GGBound AI predicts microbial life limits from genomes with 4B-parameter agent

A team of researchers from multiple institutions has introduced GGBound, a genome-grounded AI agent that can predict a microbe's physiological life boundaries—like viable temperature, pH, salinity, and carbon source usage—directly from its DNA sequence. Traditional methods require weeks of lab culturing. GGBound instead uses a 4-billion-parameter LLM (built on Qwen) that takes genome embeddings from the LucaOne model, then reasons over a similarity-based retrieval-augmented generation (RAG) module and a Genome-scale Metabolic Model (GEM) perturbation tool. The authors curated a strain-centric benchmark from IJSEM, NCBI, and BacDive with 1,525 strains and 6,448 instances covering viability intervals, optima, substrate utilization, categorical traits, and morphology.

GGBound was trained through a three-stage pipeline: gene-text alignment, supervised fine-tuning on distilled agent trajectories, and GRPO with a novel counterfactual reward that reinforces the policy only when the authentic genome embedding improves correct token generation compared to a zero-gene ablation. Despite its modest 4B parameter count, GGBound matches or surpasses much larger frontier LLMs (e.g., 70B+ models) on the benchmark. Ablation studies confirmed that the genome-token fusion, dynamic tool use, and counterfactual reward each provide significant, distinct gains. This work effectively formulates life-boundary prediction as a unified genome-to-physiology task, offering a powerful new tool for biotechnology, ecology, and synthetic biology.