NSF grant ~$1.69M: predicting the "start switch" of plant genomes with AI for crop improvement — BIO-AI (Michigan State)

What makes this award interesting is that it points AI not at language or images but at how a plant genome begins to be read. As the program name BIO-AI signals, it throws machine learning at an unsolved problem in biology.

1One genome, different uses in every cell

The underlying biology is elegant. Every cell in a plant carries the same blueprint yet uses it differently, and the key to that control is where a gene begins to be read — the transcription start site (TSS). The TSS acts as a switch determining not only when a gene turns on but which version of a protein it ultimately produces. Yet no complete map of these central switches exists across plant diversity, from algae to crops.

2The BioLearnTSS framework

The core of the project is BioLearnTSS, a machine-learning framework that predicts, directly from DNA sequence, TSS locations and core-promoter structure. Training on flowering plants and then testing generalizability on a conifer, a green alga, and a moss — lineages far apart — sits squarely at the confluence of modern genomics and AI: reading function from sequence.

It also examines how alternative TSS usage shapes the proteome by altering splicing and translation efficiency, focused on maize.

3Testing in maize and Arabidopsis

Experiments accompany the modeling. Using genome editing in maize and the model plant Arabidopsis, the team directly tests how specific regulatory DNA elements within the core promoter influence TSS selection, gene expression, and protein production.

The work is organized around three aims: ML models for TSS prediction across lineages; the effect of alternative TSS usage on splicing and translation; and how core-promoter elements shape TSS selection, mRNA accumulation, and splicing.