Protein Function Prediction: Capabilities and Limitations of Alignment-based Approaches
Year: 2026
Type: preprint
Abstract: Computational prediction of protein function is increasingly dominated by deep learning architectures, often premised on the assumption that alignment-based methods leveraging sequence homology fail on highly divergent sequences. In this study, we reassess the capacity of alignmentbased annotation transfer, demonstrating that these approaches can surpass state-of-the-art deep learning models when provided with appropriate contextual data. We identify that alignment-based methods fail primarily due to a lack of annotations to transfer, an issue affecting up to 73% of low-homology proteins. To resolve this, we introduce a framework that integrates multi-template alignment, manually curated non-experimental labels, and STRING protein-protein interaction networks. This methodology largely reduces the fraction of proteins left unannotated and achieves superior performance across three benchmark datasets, including a custom low-homology dataset. Furthermore, a longitudinal analysis of database updates reveals that static benchmarks systematically underestimate predictive value; many alignment-based "false positives" are, in fact, early accurate predictions later confirmed by experimental updates. Our findings indicate that alignment-based approaches remain highly competitive and offer a robust, explainable baseline for future functional annotation developments.
Source: HAL
Language: English
Subfield: Molecular Biology
Domain: Life Sciences
Open Access status: green
OpenAlex ID: W7213738775