New Publication in Genome Biology

Our latest study evaluates how modern DNA foundation models capture the functional impact of genetic variants.

We are thrilled to announce that our latest research, “Benchmarking DNA foundation models for zero-shot variant effect prediction shows the importance of context, training, and architecture” has been published in Genome Biology.

Large language models have revolutionized how we process text, and DNA foundation models aim to do something conceptually similar with genomic sequences. But can these models truly recognize when a single nucleotide change carries a functional consequence?

To answer this question, researchers at COMBINE Lab (Prof. Alberto Magi and Marta Baragli, and former members Ilaria Alfisi and Francesca Ciapi) systematically benchmarked prominent DNA foundation models, including Nucleotide Transformer, DNABERT, HyenaDNA, and Evo 2, across hundreds of thousands of genetic variants from ClinVar.

Key Findings:

• Architecture & Training Matter: Bigger models are not necessarily better. Models trained by reconstructing masked DNA sequences substantially outperformed those trained to predict the next nucleotide. Furthermore, training on multi-species genomes provided a distinct advantage.
• Understanding Genomic Context: The top-performing models demonstrate a genuine capacity to capture biological structure, encoding not only the local effect of a mutation but also its impact on the surrounding sequence.
• A Foundation for Next-Gen Predictors: Rather than serving solely as direct predictors, these information-rich representations provide a powerful foundation for training specialized models to assess variant pathogenicity and functional impact.

This study underscores COMBINE Lab's commitment to advancing representation learning and AI applications in precision oncology and computational genomics.

Read the full open-access paper in Genome Biology.