The revolution brought by artificial intelligence in biology and medicine is gaining a new, crucial chapter. Over the past few years, we have witnessed a boom in powerful tools, so-called protein language models, which have fundamentally changed the way scientists approach drug research, vaccine development, and the understanding of the very foundations of life. These sophisticated systems, based on the architecture of large language models (LLMs) like those that power popular chatbots, have shown a stunning ability to predict the structure and function of proteins with incredible accuracy. Despite their success, one fundamental problem remained unsolved and posed a significant obstacle – their complete opacity. Scientists were getting extremely accurate answers but had no insight into how the model reached those conclusions. They were working with a kind of "black box," which limited trust and the possibility of further refinement.

A recent study, originating from a laboratory at the prestigious Massachusetts Institute of Technology (MIT), marks a turning point in solving this problem. The research team has successfully applied an innovative technique that, for the first time, allows scientists to peek inside this "black box" and precisely determine which protein features the artificial intelligence considers when making its predictions. This breakthrough not only increases the transparency and explainability of AI models but also opens the door for accelerated development of new therapies and a better understanding of complex biological processes.

Decoding the "black box": How AI makes decisions

Understanding the decision-making process within these models is crucial for their further application. The MIT team, led by Onkar Gujral as the lead author and mentored by Bonnie Berger, a distinguished professor of mathematics and head of the Computation and Biology group, has developed a method that demystifies the inner workings of protein language models. Their work, published in the prestigious scientific journal Proceedings of the National Academy of Sciences, has the potential to transform how these powerful tools are used in biomedical research.

Protein language models, whose foundations were laid back in 2018 by Professor Berger and her then-student Tristan Bepler, function by analyzing vast databases of amino acid sequences, similar to how language models analyze text. By learning the patterns and relationships between amino acids, they can predict the three-dimensional structure of a protein and its biological function. It was precisely such models that were key to the rapid development of revolutionary tools like AlphaFold, ESM2, and OmegaFold. However, the problem was that the information within the model was encoded in a very dense and incomprehensible way. Scientists could see the final result, but not the path that led to it. It was like having a genius student who always solves the most complex math problem correctly but can never show you their work.

An innovative technique that brings light into the darkness

To solve this problem, the MIT researchers turned to an algorithm known as a "sparse autoencoder." This is the first time such an approach has been successfully applied to protein language models. The principle of operation is elegant and powerful. In standard models, information about a specific protein is encoded through the activation of a relatively small number of "nodes" within the neural network, for example, 480. In such a dense representation, each individual node must encode multiple different protein features simultaneously, making interpretation practically impossible.

The sparse autoencoder works in the opposite way: it drastically expands the representation space. Instead of 480 nodes, the model now uses, for example, 20,000 nodes. At the same time, the algorithm introduces a "sparsity constraint" which ensures that only a small number of these nodes are activated to describe the protein. This allows the information, which was previously compressed, to be "spread out." The consequence is that a single specific feature of a protein, which was previously encoded across multiple different nodes, can now occupy its own, unique node. "In a sparse representation, the neurons that fire do so in a more meaningful way," explains Gujral. Before this method, the networks packed information so tightly that it was impossible to decipher the role of individual neurons.

The role of artificial intelligence in interpreting itself

After obtaining these "purified" and sparse representations for thousands of different proteins, the scientists faced a new challenge: how to understand what each of these activated nodes means. For this purpose, they used the help of another artificial intelligence, an assistant known as Claude. Claude's task was to compare the sparse representations with the already known characteristics of each protein, such as its molecular function, the family it belongs to, or its location within the cell.

By analyzing a vast number of examples, Claude was able to link the activation of specific nodes with concrete biological properties and then describe them in simple, human-understandable language. For example, the algorithm could generate a description like: "This neuron appears to detect proteins involved in the transmembrane transport of ions or amino acids, particularly those located in the plasma membrane." Through this process, the nodes became "interpretable," and for the first time, scientists gained a clear insight into what the model "thinks." It turned out that the features most commonly encoded by the models are the protein family and specific functions, including various metabolic and biosynthetic processes.

Practical implications: From faster drug discovery to new biological insights

This advancement has far-reaching consequences. Understanding the features that a particular protein model encodes allows researchers to choose the most appropriate model for a specific task. Whether it's identifying new target molecules for drugs or designing more effective vaccines, it is now possible to use a tool that is best "tuned" to solve a specific problem. This directly accelerates and reduces the cost of the entire research and development process.

For example, in a 2021 study, Professor Berger's team used a protein language model to predict which parts of viral surface proteins were least likely to mutate. By doing so, they identified promising targets for the development of universal vaccines against influenza, HIV, and SARS-CoV-2. With the new method for interpretation, it is now possible not only to get such a prediction but also to understand on the basis of which biochemical and structural properties the model made that decision, which provides an additional level of confirmation and directs further laboratory research.

Furthermore, analyzing the features that the model independently recognizes as important could one day lead to completely new biological discoveries. It is possible that artificial intelligence, by analyzing patterns in data that the human eye cannot perceive, will identify previously unknown protein functions or discover new connections between different biological pathways. "One day, when the models become even more powerful, we might learn more about biology than we currently know, precisely by opening up the models themselves," Gujral concludes optimistically. This technology promises not only to help us find answers to known questions but also to pose entirely new ones that will shape the future of science.