PATHOS: Predicting variant pathogenicity by combining protein language models and biological features

• Our analysis reveals that fine-tuning Protein Language Models does not improve performance on pathogenicity prediction compared to using original embeddings. • PATHOS integrates ESMC 600M and Ankh2 Large Protein Language Models with biological features for optimal pathogenicity prediction. • It outperforms 65 state-of-the-art predictors, achieving a record MCC of 0.591 on clinical data. • High accuracy and robust generalization are maintained on new sequences absent from the training dataset. • A user-friendly web server provides 140 million precomputed predictions for the human proteome. Predicting the pathogenic impact of missense variants is essential for understanding and diagnosing genetic diseases. These approaches have undergone significant evolution, with the latest methodologies based on deep learning approaches. Nonetheless, only a limited number use the potential of Protein Language Models (PLMs), which have demonstrated strong performance across various protein-related tasks. A new predictor, called PATHOS, was developed; it combines embeddings from an optimal set of two PLMs, namely ESM C 600M and Ankh 2 Large. Their embeddings were combined with additional crucial biological features such as phylogenetic probabilities, allele frequency, and protein annotations; they were aggregated using a fully connected layer architecture. Compared to 65 other predictors on clinical data, PATHOS outperforms state-of-the-art performance. It achieves a Matthews Correlation Coefficient (MCC) of 0.591 on a manually and carefully curated clinical dataset and 0.826 on a ClinVar dataset, surpassing other leading tools. Furthermore, case studies on the progesterone receptor and the KCNQ1 ion channel illustrate that PATHOS can identify functionally critical regions and known pathogenic mutations missed by other leading predictors like AlphaMissense. To ensure broad accessibility and facilitate use by non-specialists, a user-friendly web server containing a database of 140 million precomputed predictions from human protein from Swiss-Prot was provided. The web server is available at: https://dsimb.inserm.fr/PATHOS/