VariantMedium: sensitive and generalizable somatic point mutation calling with 3D DenseNets trained and evaluated on experimental data.
Summary
Accurate identification of somatic variants is crucial in oncology, yet existing methods often struggle with high sensitivity in certain genomic regions. VariantMedium, a novel tool, combines a tree-based classifier with a 3D DenseNet convolutional neural network. This model was trained and validated on a large dataset of experimentally confirmed variants and further refined using an active learning strategy. It demonstrates superior sensitivity and comparable or better F1 scores than benchmark tools, especially in genomic regions prone to high sequencing error rates.
Analysis
From a bioinformatic perspective, VariantMedium represents an incremental yet significant application of neural network architectures, combining a tree-based classifier with a 3D DenseNet. Its novelty lies in the integration of an active learning strategy, where predictions are experimentally confirmed via targeted deep sequencing to refine the model. The tool was benchmarked against established variant callers like Mutect2 and Strelka2, demonstrating superior sensitivity and similar or better F1 scores, particularly in challenging genomic regions. The model was trained and validated on 336,839 variants from 2,956 WES/WGS samples and evaluated on 118,887 variants from two independent deep sequencing studies. The code is provided open-source, ensuring reproducibility and facilitating its potential clinical deployment for precision medicine by improving somatic mutation detection.