Protein Model Test Sheds Light on AI Black Box

Emory researchers develop a method to quantify the reliability of protein language models used in biology

Apr. 2, 2026 at 10:14am

Got story updates? Submit your updates here. ›

Computational biologists at Emory University have developed a new framework to test the accuracy of AI language models used to analyze complex biological data like proteins. The method compares how the language model encodes natural proteins versus synthetic random proteins, allowing researchers to identify areas of uncertainty or low-quality embeddings in the model's 'black box'. This provides a way to quantify the reliability of the model's predictions and improve the development of more accurate AI tools for studying genomics and metagenomics.

Why it matters

AI language models are revolutionizing biology by treating complex biological data like a language, but a critical gap has been the lack of a method to estimate the reliability of these predictions. This new framework developed at Emory provides a way to 'shine a light into the black box of AI' and better understand how language models work, which is essential for improving their reliability and developing more robust models to study the vast complexity of genomes and metagenomes.

The details

The key to the new testing method lies in understanding how evolution shapes proteins. The researchers compared how a protein language model would classify biologically meaningful proteins found in nature versus randomly generated synthetic proteins. By visualizing the model's latent space, they found that natural proteins were grouped together, segregated from the 'junkyard' of low-quality synthetic protein embeddings. They then quantified this relationship into a 'random neighbor score' that reflects the model's confidence in each protein embedding. Applying this method allows for more precise measurements of a language model's accuracy, which can then be used to enhance the machine learning process and develop higher quality AI tools for biological research.

  • The study was published in the journal Nature Methods on April 2, 2026.

The players

Yana Bromberg

Senior author of the paper and professor of biology and computer science at Emory University.

R. Prabakaran

First author of the study and a postdoctoral fellow in the Bromberg lab at Emory University.

Emory University

The institution where the computational biology research was conducted.

Got photos? Submit your photos here. ›

What they’re saying

“To the best of our knowledge, our framework is the first generalized method to quantify protein sequence embedding reliability.”

— Yana Bromberg, Senior author, professor of biology and computer science

“Our method is a simple, elegant solution to a complex problem. It's a foundational method with a lot of scope for a range of language models in science.”

— R. Prabakaran, First author, postdoctoral fellow

“We are shining a light into the black box of AI. Better understanding how a language model works allows you to find ways to keep improving its reliability and to develop better models.”

— Yana Bromberg, Senior author, professor of biology and computer science

What’s next

The researchers plan to apply their new method to further refine and improve the reliability of AI language models used in a range of biological research applications, from studying the human genome to analyzing complex microbial communities.

The takeaway

This new framework provides a critical tool for quantifying the reliability of AI language models in biology, allowing researchers to identify areas of uncertainty and improve the development of more accurate and robust computational tools to tackle the vast complexity of genomics and metagenomics.