For human mutations that are predicted to result in an amino acid substitution we provide SIFT and PolyPhen predictions for the effect of this substitution on protein function. We compute the predictions for each of these tools for all possible single amino acid substitutions in the Ensembl human proteome. This means we can provide predictions for novel mutations for VEP and API users. We were able to compute predictions from at least one tool for over 95% of the human proteins in Ensembl. SIFT predictions are also available for chicken, cow, dog, horse, mouse, pig, rat sheep and zebrafish.

These tools are developed by external groups and we provide a brief explanation of the approach each takes below, and how we run it in Ensembl. For much more detail please see the representative papers listed below, and the relevant publications available on each tool's website.

SIFT

SIFT predicts whether an amino acid substitution is likely to affect protein function based on sequence homology and the physico-chemical similarity between the alternate amino acids. The data we provide for each amino acid substitution is a score and a qualitative prediction (either 'tolerated' or 'deleterious'). The score is the normalized probability that the amino acid change is tolerated so scores nearer 0 are more likely to be deleterious. The qualitative prediction is derived from this score such that substitutions with a score < 0.05 are called 'deleterious' and all others are called 'tolerated'.

We ran SIFT version 5.2.2 following the instructions from the authors and used SIFT to choose homologous proteins rather than supplying them ourselves. We used all protein sequences available from UniRef90 (release 2014_11) as the protein database.

List of species with SIFT data:

• Bos taurus
• Canis familiaris
• Danio rerio
• Equus caballus

• Gallus gallus
• Homo sapiens
• Mus musculus
• Ovis aries

• Rattus norvegicus
• Sus scrofa

 

PolyPhen

PolyPhen-2 predicts the effect of an amino acid substitution on the structure and function of a protein using sequence homology, Pfam annotations, 3D structures from PDB where available, and a number of other databases and tools (including DSSP, ncoils etc.). As with SIFT, for each amino acid substitution where we have been able to calculate a prediction, we provide both a qualitative prediction (one of 'probably damaging', 'possibly damaging', 'benign' or 'unknown') and a score. The PolyPhen score represents the probability that a substitution is damaging, so values nearer 1 are more confidently predicted to be deleterious (note that this the opposite to SIFT). The qualitative prediction is based on the False Positive Rate of the classifier model used to make the predictions.

We ran PolyPhen-2 version 2.2.2, release 405c (available here) following all instructions from the authors and using the UniProtKB UniRef100 (release 2013_10) non-redundant protein set as the protein database and DSSP (snapshot 22-Nov-2013) and PDB (snapshot 22-Nov-2013) as the structural databases. When computing the predictions we store results for the classifier models trained on the HumDiv and HumVar datasets. Both result sets are available through the variation API which defaults to HumVar if no selection is made. (Please refer to the PolyPhen website or publications for more details of the classification system).

List of species with PolyPhen data:

• Homo sapiens

 

Inconsistent predictions

PolyPhen and SIFT results are heavily dependant on sequence conservation estimates derived from protein sequence alignments and using different versions of the protein databases can result in substantial variance in the predictions and scores obtained. Ensembl releases are not synchronised with updates to the SIFT and PolyPhen web servers, so differences are to be expected. Changes in calls will also be seen between some Ensembl releases due to updates in the software and databases used.

 

Prediction data format

The SIFT and PolyPhen predictions are precomputed and stored in the variation databases and predictions are accessible in the variation API by using the sift_prediction, sift_score, polyphen_prediction and polyphen_score methods on a Bio::EnsEMBL::Variation::TranscriptVariationAllele object. For users wanting to access the complete set of predictions from the MySQL database or API, an explanation of the format used is provided here.

The predictions and associated scores are stored as a matrix, with a column for each possible alternate amino acid and a row for each position in the translation. Each prediction for a position and amino acid is stored as a 2-byte value which encodes both the qualitative prediction and score encoded as described below. A simple example matrix is shown in the figure below, though note we only show the decoded score rather than the actual binary value stored in the database.

Prediction matrices can be fetched and manipulated in a user-friendly manner using the variation API, specifically using the ProteinFunctionPredictionMatrixAdaptor which allows you to fetch a prediction matrix using either a transcript or a translation stable ID. This adaptor returns a ProteinFunctionPredictionMatrix object and you can use the get_prediction method to retrieve a prediction for a given position and amino acid. Users who want to use entire matrices should use the deserialize method to decode an entire binary formatted matrix into a simple Perl hash. Please refer to the API documentation for both of these classes for more details. For users who require direct access to the MySQL database (for instance because they are accessing the database in a language other than Perl) we provide a description of the binary format used below.

Prediction matrices for each translation from each of SIFT and PolyPhen are stored in the protein_function_predictions table. The analysis used to calculate the predictions is identified in the analysis_attrib_id column which refers to a value stored in the attrib table, and the protein sequence the predictions apply to is identified by the translation_md5_id column which refers to a hexadecimal MD5 hash of the sequence stored in the translation_md5 table. The prediction matrices are stored in the prediction_matrix column as a gzip compressed binary string. Once uncompressed, this string contains a 40 byte substring for each row in the matrix concatenated together in position order. Each row is composed of 20 2-byte predictions, one for each possible alternative amino acid in alphabetical order, though note that the value for the amino acid that matches the reference amino acid is identified as a 2-byte value with all bits set, or 0xFFFF in hexadecimal notation. The prediction itself is packed as a 16 bit little-endian ("VAX" order, or "v" format if using the perl pack subroutine) unsigned short value. The top 2 bits of this short are used to encode the qualitative prediction (PolyPhen has 4 possible values, while SIFT has just 2), and the bottom 10 bits are used to encode the prediction score. To decode the qualitative prediction you should mask off all bits except the top 2, and shift the resulting short right by 14 bits and treat this as an integer between 0 and 3. The corresponding prediction can then be looked up in the table below. To decode the prediction score you should mask off the top 6 bits and the resulting value can be treated as a number between 0 and 1000, which should be divided by 1000 to give a 3 decimal place score (casting to a floating point type if necessary). Bits 11-14 are not used, except to encode the "same as reference" dummy prediction 0xFFFF.

A diagram of the encoding scheme is shown above. In this example from a polyphen prediction, the top prediction bits are 0b01 which in decimal is the number 1, which corresponds to "possibly damaging" in the table below. The score bits are 0b1110001010 which in decimal is the number 906, which when divided by 1000, gives a score of 0.906.

Tool	Numerical value	Qualitative prediction
PolyPhen	0	"probably damaging"
PolyPhen	1	"possibly damaging"
PolyPhen	2	"benign"
PolyPhen	3	"unknown"
SIFT	0	"tolerated"
SIFT	1	"deleterious"

To retrieve a prediction for a particular amino acid substitution at a given position in a translation, first you must retrieve the binary matrix from the database and uncompress it using gzip. You can calculate the offset into this string by multiplying the desired position (starting at 0) by 20 and then adding the index of the desired amino acid in an alphabetical ordering of amino acids (also starting at 0), and then multiply this value by 2 to take into account the fact that each prediction uses 2 bytes. Each matrix also includes a 3 byte header used check if the data is corrupted etc. so you will also need to add 3 to the calculated offset. The 2 bytes at the calculated position can then be extracted and the approach described above can be used to retrieve the prediction and score. A perl implementation of this scheme can be found in the Bio::EnsEMBL::Variation::ProteinFunctionPredictionMatrix module in the variation API and should hopefully inform attempts to reimplement this scheme in other languages.