The promise of AlphaFold for gene structure annotation.

Most new genomes lack annotation, automated methods are error-prone, and few genomes are ever manually curated due to time and cost. Protein structure predictions may offer a new route to assess and improve gene models without requiring experimental data. Here, we explore whether scores from protein structure prediction can aid in scoring gene model quality. We chose three species (Fusarium graminearum, Toxoplasma gondii, and Aspergillus fumigatus) from the VEuPathDB database that have collectively undergone more than 1000 manual curation events. We modelled translations of the gene models with AlphaFold 3, before and after curation, collecting various scores. Then we carried out structure searching of the PDB with Foldseek and sequence-based domain identification using InterProScan. We profiled the scores produced by these methods to identify those best for gene model assessment. AlphaFold 3 scores strongly favoured manually improved over pre-improvement gene models, supporting 65-84% of manually-curated changes. Combining scores across multiple tools (AlphaFold 3, Foldseek and InterProScan) provided further improvements in model scoring. Overall, the most discriminative scores combined the outputs of AlphaFold 3 and Foldseek. Importantly, we find that scores from the much faster Protenix-Mini retain the same discriminatory power as those from AlphaFold 3. Our results, therefore, highlight the potential of scores derived from deep learning-based protein structure prediction for scoring gene models in the absence of experimental data.