not assume one set of cutoffs applies to every run. |
| REVEL | 0 to 1; higher scores generally rank missense variants as more likely to be deleterious. | Values near or above 0.5 are often used as a screening threshold, but the chosen cutoff depends on the intended application and validation data. |
| CADD PHRED-like score | Higher values indicate a higher rank for predicted deleteriousness among possible substitutions. | A score of 10 corresponds approximately to the top 10%, 20 to the top 1%, and 30 to the top 0.1% of the ranked set—not to a 10%, 1%, or 0.1% probability of disease. |
SIFT uses sequence-based information to assess whether an amino-acid substitution is predicted to be tolerated. PolyPhen-2 uses sequence and protein-related features to classify possible effects. REVEL combines information from multiple prediction methods to rank missense variants. CADD integrates diverse annotations into a score that ranks variants by predicted deleteriousness. These brief descriptions identify the outputs’ general intent, not a guarantee that a particular variant has a measurable functional defect.
For a CADD PHRED-like score \(C\), the rank interpretation can be expressed as a transformed fraction:
This fraction is approximately the proportion of the ranked set at least as highly ranked, subject to the score’s ranking convention and reference set. It is not the probability that the variant is pathogenic. Raw CADD scores and PHRED-like scores are different fields; the transformation above applies to the PHRED-like scale.
Worked Example: Reading a CADD PHRED-Like Score
Suppose a hypothetical missense variant has a CADD PHRED-like score of 24. Interpret its rank scale without treating it as a clinical probability.
Using the rank-fraction relation:
This is about 0.398% of the ranked set, or roughly the top 0.4% under the score’s ranking convention. It means the variant receives a relatively high CADD rank among the variants in the relevant reference set. It does not mean there is a 24% or 99.6% probability that the variant is pathogenic, nor does it show that the variant causes a phenotype.
Thresholds Are Decision Rules, Not Natural Boundaries
A threshold turns a continuous score into a category for a particular purpose. For SIFT, the commonly used 0.05 cutoff gives a convenient predicted-deleterious versus predicted-tolerated label. It does not mark a biological discontinuity: scores of 0.049 and 0.051 are nearly identical numerical outputs even though they fall on opposite sides of that convention.
PolyPhen-2’s categories similarly depend on its selected model and classification thresholds. REVEL cutoffs are also chosen for a task; a value of 0.5 can be a useful screening rule in some settings, but it is not a universal boundary between harmless and harmful variation. CADD’s commonly quoted values describe rank ranges, not a classifier with a universal clinical cutoff. A threshold should therefore be reported with its tool, version or model, and purpose. It should not be presented as a universal rule for interpreting every variant.
A probability has a distinct interpretation. If a method were calibrated so that a prediction of 0.8 corresponded to an outcome frequency of 80% among comparable predictions, then 0.8 could be interpreted probabilistically for that outcome and setting. The usual SIFT, PolyPhen-2, REVEL, and CADD scores are not automatically calibrated in this way. Their numerical ranges alone do not justify probability language.
Worked Example: A Borderline SIFT Score
For a hypothetical missense substitution, the reported SIFT score is 0.048. The report uses the conventional SIFT threshold of 0.05.
Because \(0.048 \leq 0.05\), the substitution falls in the commonly used predicted deleterious category. This statement describes the SIFT classification under that threshold. It does not mean there is a 4.8% chance of a protein effect, or that the variant has been shown to cause disease.
If another variant has a SIFT score of 0.052, it falls just above the same cutoff and is commonly labelled predicted tolerated. The difference in labels should not be mistaken for strong evidence that the two variants have sharply different biological effects: their scores differ by only 0.004, and a threshold is a simplifying decision rule.
Compare Evidence Without Treating Scores as Votes
Different tools can give different predictions because they use different inputs, models, and target labels. Even when several tools point in the same direction, their outputs are not necessarily independent confirmations: some methods incorporate related or overlapping information. A careful interpretation describes the pattern and its limitations rather than counting how many tools cross a cutoff.
The direction of each scale must be checked before interpreting agreement. For example, a low SIFT score and a high REVEL score both point toward a more concerning prediction, whereas a high SIFT score and a high REVEL score do not. A PolyPhen-2 label should be read with its selected model and thresholds. A high CADD rank adds a ranking-based signal, but it does not resolve what the other tools predict about a particular protein change.
Worked Example: Mixed Predictions for One Missense Variant
A hypothetical missense substitution has SIFT \(=0.12\), PolyPhen-2 \(=0.78\), REVEL \(=0.42\), and CADD PHRED-like score \(=18\). Assume the listed values come from the stated tools and that the PolyPhen-2 model is identified in the analysis record.
SIFT \(=0.12\) is above the conventional 0.05 cutoff, so it is labelled predicted tolerated by that rule. PolyPhen-2 \(=0.78\) is higher on its 0-to-1 scale, but its category depends on the selected model’s thresholds; it should not be assigned a category without checking them. REVEL \(=0.42\) is below the often-used 0.5 screening cutoff. For CADD:
Thus, the CADD value corresponds approximately to the top 1.585%, or about 1.6%, of the ranked set. This is a relatively high rank, not a probability of pathogenicity. Taken together, these outputs are mixed rather than a unanimous strong prediction. The appropriate conclusion is to report the tool-specific results and avoid converting them into a definitive functional or clinical claim.
Frequency and Transcript Context Change the Question
Allele frequency and in silico score describe different properties. Frequency estimates how common an allele is in a specified population sample; a coding-effect score estimates or ranks a predicted molecular effect. A rare variant is not necessarily damaging, and a common variant is not necessarily functionally inert. Frequency can be relevant to whether a variant is compatible with a particular disease model, but that requires disease-specific assumptions about prevalence, penetrance, inheritance, and population sampling. It cannot be decided from the score alone.
Transcript context is equally important. A score for a missense change refers to a particular amino-acid substitution in a particular protein sequence. If a variant is missense in one transcript but synonymous, absent, or outside the coding sequence in another, the score does not describe every transcript equally. Confirm that the genome build, alleles, codon, protein change, and transcript identifier match the intended consequence. A score cannot repair an incorrect consequence assignment.
Worked Example: Different Frequencies and Transcript Contexts
Consider two hypothetical variants from the same locus. Variant A has estimated allele frequency \(0.0002\) in a specified reference population and is missense in the protein-coding transcript being assessed. Its SIFT score is 0.02 and its REVEL score is 0.82. Variant B has estimated allele frequency \(0.08\); it is missense in a minor transcript but does not alter the protein in the transcript used for the main analysis. Its REVEL score, computed for the minor-transcript protein change, is 0.71.
The frequency difference is:
In that reference population, variant B’s estimated frequency is 400 times variant A’s. Variant A’s SIFT score is below 0.05, and both variants have REVEL values above the commonly used 0.5 screening threshold. Those threshold comparisons describe predictions for their respective protein substitutions; they do not make the variants equivalent.
The frequency contrast and transcript distinction must be reported separately. Variant B’s score applies to the minor-transcript protein change, not automatically to the main-analysis transcript. Variant A’s rarity does not establish that it is damaging, and variant B’s higher frequency does not by itself prove that its minor-transcript effect is irrelevant. The population, transcript, disease model, and prediction task determine what additional interpretation is justified.
A Practical Interpretation Sequence
A disciplined reading prevents a plausible-looking score from becoming an unsupported conclusion. Before describing a score as evidence for a coding effect, check the variant representation and the score’s meaning. Then keep the prediction distinct from frequency, association evidence, and clinical interpretation.
Check genome build, reference and alternate alleles, transcript identifier, codon, and protein change. Confirm that the score was calculated for that consequence.
Distinguish, for example, CADD PHRED-like values from raw values; note that low SIFT and high PolyPhen-2 or REVEL values point toward stronger predicted impact.
Name the tool, model or version when available, and explain that the threshold is a classification or screening convention, not a universal biological boundary.
Describe computational predictions, allele frequency, and transcript context on their own terms. Do not turn one into a probability or conclusion about another.
Common Mistakes and Why They Matter
- Reading a score as a probability. “REVEL is 0.8, so the variant has an 80% chance of being pathogenic” is not justified by the score’s range. Report it as a REVEL score and use probability language only when calibration for the specified outcome and setting has been demonstrated.
- Using one direction for every scale. A low SIFT score and a high REVEL score can both suggest stronger predicted impact. Check each tool’s direction rather than interpreting “higher is worse” by default.
- Treating a cutoff as proof. Crossing 0.05 for SIFT or 0.5 for REVEL gives a tool-specific classification under a chosen rule. It does not establish a biological effect or disease causation.
- Equating rare with damaging. Low frequency and a high predicted-effect score are distinct observations. Neither alone establishes that a variant explains an association or phenotype.
- Ignoring the transcript behind the score. A prediction for one protein isoform may not apply to another. Verify the transcript and protein consequence before comparing variants or reporting a score.
- Counting correlated predictions as independent evidence. Agreement among tools can be informative, but it is not equivalent to several independent experiments. State what the tools predict without multiplying or summing their scores as though they were independent probabilities.
Check Your Understanding
Answer each question by distinguishing the score’s scale from the biological or clinical conclusion.
- A missense variant has SIFT \(=0.03\). What is its conventional SIFT classification, and what does that classification not establish?
- A CADD PHRED-like score is 20. What rank interpretation is commonly associated with this value, and why is it not a 20% probability of disease?
- Why is a REVEL score of 0.8 not automatically an 80% probability that a variant is pathogenic?
- Two variants have similar REVEL scores, but one is rare and missense in the transcript being studied while the other is more frequent and missense only in a minor transcript. Name two reasons not to treat them as equivalent evidence.
- Why can two tools that both predict a damaging effect not necessarily be counted as independent confirmations?