The challenge: agricultural systems in a changing climate

We recently announced a $2.5 million grant with support from the Gates Foundation to develop a foundation model for crops aimed at addressing one of humanity's most urgent challenges: adapting agricultural systems to a rapidly changing climate. Crop genomes are enormous and profoundly understudied. They are not linear strings of nucleotides but vast, structured, and dynamic systems in which phenotype emerges from interactions spanning millions of base pairs, regulatory elements, chromatin architecture, and coordinated expression programs. Yet most computational approaches still analyze these systems in fragments, isolating SNPs, genes, or regions without modeling the broader biological context in which function truly arises. At Converge Bio, we believe that to understand genotype-phenotype relationships in a meaningful way, AI must operate at a biological scale.

Why long-context modeling matters in plant genomics

This conviction is driving our development of a multimodal, long-context foundation model purpose-built for plant genomics. In natural language, meaning depends on context across paragraphs and documents; in genomics, function depends on context across megabases. Complex agronomic traits such as heat resilience or yield are rarely governed by a single gene. They emerge from long-range regulatory interactions, co-expression networks, epistatic effects, and structural genomic organization acting in concert. Traditional machine learning pipelines typically treat these components separately, stitching together independent analyses into downstream heuristics. Long-context transformer architectures allow us to process them jointly, modeling dependencies across entire genomic regions rather than isolated loci. We are evaluating and extending state-of-the-art DNA modeling approaches capable of capturing long-range structure, enabling our models to reason simultaneously across entire QTL regions, regulatory sequences, gene neighborhoods, and integrated transcriptomic signals. Instead of asking whether a particular SNP is statistically significant, we can ask a more biologically grounded question: given a genomic region and its corresponding transcriptomic profile, which variant is most likely to be causally responsible for the observed phenotype?

Pretrain broadly, fine-tune precisely

The core technical principle underlying this effort is straightforward but powerful: pretrain broadly, fine-tune precisely. We are assembling large-scale crop genomic and transcriptomic datasets and harmonizing them into a unified training corpus. The objective of pretraining is not immediate prediction, but representation learning. Through self-supervised objectives such as masked and causal language modeling adapted to nucleotide sequences and expression profiles, the model learns the underlying statistical and structural regularities of plant genomes before being asked any task-specific question. Only after this large-scale pretraining phase do we fine-tune the system for causal SNP prioritization within QTL windows, multimodal genotype-phenotype mapping, and candidate gene ranking. This mirrors the paradigm that transformed natural language processing and protein modeling, now applied to the complexity of crop genomics through Converge Bio's integrated foundation-model stack.

Interpretability as a first-class design constraint

In biology, prediction without interpretability is insufficient. We therefore treat explainability as a first-class design constraint rather than an afterthought. By leveraging attention mechanisms and embedding-level analyses, we can map model predictions back to specific sequence segments, identify regulatory regions that drive decisions, and highlight expression features that influence prioritization. Transformer-based architectures provide a direct mapping between input regions and output inferences, enabling a transparent link between genomic context and predicted causality. This transforms the model from a black-box scoring engine into a scientific collaborator. Breeders can understand why a gene was prioritized. Biologists can generate testable mechanistic hypotheses. Experimental teams can design targeted CRISPR edits grounded in model-derived insights.

What success looks like

Our target outcome is ambitious and measurable. By the end of this effort, the fine-tuned model will accept a QTL region alongside a matched transcriptomic profile, rank candidate causal SNPs or genes, and place the true causal variant among the top candidates in the large majority of benchmark cases. This represents a translational shift from correlation-based mapping toward AI-driven causal prioritization operating at genomic scale. When successful, this approach has the potential to compress breeding cycles dramatically, accelerating the development of climate-resilient crops.

Reshaping causal gene discovery in agriculture

At Converge Bio, we see this as a defining opportunity to bring long-context, multimodal foundation models into the heart of agricultural genomics and fundamentally reshape how causal gene discovery is performed.