New scoring method

We’ve changed quite a few things for this release. For a start, we’ve implemented a new scoring method that assigns every query protein a single overall score, rating its overall functional similarity to the reference set. This is calculated by forming an ‘optimal score profile’, a vector with one element in it for each predictor in the query:

So each query protein has a cell for each predictor, which will contain a p-value.

Then, we populate each cell with the best prediction (lowest p-value) from that predictor, between the query protein and any reference protein, leaving blanks where a predictor has made no predictions at all about a given query protein. For example:

Finally, using Fisher’s method (unweighted), we calculate an overall score for each query protein by combining these p-values:

The higher this score is, the more closely associated the query protein is with the biological function performed by the reference proteins.

The web interface now uses this method to score and rank the results, and the front-end web service supports an operation called RetrieveSetwiseScores so you can use it in your scripts.

More new features in web interface

We’ve also added a feature that lets you select a GO term if you want to supply a reference set automatically. This queries the EBI’s web services for all the terms matching your search string, and when you select one, it retrieves all the human proteins associated with that term or any of its descendants by high-quality annotations (IDA, IEP, IGI, IMP, RCA, TAS). We would be grateful for any feedback on using this new experimental feature.

Follow us on Twitter

We now have a Twitter account where regular status updates appear: http://twitter.com/FuncNet

We hope to bring them back online by around 10:00 tomorrow morning.

Apologies for any inconvenience.

This supplies the overall Fisher’s Method score for each protein pair, as with RetrievePairwiseScores, but also gives you the scores from each predictor from which the overall scores were calculated. This enables service consumers to integrate FuncNet results with their own analysis methods at a finer level of granularity, and allows for filtering of predictors, score integration using alternative methods, etc.

When interpreting the results, remember that only the p-value for each prediction is comparable between predictors. The raw scores scales and distributions vary between predictors, and are provided for informational purposes only, and for complete transparency in case you want to train a classifier on them for example. Remember too that not all predictors actually yield a single raw score, in which case this field will be set to zero.

See the example queries page for more details on this query.