Structure-aware transfer of genomic foundation representations across population and cross-clade classification
Year: 2026
Type: preprint
Abstract: Genomic foundation models provide transferable DNA sequence representations, but their application to sample-level classification requires effective integration of sequencing reads. We present EvoGR (Evo 2-based Genomic Representation), a read-level transfer framework using a shared frozen Evo 2 encoder and Permutation-invariant Attention Aggregation (PAA). Mean pooling converts variable-length single-stranded reads into fixed-dimensional embeddings, while PAA integrates variable-sized, unordered read sets. EvoGR addresses two complementary tasks: population classification using AIM-enriched reads and four-class genome-origin classification involving human, non-human animal, plant, and microorganism sequences. Inference requires only single-stranded reads, without reference genomes, SNP coordinates, or explicit REF/ALT representations. EvoGR achieved competitive performance across both tasks and maintained strong classification performance on real human sequencing data, demonstrating transfer from simulated to real sequencing inputs. Comparisons across genomic foundation models identified Evo 2 as the strongest encoder under the evaluated settings. Ablation experiments demonstrated the contribution of PAA, while read-length sensitivity analyses showed that longer reads generally improved classification performance, with diminishing gains at greater lengths. These findings establish EvoGR as a unified framework for read-level genomic representation transfer across distinct biological classification tasks.
Source: Research Square
Authors: Xiaoyuan Liu, Tianyi Zhao
Institution: Harbin Institute of Technology
Language: English
Subfield: Molecular Biology
Domain: Life Sciences
Open Access status: green
OpenAlex ID: W7233793647
