PolyPhen-2: How to Run It and Read the Output (HumDiv vs HumVar)

PolyPhen-2 (Polymorphism Phenotyping v2) is a free web server that predicts the possible impact of an amino acid substitution on the structure and function of a human protein, using physical and comparative considerations. It is meant for human non-synonymous SNPs (missense variants). This guide shows how to run a single query and a batch query, how to choose between the HumDiv and HumVar models, and how to read the output. PolyPhen-2 produces a computational prediction only; it is not a clinical interpretation, and it should be read alongside other evidence.

What PolyPhen-2 is used for in research

Researchers use PolyPhen-2 to prioritise missense variants found in sequencing or genotyping studies by estimating whether a substitution is likely to be damaging to the protein. The server’s own description states that it predicts the possible impact of an amino acid substitution on the structure and function of a human protein (PolyPhen-2 web server, genetics.bwh.harvard.edu/pph2, read 2026-11-08). The method was introduced by Adzhubei et al. in Nature Methods (2010) and is documented step by step in Adzhubei, Jordan and Sunyaev, “Predicting functional effect of human missense mutations using PolyPhen-2”, Current Protocols in Human Genetics (2013), Unit 7.20, doi:10.1002/0471142905.hg0720s76.

Variant frequency resources such as gnomAD and the 1000 Genomes Project are commonly consulted alongside a predictor, since a prediction says nothing about how common a variant is in a population.

Is PolyPhen the same as PolyPhen-2?

“PolyPhen-2” is the current server and the name you should search for and cite. The server is titled “Polymorphism Phenotyping v2”, and the version suffix is part of the name. When a paper or annotation tool says “PolyPhen”, check which version and which model (HumDiv or HumVar) it reports.

Access: official source only

PolyPhen-2 is run from the official web server at genetics.bwh.harvard.edu/pph2. No installation is needed for the web interface. The page links a batch query page in its navigation, and its support contact is listed on the server itself. Many annotation pipelines also report PolyPhen-2 scores; those downstream tools may apply their own score bands, so always confirm which labels and cutoffs the tool you use applies.

Step-by-step: run a single query

The query form on the server accepts the following fields (read 2026-11-08):

1. Open the PolyPhen-2 web server and use the query form.
2. In Protein or SNP identifier, enter a protein identifier or SNP identifier. Alternatively, paste a sequence into Protein sequence in FASTA format.
3. Enter the Position of the substitution in the protein.
4. Under Substitution, pick the wild-type residue (AA1) and the variant residue (AA2) from the lists.
5. Optionally add a Query description to label the job.
6. Optionally open Display advanced query options. These let you sort hits by identity, E-value or mapping to mismatch; request structural parameters and contacts for the first hit or all hits; and set the minimal alignment length, minimal identity, maximal gap length and contact threshold (in Angstroms).
7. Submit the query and wait for the results page.

Step-by-step: run a batch query

For many variants, use the batch query page linked from the server’s navigation. The 2013 protocol describes three input formats (Adzhubei et al. 2013):

• Protein substitutions: UniProt identifier, position, wild-type amino acid (AA1), variant amino acid (AA2).
• dbSNP identifiers (rsIDs).
• Genomic coordinates: chromosome:position, reference allele and alternative allele. Chromosome position coordinates are 1-based, and the default assembly is GRCh37/hg19.

Columns may be tab- or space-separated. Check the current batch page for any other accepted formats before relying on one not listed here. The protocol also describes a third use: searching precomputed predictions for the whole human exome space.

HumDiv or HumVar?

PolyPhen-2 provides two trained models. HumDiv is the default. According to the Current Protocols unit, HumDiv is preferred for evaluating rare alleles and for dense mapping of regions identified by genome-wide association studies (see our GWAS guide), while HumVar is better suited for diagnostics of Mendelian diseases. Choose the model that matches your question and report which one you used, since the two give different scores for the same variant.

Output and how to read it

The key score is pph2_prob, described in the protocol as the classifier probability of the variation being damaging. The server turns it into a ternary classification:

Label	Meaning in the protocol
Probably damaging	At or below the lower false positive rate (FPR) cutoff; the more confident call.
Possibly damaging	Between the lower and higher FPR cutoffs.
Benign	Above the higher FPR cutoff.
Unknown	Insufficient data to make a prediction.

The FPR thresholds are 5% and 10% for HumDiv and 10% and 20% for HumVar (Adzhubei et al. 2013). Because the labels are defined by FPR thresholds that differ by model, the numeric score at which a variant changes label is not one fixed number. Avoid quoting a universal score cutoff; report the label together with the score and the model.

Which batch output file to read

• Short: 11 columns, with the prediction, probability, FPR and TPR. Start here.
• Full: 56 columns, adding structural, conservation and sequence features.
• SNPs: 38 columns of MapSNPs annotations.
• Logs: job messages, useful when variants fail to map.

Troubleshooting

• “Unknown” predictions: the protocol defines this as insufficient data; it is not a benign call.
• Coordinates do not map: genomic input is 1-based and defaults to GRCh37/hg19. Coordinates from another assembly will not match; check your assembly first and consult the Logs file.
• Different results in different tools: annotation tools may report the HumDiv or HumVar score and apply their own bands. Compare like with like.

Free alternatives and complements

Predictors answer the same question with different methods, and agreement among several is often reported alongside a PolyPhen-2 call. The table lists each tool’s primary publication.

Tool	Primary publication
SIFT	Ng PC, Henikoff S. Nucleic Acids Res 2003;31(13):3812-3814
CADD	Kircher M et al. Nat Genet 2014;46:310-315
REVEL	Ioannidis NM et al. Am J Hum Genet 2016;99:877-885
AlphaMissense	Cheng J et al. Science 2023;381:eadg7492

Check each tool’s own site for current access terms before use. Our related guide to DAVID covers a different step, functional annotation of gene lists, and computational reproducibility explains why model, assembly and date should be recorded with every prediction.

FAQ

What does a PolyPhen-2 score mean?

The score pph2_prob is the classifier probability that the substitution is damaging. The label (probably damaging, possibly damaging, benign, unknown) is assigned from FPR thresholds that depend on whether HumDiv or HumVar was used.

Should I use HumDiv or HumVar?

HumDiv is the default and is preferred for rare alleles and dense mapping of GWAS regions; HumVar is better suited to diagnostics of Mendelian disease (Adzhubei et al. 2013).

Can PolyPhen-2 diagnose a disease or tell me a variant is pathogenic?

No. It is a computational prediction of the effect of an amino acid substitution. It does not replace clinical interpretation, and this guide gives no medical advice.

Does PolyPhen-2 accept VCF files?

The batch input formats documented in the 2013 protocol are protein substitutions, rsIDs and genomic coordinates; VCF is not among them.

Sources

• PolyPhen-2 web server, genetics.bwh.harvard.edu/pph2, read 2026-11-08.
• Adzhubei I, Jordan DM, Sunyaev SR. Predicting functional effect of human missense mutations using PolyPhen-2. Curr Protoc Hum Genet 2013, Unit 7.20. PMC4480630, read 2026-11-08.
• Adzhubei IA et al. A method and server for predicting damaging missense mutations. Nat Methods 2010;7:248-249.