A Transferable Genomic Language Model Framework for Fungal Gene Essentiality Prediction
- 1Dartmouth College
Abstract
Predicting biological function from genomic sequence remains a major challenge in computational and systems biology. Here, we tested whether the genomic language model Evo2, which encodes context-dependent DNA sequence patterns into embeddings, enables prediction of essential genes in fungi, a phenotype central to fungal biology and antifungal target discovery. We found that model performance was constrained not by the type or complexity of the downstream classifier, but by the biological information contained in Evo2 DNA embeddings. Specifically, the information recoverable from these embeddings progressively declined for biological features further downstream of DNA sequence, revealing a bottleneck for predicting higher-order cellular phenotypes. We alleviated this bottleneck by integrating Evo2 embeddings with two sequence-informed, system-level features: ortholog-based essentiality and protein-protein interactions. The multimodal framework demonstrated consistent performance both within and across three evolutionarily divergent yeasts (Candida albicans, Saccharomyces cerevisiae, and Schizosaccharomyces pombe), and its predictions were supported by published experimental evidence when transferred to the filamentous mold Aspergillus fumigatus. These results establish our framework as a transferrable tool for predicting essential genes across fungal genomes, including species with limited or no experimentally determined essentially data.
Related
- GPBench: A benchmark framework for genomic prediction
- PathoBERT: A Hybrid Attention-Based Genomic Language Model for Read-Level Bacterial Pathogenicity Prediction
- MicrobeRT: A Framework and Suite of Fine-tuned Language Models for Genomic Sequences
- Toward a Transferable Representation of Protein Dynamics for Functional Prediction and Mechanistic Discovery
- Few-shot regulatory sequence design with a genomic language model
BibTeX
@misc{liao2026transferable,
title = {{A Transferable Genomic Language Model Framework for Fungal Gene Essentiality Prediction}},
author = {C. Liao and H. Thomas},
howpublished = {bioRxiv},
year = {2026},
month = sep,
doi = {10.64898/2026.09.03.749190},
url = {https://www.biorxiv.org/content/10.64898/2026.09.03.749190v1},
note = {CC BY},
}
RIS
TY - UNPB
TI - A Transferable Genomic Language Model Framework for Fungal Gene Essentiality Prediction
AU - C. Liao
AU - H. Thomas
PY - 2026
DA - 2026/09/03
PB - bioRxiv
DO - 10.64898/2026.09.03.749190
UR - https://www.biorxiv.org/content/10.64898/2026.09.03.749190v1
AB - Predicting biological function from genomic sequence remains a major challenge in computational and systems biology. Here, we tested whether the genomic language model Evo2, which encodes context-dependent DNA sequence patterns into embeddings, enables prediction of essential genes in fungi, a phenotype central to fungal biology and antifungal target discovery. We found that model performance was constrained not by the type or complexity of the downstream classifier, but by the biological information contained in Evo2 DNA embeddings. Specifically, the information recoverable from these embeddings progressively declined for biological features further downstream of DNA sequence, revealing a bottleneck for predicting higher-order cellular phenotypes. We alleviated this bottleneck by integrating Evo2 embeddings with two sequence-informed, system-level features: ortholog-based essentiality and protein-protein interactions. The multimodal framework demonstrated consistent performance both within and across three evolutionarily divergent yeasts (Candida albicans, Saccharomyces cerevisiae, and Schizosaccharomyces pombe), and its predictions were supported by published experimental evidence when transferred to the filamentous mold Aspergillus fumigatus. These results establish our framework as a transferrable tool for predicting essential genes across fungal genomes, including species with limited or no experimentally determined essentially data.
ER -