Pitt immunologists have developed a new protein language model called SWING — sliding window interaction grammar — that accurately predicts how proteins interact.
“We see SWING as a widely useable platform for all biologists to study protein–protein interactions in their favorite contexts,” said co-senior author Jishnu Das, Ph.D., assistant professor of immunology at Pitt, who said the research reflects the interdisciplinary collaboration between his lab’s expertise in machine learning and co-senior author Alok Joglekar’s focus on antigen discovery.
Protein language models are algorithms that aim to predict protein structure and function by analyzing their amino acid sequences, like understanding a sentence based on the order of words.
Unlike traditional protein language models that treat proteins as isolated units, SWING captures the “grammar” of how proteins interact by using a sliding window approach to pair amino acids across proteins and generate a language-like representation of the paired sequences.
“Virtually no protein acts in isolation,” said Joglekar, Ph.D., assistant professor of immunology. “Our model doesn’t focus on the proteins as words in a dictionary, but rather the grammar which governs how those words form sentences.”
In a new Nature Methods paper, the researchers — including co-first authors and graduate students in Das’s lab Jane Siwek, Alisa Omelchenko and Prabal Chhibbar — showed that SWING accurately predicted binding between peptides and major histocompatibility complex, a key step in adaptive immunity that allows presentation of peptides from foreign substances to T cells.
According to the researchers, this application could be useful for predicting vaccine design, understanding mechanisms of immune attack and tolerance and understanding genetic risk of disease.
“SWING can learn relationships between proteins in the absence of structural information or extensive training datasets,” said Das, highlighting its potential to scale across biological systems and species.
Das and Jogkelar hope to expand SWING’s capabilities to model antibody–antigen and small molecule–protein interactions, making it a versatile tool for biologists across disciplines.
SWING is also described in a Research Briefing