Stowers scientists unveil new AI model interpretation method

Researchers at the Stowers Institute for Medical Research have developed a novel method called PISA (pairwise influence by sequence attribution) that enhances the interpretability of AI models used in genomic studies. This advancement allows scientists to better understand how these models make predictions based on DNA sequences, separating experimental biases from biological signals.

PISA provides a detailed mapping of a model’s predictions at the single DNA base level, revealing the specific influences of each base on the model's output. This high-resolution approach enables researchers to identify and mathematically eliminate biases in experimental data, particularly in nucleosome mapping, which has traditionally required expensive and complex sequencing methods.

In their application of PISA, the Stowers team detected a significant technical bias in nucleosome mapping data. By removing this bias, they uncovered DNA sequences that play critical roles in positioning nucleosomes and delineating larger 3D chromatin domains. This new insight could help a deeper understanding of gene regulation and genetic diseases, areas of ongoing interest for biologists.

The study, published in Nature Communications, was led by Julia Zeitlinger, Ph.D., in collaboration with Anshul Kundaje, Ph.D., from Stanford University, and first author Charles McAnany, Ph.D. Zeitlinger noted, “We want to understand what the models learn and connect it to biological mechanisms.” The method builds on the existing BPNet framework, enhancing its capabilities for biological applications.

Traditionally, deep learning models have been challenging to interrogate due to their complex nature. PISA addresses this issue by providing a fine-grained view of model predictions, which can help in generating new hypotheses for biological experiments. The ability to visualize the specific contributions of each DNA base allows researchers to connect predictions back to biological mechanisms more effectively.

By applying PISA to various genomic data types, including MNase-seq, the researchers demonstrated its potential to disentangle biological insights from experimental noise. This capability represents a significant advancement in genomic research, offering a systematic approach for biologists to use AI models in their studies.

This announcement was distributed via PR Newswire and reflects ongoing efforts to integrate AI advancements into biological research.

Have news to announce?

Get your launch, funding round, or partnership covered on GAX Online, in front of tech buyers and the AI answer engines that cite us.