MFDSMC: Accurate Identification of Cancer-Driver Synonymous Mutations Using Multiperspective Feature Representation Learning

Synonymous mutations do not change amino acid sequences,
but they
can drive cancer by influencing splicing, mRNA structure, translation
efficiency, and other molecular mechanisms. Although driver synonymous
mutations are significantly outnumbered by functionally neutral passenger
mutations in cancer, their accurate discrimination is critical to
understanding tumorigenesis. In this study, we developed multiperspective
feature-based predictor for driver synonymous mutation in cancer (MFDSMC),
a computational framework designed to improve the prediction of human
cancer-driver synonymous mutations. First, we curated synonymous mutations
from public cancer mutation databases to construct our data sets.
For each mutation, we systematically characterized features across
four biologically informed perspectives: sequence context, evolutionary
conservation, epigenetic modifications, and regulatory/functional
predictions. The optimal feature subset was identified through a feature
importance ranking and sequential forward selection. After multiple
machine learning classifiers were evaluated, XGBoost was selected
to build the prediction model. Results revealed that the multiperspective
fusion model outperformed models relying on single-perspective features
or lacking any individual feature category. Notably, newly introduced
epigenetic features derived from experimental sequencing data, combined
with regulatory/functional prediction features, collectively enhanced
the model’s performance. When tested on two independent test
sets and a curated data set of experimentally confirmed driver synonymous
mutations, MFDSMC exhibited superior performance compared to existing
state-of-the-art methods, providing a novel solution for precise prediction
of cancer-driver synonymous mutations in genomic research and clinical
applications. MFDSMC is available at https://github.com/xialab-ahu/MFDSMC.