![\"Bonnie](\"https:\/\/www.newsbeep.com\/us\/wp-content\/uploads\/2025\/09\/bonnie-berger-headshot-m.jpg\" "\\"Bonnie")
<\/p>\nFor large language models (LLMs) like ChatGPT, accuracy often means complexity. To be able to make good predictions, ChatGPT must deeply understand the concepts and features that are associated with each word\u2014but how it gets to this point is typically a black box. <\/p>\n
Similarly, protein language models (PLMs), which are LLMs used by protein scientists, are dense with information. Scientists often have a hard time understanding how these models solve problems, and as a result, they struggle to judge the reliability of the models\u2019 predictions. <\/p>\n
<\/p>\n
Bonnie Berger is a mathematician and computer scientist at the Massachusetts Institute of Technology. She\u2019s interested in using large language models to study proteins.<\/p>\n
Bonnie Berger<\/p>\n
\u201cThese models give you an answer, but we have no idea why they give you that answer,\u201d said Bonnie Berger<\/a>, a mathematician and computer scientist at the Massachusetts Institute of Technology. Because it\u2019s difficult to assess the models\u2019 performance, \u201cpeople either put zero trust or all their trust in these protein language models,\u201d Berger said. She believes that one way to calm these qualms is to try to understand how PLMs think.<\/p>\n Continue reading below…<\/p>\n Recently, Berger\u2019s team applied a tool called sparse autoencoders, which are often used to make LLMs more interpretable, to PLMs.1 By making the dense information within PLMs sparser, the researchers could uncover information about a protein\u2019s family and its functions from a single sequence of amino acids. This work, published in the Proceedings of the National Academy of Sciences, may help scientists better understand how PLMs come to certain conclusions and increase researchers\u2019 trust in them.<\/p>\n James Fraser is a biophysicist at the University of California, San Francisco who uses computational approaches to study protein conformation. He was not involved in the study. <\/p>\n James Fraser<\/p>\n \u201c[This study] tells us a lot about what the models are picking up on,\u201d said James Fraser<\/a>, a biophysicist at the University of California, San Francisco who was not involved in the study. \u201cIt\u2019s certainly cool to get this kind of look under the hood of what was previously kind of a black box.\u201d <\/p>\n Berger thought that part of people\u2019s excitement about PLMs had come from AlphaFold<\/a>\u2019s success. But while both PLMs and AlphaFold are AI tools, they work quite differently. AlphaFold predicts protein structure by aligning a lot of protein sequences. Models like these typically boast a high level of accuracy, but researchers must spend considerable time and resources to train them.<\/p>\n On the other hand, PLMs are designed to predict features of a protein, like how it interacts with other proteins, from a single sequence. PLMs learn the relationship between protein sequence and function instead of the relationship between different protein sequences. While they learn much faster, they may not be as accurate. <\/p>\n \u201cWhen large language models that only take a single sequence came along, people thought, \u2018We should believe this too,\u2019\u201d Berger said. \u201cBut now, they\u2019re at the stage of, \u2018Oh my gosh, they\u2019re not always right.\u2019\u201d To know when PLMs are right or wrong, researchers first need to understand them. <\/p>\n PLMs are highly complex. Each neuron in the neural network\u2014AI\u2019s equivalent of a brain\u2014is assigned to more than one discrete unit of information, called tokens. Conversely, multiple neurons often process each token.<\/p>\n Onkar Gujral is a fifth-year mathematics PhD student at the Massachusetts Institute of Technology, advised by Bonnie Berger. He was the lead author of the study. <\/p>\n Onkar Gujral<\/p>\n![\"James](\"https:\/\/www.newsbeep.com\/us\/wp-content\/uploads\/2025\/09\/james-fraser-headshot-m.jpg\" "\\"James")
<\/p>\n![\"Onkar](\"https:\/\/www.newsbeep.com\/us\/wp-content\/uploads\/2025\/09\/onkar-gujral-headshot-m.jpg\" "\\"Onkar")
<\/p>\n