D. Pal, A. Odell, A. Singh, A. M. Thompson, A. Ross, A. Thessen
Developed an environment-aware DNA-LLM by fine-tuning AgroNT with stress-specific prompt tokens to capture stress-conditioned sequence representations. Prioritized genomic regions associated with grain yield in maize under heat and drought by analyzing stress-induced shifts in sequence representations and transformer attention patterns; findings supported by differential gene-expression and overlap with QTLs. Identified stress-associated motifs enriched in model-emphasized regions, with prioritized loci proximal to genes involved in stress adaptation pathways.
Abiotic stresses such as heat and drought severely reduce maize productivity, yet identifying genomic regions that confer stress resilience remains a challenge. Inspired by advances in Large Language Models (LLMs), Genomic Foundation Models (GFMs) have recently emerged as a promising approach for capturing regulatory patterns through large-scale pre-training on DNA sequences. However, their application to plant stress-response analysis remains unexplored. This study presents an environment-aware DNA-LLM that adapts AgroNT, a transformer-based GFM pre-trained on diverse plant genomes, by incorporating stress-specific prompt tokens. Through parameter-efficient fine-tuning, the model learns stress-conditioned sequence representations that form distinct clusters in the embedding space across environmental contexts. By combining stress-induced shifts in these sequence representations relative to control conditions with transformer attention patterns, we prioritized putative heat- and drought-responsive genomic regions associated with grain yield in the Genomes-to-Fields (G2F) panel. Prioritized regions were supported by spatiotemporal differential gene-expression evidence and overlap with stress-associated quantitative trait loci. They were further characterized through transcription-factor family analysis and regulatory motif enrichment. Attention-guided analysis additionally identified stress-associated motifs enriched within model-emphasized sequence regions. Overall, the prioritized loci were proximal to genes involved in transcriptional regulation, signaling, and metabolic pathways relevant to abiotic-stress adaptation, demonstrating the potential of stress-conditioned transformer-based sequence modeling for environment-aware genome-to-phenome analysis.