An expanded evaluation of protein function prediction methods shows an improvement in accuracy

Abstract

Background: A major bottleneck in our understanding of the molecular underpinnings of life is the assignment of function to proteins. While molecular experiments provide the most reliable annotation of proteins, their relatively low throughput and restricted purview have led to an increasing role for computational function prediction. However, assessing methods for protein function prediction and tracking progress in the field remain challenging.

Results: We conducted the second critical assessment of functional annotation (CAFA), a timed challenge to assess computational methods that automatically assign protein function. We evaluated 126 methods from 56 research groups for their ability to predict biological functions using Gene Ontology and gene-disease associations using Human Phenotype Ontology on a set of 3681 proteins from 18 species. CAFA2 featured expanded analysis compared with CAFA1, with regards to data set size, variety, and assessment metrics. To review progress in the field, the analysis compared the best methods from CAFA1 to those of CAFA2.

Conclusions: The top-performing methods in CAFA2 outperformed those from CAFA1. This increased accuracy can be attributed to a combination of the growing number of experimental annotations and improved methods for function prediction. The assessment also revealed that the definition of top-performing algorithms is ontology specific, that different performance metrics can be used to probe the nature of accurate predictions, and the relative diversity of predictions in the biological process and human phenotype ontologies. While there was methodological improvement between CAFA1 and CAFA2, the interpretation of results and usefulness of individual methods remain context-dependent.

Keywords: Disease gene prioritization; Protein function prediction.

Fig. 1

Time line for the CAFA2 experiment

Fig. 2

CAFA2 benchmark breakdown. a The benchmark size for each of the four ontologies. b Breakdown of benchmarks for both types over 11 species (with no less than 15 proteins) sorted according to the total number of benchmark proteins. For both panels, dark colors (blue, red, and yellow) correspond to no-knowledge (NK) types, while their light color counterparts correspond to limited-knowledge (LK) types. The distributions of information contents corresponding to the benchmark sets are shown in Additional file 1. The size of CAFA 1 benchmarks are shown in gray. BPO Biological Process Ontology, CCO Cellular Component Ontology, HPO Human Phenotype Ontology, LK limited-knowledge, MFO Molecular Function Ontology, NK no-knowledge

Fig. 3

CAFA1 versus CAFA2 (top methods). A comparison in F _max between the top-five CAFA1 models against the top-five CAFA2 models. Colored boxes encode the results such that (1) the colors indicate margins of a CAFA2 method over a CAFA1 method in F _max and (2) the numbers in the box indicate the percentage of wins. For both the Molecular Function Ontology (a) and Biological Process Ontology (b) results: A CAFA1 top-five models (rows, from top to bottom) against CAFA2 top-five models (columns, from left to right). B Comparison of Naïve baselines trained respectively on SwissProt2011 and SwissProt2014. C Comparison of BLAST baselines trained on SwissProt2011 and SwissProt2014

Fig. 4

Overall evaluation using the maximum F measure, F _max. Evaluation was carried out on no-knowledge benchmark sequences in the full mode. The coverage of each method is shown within its performance bar. A perfect predictor would be characterized with F _max=1. Confidence intervals (95 %) were determined using bootstrapping with 10,000 iterations on the set of benchmark sequences. For cases in which a principal investigator participated in multiple teams, the results of only the best-scoring method are presented. Details for all methods are provided in Additional file 1

Fig. 5

Precision–recall curves for top-performing methods. Evaluation was carried out on no-knowledge benchmark sequences in the full mode. A perfect predictor would be characterized with F _max=1, which corresponds to the point (1,1) in the precision–recall plane. For cases in which a principal investigator participated in multiple teams, the results of only the best-scoring method are presented

Fig. 6

Overall evaluation using the minimum semantic distance, S _min. Evaluation was carried out on no-knowledge benchmark sequences in the full mode. The coverage of each method is shown within its performance bar. A perfect predictor would be characterized with S _min=0. Confidence intervals (95 %) were determined using bootstrapping with 10,000 iterations on the set of benchmark sequences. For cases in which a principal investigator participated in multiple teams, the results of only the best-scoring method are presented. Details for all methods are provided in Additional file 1

Fig. 7

Overall evaluation using the averaged AUC over terms with no less than ten positive annotations. The evaluation was carried out on no-knowledge benchmark sequences in the full mode. Error bars indicate the standard error in averaging AUC over terms for each method. For cases in which a principal investigator participated in multiple teams, the results of only the best-scoring method are presented. Details for all methods are provided in Additional file 1. AUC receiver operating characteristic curve

Fig. 8

Averaged AUC per term for Human Phenotype Ontology. a Terms are sorted based on AUC. The dashed red line indicates the performance of the Naïve method. b The top-ten accurately predicted terms without overlapping ancestors (except for the root). AUC receiver operating characteristic curve

Fig. 9

Performance evaluation using the maximum F measure, F _max, on eukaryotic (left) versus prokaryotic (right) benchmark sequences. The evaluation was carried out on no-knowledge benchmark sequences in the full mode. The coverage of each method is shown within its performance bar. Confidence intervals (95 %) were determined using bootstrapping with 10,000 iterations on the set of benchmark sequences. For cases in which a principal investigator participated in multiple teams, the results of only the best-scoring method are presented. Details for all methods are provided in Additional file 1

Fig. 10

Similarity network of participating methods for BPO. Similarities are computed as Pearson’s correlation coefficient between methods, with a 0.75 cutoff for illustration purposes. A unique color is assigned to all methods submitted under the same principal investigator. Not evaluated (organizers’) methods are shown in triangles, while benchmark methods (Naïve and BLAST) are shown in squares. The top-ten methods are highlighted with enlarged nodes and circled in red. The edge width indicates the strength of similarity. Nodes are labeled with the name of the methods followed by “-team(model)” if multiple teams/models were submitted

Fig. 11

Case study on the human ADAM-TS12 gene. Biological process terms associated with ADAM-TS12 gene in the union of the three databases by September 2014. The entire functional annotation of ADAM-TS12 consists of 89 terms, 28 of which are shown. Twelve terms, marked in green, are leaf terms. This directed acyclic graph was treated as ground truth in the CAFA2 assessment. Solid black lines provide direct “is a” or “part of” relationships between terms, while gray lines mark indirect relationships (that is, some terms were not drawn in this picture). Predicted terms of the top-five methods and two baseline methods were picked at their optimal F _max threshold. Over-predicted terms are not shown