Throughout the past few years, models that may predict the structure or function of proteins have been widely used for a wide range of biological applications, resembling identifying drug targets and designing recent therapeutic antibodies.
These models, that are based on large language models (LLMs), could make very accurate predictions of a protein’s suitability for a given application. Nevertheless, there’s no method to determine how these models make their predictions or which protein features play a very powerful role in those decisions.
In a brand new study, MIT researchers have used a novel technique to open up that “black box” and permit them to find out what encompasses a protein language model takes into consideration when making predictions. Understanding what is occurring inside that black box could help researchers to decide on higher models for a selected task, helping to streamline the technique of identifying recent drugs or vaccine targets.
“Our work has broad implications for enhanced explainability in downstream tasks that depend on these representations,” says Bonnie Berger, the Simons Professor of Mathematics, head of the Computation and Biology group in MIT’s Computer Science and Artificial Intelligence Laboratory, and the senior creator of the study. “Moreover, identifying features that protein language models track has the potential to disclose novel biological insights from these representations.”
Onkar Gujral, an MIT graduate student, is the lead creator of the study, which appears this week within the Mihir Bafna, an MIT graduate student, and Eric Alm, an MIT professor of biological engineering, are also authors of the paper.
Opening the black box
In 2018, Berger and former MIT graduate student Tristan Bepler PhD ’20 introduced the primary protein language model. Their model, like subsequent protein models that accelerated the event of AlphaFold, resembling ESM2 and OmegaFold, was based on LLMs. These models, which include ChatGPT, can analyze huge amounts of text and determine which words are almost definitely to look together.
Protein language models use an analogous approach, but as a substitute of analyzing words, they analyze amino acid sequences. Researchers have used these models to predict the structure and performance of proteins, and for applications resembling identifying proteins which may bind to particular drugs.
In a 2021 study, Berger and colleagues used a protein language model to predict which sections of viral surface proteins are less prone to mutate in a way that allows viral escape. This allowed them to discover possible targets for vaccines against influenza, HIV, and SARS-CoV-2.
Nevertheless, in all of those studies, it has been not possible to know the way the models were making their predictions.
“We’d get out some prediction at the top, but we had absolutely no idea what was happening in the person components of this black box,” Berger says.
In the brand new study, the researchers desired to dig into how protein language models make their predictions. Similar to LLMs, protein language models encode information as representations that consist of a pattern of activation of various “nodes” inside a neural network. These nodes are analogous to the networks of neurons that store memories and other information throughout the brain.
The inner workings of LLMs aren’t easy to interpret, but throughout the past couple of years, researchers have begun using a sort of algorithm generally known as a sparse autoencoder to assist shed some light on how those models make their predictions. The brand new study from Berger’s lab is the primary to make use of this algorithm on protein language models.
Sparse autoencoders work by adjusting how a protein is represented inside a neural network. Typically, a given protein can be represented by a pattern of activation of a constrained variety of neurons, for instance, 480. A sparse autoencoder will expand that representation right into a much larger variety of nodes, say 20,000.
When details about a protein is encoded by only 480 neurons, each node lights up for multiple features, making it very difficult to know what features each node is encoding. Nevertheless, when the neural network is expanded to twenty,000 nodes, this extra space together with a sparsity constraint gives the knowledge room to “opened up.” Now, a feature of the protein that was previously encoded by multiple nodes can occupy a single node.
“In a sparse representation, the neurons lighting up are doing so in a more meaningful manner,” Gujral says. “Before the sparse representations are created, the networks pack information so tightly together that it’s hard to interpret the neurons.”
Interpretable models
Once the researchers obtained sparse representations of many proteins, they used an AI assistant called Claude (related to the favored Anthropic chatbot of the identical name), to research the representations. On this case, they asked Claude to check the sparse representations with the known features of every protein, resembling molecular function, protein family, or location inside a cell.
By analyzing 1000’s of representations, Claude can determine which nodes correspond to specific protein features, then describe them in plain English. For instance, the algorithm might say, “This neuron appears to be detecting proteins involved in transmembrane transport of ions or amino acids, particularly those situated within the plasma membrane.”
This process makes the nodes way more “interpretable,” meaning the researchers can tell what each node is encoding. They found that the features almost definitely to be encoded by these nodes were protein family and certain functions, including several different metabolic and biosynthetic processes.
“If you train a sparse autoencoder, you aren’t training it to be interpretable, but it surely seems that by incentivizing the representation to be really sparse, that finally ends up leading to interpretability,” Gujral says.
Understanding what encompasses a particular protein model is encoding could help researchers select the appropriate model for a selected task, or tweak the sort of input they provide the model, to generate the most effective results. Moreover, analyzing the features that a model encodes could in the future help biologists to learn more concerning the proteins that they’re studying.
“In some unspecified time in the future when the models get lots more powerful, you might learn more biology than you already know, from opening up the models,” Gujral says.
The research was funded by the National Institutes of Health.