BFPSM: A biological language model-based framework for pathogenic synonymous mutation prediction

Accurate prediction of pathogenic synonymous mutations is essential for advancing precision medicine. To address this challenge, we propose BFPSM, a deep learning framework that systematically integrates heterogeneous biological language models for pathogenic synonymous mutations prediction based on sequences only. Specifically, BFPSM extracts embeddings from three DNA language models and one RNA language model to capture evolutionary conservation, long-range dependencies, reverse-complement information and splicing-related information, integrating them via feature concatenation to capture diverse biological contexts. Additionally, to fully exploit these multi-scale embeddings, high-dimensional outputs are processed by dynamic multi-scale convolutional neural network-bidirectional long short-term memory and lower‑dimensional ones by text convolutional neural network. This design enables multi-view feature extraction and enhances prediction robustness. Evaluated on the complete test set, BFPSM achieved an AUC of 0.928 and an AUPR of 0.941. On the largest common subset of the test set, the corresponding AUC and AUPR were 0.938 and 0.949, respectively. These results demonstrate that BFPSM outperforms state-of-the-art methods at the sequence level, highlighting its utility for clinical and functional variant prioritization. The source code has been made available on https://github.com/xialab-ahu/BFPSM.