Broadening Pfam Protein Sequence Annotations

Pfam is a database of conserved protein families or domains commonly used for genome annotation and sequence classification. It comprises two parts: (1) Pfam-A families, which are fully annotated and consist of a representative seed alignment, HMMs, and a full alignment comprising all sequences that score above the curated threshold; (2) Pfam-B families, which are automatically generated clusters of domains not matched by Pfam-A but that often indicate conserved sequence regions. Pfam release 23.0 predicts at least one Pfam-A domain on 74% of the sequences in UniProtKB, and predicts either a Pfam-A or Pfam-B domain on 93% of the sequences in UniProtKB.

With the ever increasing rate of deposition of new proteins of all qualities into the underlying repositories, it is essential that Pfam continues to grow in order to maintain its coverage. We have used a number of strategies to improve the annotation provided by Pfam, and these include both building new families and expanding existing ones. Pfam has also greatly benefited from contributions from its user community. New family and functional annotation submissions from an _S. pombe_ curator have ensured that Pfam has a high coverage - 83% - of the _S. pombe_ proteome. 

Many of the early Pfam-A models have not been altered since they were first deposited. As the diversity of the sequence databases grows, the diversity within a Pfam seed alignment can become too narrow for representing the breadth of sequences that should belong to that family. The result is that some of the early Pfam-A HMMs fail to detect remote homologues. To address this problem we have rebuilt a large proportion of Pfam-A families, which has increased the Pfam-A coverage by 1-2%. Another strategy we have used has been that of targeted building, where a particular system or complex is examined in detail to ensure families exist for all components and annotation is consistent. In terms of building new Pfam-A families, the two major starting points are Pfam-B clusters and novel structures. From these we have built ~1000 families between releases 22.0 and 23.0, and a further 800 families since release 23.0.

Between Pfam releases 22.0 and 23.0 we have changed the the way in which Pfam-B families are generated. Previously, Pfam-B families were created from PRODOM clusters that were based on a much smaller sequence database than the one upon which Pfam was built. We now use the ADDA algorithm that generates clusters from the same underlying sequence database as Pfam is based on, thus resulting in a more comprehensive Pfam-B contribution. This has increased the sequence coverage contributed by Pfam-B substantially from 3.9% to 11.8%. 

In a further drive to improve coverage, Pfam is currently evaluating a new release of the HMMER software (HMMER3) used to construct and search the Pfam HMMs. Early results show that HMMER3 is ~100 fold faster and has increased specificity and sensitivity compared with HMMER2. 
&#xa