Within the past few years, models that can predict protein structure or function have been widely used for a variety of biological applications, such as identifying drug targets and designing new therapeutic antibodies.
These models are based on large-scale language models (LLMs) and can very accurately predict the suitability of proteins for specific applications. However, there is no way to determine how these models make predictions, or which protein characteristics play the most important role in these decisions.
In a new study, MIT researchers used new techniques to open its “black box” and allow them to determine the features that protein language models take into account when making predictions. Understanding what is happening within that black box can help researchers choose better models for specific tasks and streamline the process of identifying new drugs and vaccine targets.
“Our work has broadly impacted the increased explanability of downstream tasks that rely on these representations,” says Bonnie Berger, head of mathematics, director of the Computation and Biology Group at MIT’s Institute of Computer Science and Artificial Intelligence. “In addition, we can identify features that the Protein Language Models track may reveal new biological insights from these representations.”
MIT graduate student Onkar Gujral is the lead author of the study that will be featured this week. Proceedings of the National Academy of Sciences. Mihir Bafna, a graduate student at MIT, and Eric Alm, a professor of bioengineering, are also authors of the paper.
Open the black box
In 2018, Berger and former MIT graduate student Tristan Bepler PhD ’20 introduced their first protein language model. Like subsequent protein models that accelerated the development of AlphaFolds such as ESM2 and OmegaFold, their models were based on LLMS. These models, including ChatGpt, analyze a huge amount of text and understand which words are most likely to appear together.
Protein language models use a similar approach, but instead of analyzing words, they analyze amino acid sequences. Researchers use these models to predict protein structure and function and identify proteins that may bind to a particular drug, among other applications.
In a 2021 study, Berger and colleagues used protein language models to predict which sections of the viral surface proteins are unlikely to mutate in a way that allows for virus escape. This allowed us to identify possible targets for vaccines against influenza, HIV, and SARS-COV-2.
However, in all these studies it was impossible to know how the model makes predictions.
“At the end, I’ll take a few predictions, but I had no idea what was going on with the individual components of this black box,” says Berger.
In the new study, researchers wanted to delve into how protein language models predict. Similar to LLMS, protein language models encode information as representations consisting of activation patterns of different “nodes” within neural networks. These nodes are similar to networks of neurons that store memories and other information in the brain.
The inner workings of LLMS are not easy to interpret, but within the past few years, researchers have been able to shed light on how those models predict using a type of algorithm known as sparse autoencoder. A new study in Berger’s Lab is the first to use this algorithm in a protein language model.
Sparse autoencoder works by regulating how proteins are represented within neural networks. Typically, certain proteins are represented by the activation patterns of, for example, 480, a constrained number of neurons. Sparse autoencoder extends its representation to much more nodes. For example, 20,000.
When protein information is encoded by only 480 neurons, each node lights up to match multiple functions, making it extremely difficult to know the function that each node is encoding. However, when the neural network is expanded to 20,000 nodes, this extra space and sparsity constraint makes the information room “spread”. This allows protein features previously encoded by multiple nodes to occupy a single node.
“In sparse expressions, neuronal illumination does so in a more meaningful way,” Gujar says. “Before a sparse representation is created, the networks are very close to information, making it difficult to interpret neurons.”
Interpretable models
When researchers got sparse representations of many proteins, they used an AI assistant called Claude (associated with the popular human chatbot of the same name) to analyze the representations. In this case, they asked Claude to compare the sparse expression with known features of each protein, including molecular function, protein family, and intracellular location.
By analyzing thousands of expressions, Claude can determine the nodes corresponding to the characteristics of a particular protein and explain them in plain English. For example, an algorithm might say, “This neuron appears to be detecting proteins involved in transmembrane transport of ions or amino acids, particularly proteins found in the plasma membrane.”
This process makes the node much more “interpretable.” This means that researchers can know what each node is encoding. They found that the features most likely to be encoded by these nodes are specific functions that involve several different metabolic and biosynthetic processes.
“When you train a sparse autoencoder, you can’t train it to be able to interpret it, but by encouraging the representation to be truly sparse, you can see that it leads to interpretability,” Gujar says.
Understanding what a particular protein model is encoding can help researchers choose the right model for a particular task, and adjust the type of input they give to the model, to produce the best results. Furthermore, analyzing the functions that models encode can help one day learn more about the proteins biologists are studying.
“At some point when a model becomes much more powerful, you can learn more about biology than you already know from opening it,” Gujar says.
This study was funded by the National Institutes of Health.
