Stowers Institute Scientists Develop PISA for Genomic AI Models

Stowers Institute for Medical Research scientists have developed PISA, a method for interpreting genomic AI models, the institute said in a press release. PISA stands for pairwise influence by sequence attribution and traces a model prediction at a single DNA base back to other bases that influenced it.

The method creates a base pair resolution map of what a genomic AI model learned, rather than only showing what it predicted. Researchers can use it to separate experimental bias from biological signals and train models with a more specific focus.

When the team applied PISA to nucleosome mapping data, it found and mathematically removed a technical bias in the data. The analysis then identified DNA sequences that position nucleosomes and also mark boundaries of larger 3D chromatin domains, which are usually found through more sequencing intensive methods.

The researchers said PISA can point to DNA sequence elements and mechanisms for further study in gene regulation and genetic disease.