Combining natural language processing and graph learning for disease-gene prediction:

Background:With the growing availability of structured genetic association data, in-silico disease-gene prediction offers a scalable approach to identify novel disease-relevant genes. This study evaluates whether graph learning, combined with natural language processing (NLP) techniques, can support the interpretation of newly identified variants and enable the prioritization of novel candidate genes. Material and Methods: A heterogeneous graph was built using disease-gene associations from the GWAS Catalog and Open Targets GWAS associations. Gene-gene edges were derived from STRING interactions. A pre-trained SapBERT model generated semantic embeddings from Experimental Factor Ontology (EFO) terms, and cosine similarities were used to construct trait-trait edges. Metapath2Vec was applied to learn node representations, and the resulting embedding similarities were used to rank genes for each trait. Performance was evaluated using non-GWAS Open Targets associations as true positives. An ablation model without trait-trait edges was also tested to assess the contribution of NLP-based similarities. Results: Table 1 reports results averaged across 1,687 EFO traits. Being a ranking task, metrics reflect ranking performance. Reciprocal rank (RR) is the inverse rank of the first true positive. Hit@k equals 1 if at least one true positive appears within the top k positions, and Prec@k is the proportion of true positives within the top k. Conclusion: The defined approach achieved promising predictive performances, and the ablation study confirmed the benefit of trait-trait similarities (p-values < 0.0001). The metric variability is being addressed through ongoing optimization.