© The Indian Express Pvt Ltd
Tags:
Journalism of Courage
The recent emergence of large language models (LLMs) has revolutionised the research on proteins — the microscopic mechanisms that are involved in virtually every important activity happening inside all living things. Scientists use a version of LLM for various tasks, such as predicting the structure or function of a protein, which contributes to the development of drugs and vaccines.
However, little is known about how these models make such predictions, as they work like a “black box”, meaning there is no way to determine what is happening inside them. This poses several kinds of issues. For instance, researchers do not know if the basis for the model’s prediction is meaningful or not, and might waste months or years on dead-end experiments.
Now, a team of researchers based at the Massachusetts Institute of Technology (the United States) has tried to shed light on the inner workings of the language models that predict the structure and function of proteins by using an innovative technique. They have described their findings in the study, ‘Sparse autoencoders uncover biologically interpretable features in protein language model representations’, which was published in the journal Proceedings of the National Academy of Sciences last month. The team included Onkar Gujral, Mihir Bafna, Eric Alm, and Bonnie Berger.
Berger, the senior author of the study, told The Indian Express over email, “This is the first work that allows us to look inside the ‘black box’ of protein language models to gain insights into why they function as they do.”
Here is a look at the technique used by the researchers.
The LLMs used for protein research are known as protein language models (pLMs). Unlike the usual LLMs, which are trained on English words, pLMs are trained on protein sequences — the linear, specific order of one-dimensional amino acids that are linked together to form a three-dimensional protein. Typically, proteins are made of a combination of 20 different kinds of amino acids, and the structure and function of each protein are governed by the arrangement of the various amino acids in it. For instance, antibody proteins fold into shapes that help them to identify and target foreign bodies, like when a key fits into a lock.
Gujral, the lead author of the study, told The Indian Express over a video call, “Protein language models are trained on protein sequences and their objective is to predict the next amino acid in a sequence, just like regular language models which predict the next word in a sequence.”
By doing so, pLMs can intuit the patterns and interrelationships between protein sequence, structure and function. They can, for example, infer how changing amino acids in certain locations of a protein sequence could impact the shape that the protein folds into.
This helps scientists save a lot of time and effort, as earlier, they had to carry out extensive experiments using X-rays, microscopes and other tools to determine the shape and function of proteins. Knowing such details about a protein makes it easier for researchers to design drugs that bind to the protein, turning it on or off.
An LLM or pLM can make predictions because it is an artificial neural network. Just as the human brain is composed of neurons that connect and send signals to each other, these artificial neural networks are composed of several interdependent layers of neurons. It is these neurons that process large amounts of data to recognise patterns and make predictions.
However, it is not easy to pinpoint exactly which individual neuron is responsible for recognising specific patterns or making a prediction using the input data. That is because neural networks have far fewer neurons than there are concepts they understand, so nearly all of them get activated for any particular concept. Put simply, neurons work together in clusters and not in isolation. This is why it is challenging to interpret and analyse the inner workings of LLMs or pLMs.
Gujral said, “For example, one can give a body of text to an LLM that may contain various concepts like famous monuments such as the Taj Mahal. One cannot pinpoint which neuron will be identifying the Taj Mahal because every neuron will be doing hundreds of things, such as trying to recognise the names of monuments, animals, places, ideas, etc. That is the reason why one cannot interpret the inner workings of an LLM.”
Gujral and his team used sparse autoencoders, which are basically smaller neural networks that are trained on the inner activity of an LLM (in this case, a pLM). These autoencoders are able to separate distinct patterns in the activity since “sparse”, meaning very small, groups of their neurons get activated together. After several such patterns, called features, have been identified, scientists can ascertain which words or proteins trigger which features. This gives a kind of mind-map of the LLM or pLM.
Using sparse autoencoders, the MIT researchers could tell what information the pLM learnt when it had been trained on protein sequences. “We could understand what the pLM had learnt on its own and what information it was making use of. This can provide us with insight into what the pLM did or did not use in applications like drug or vaccine design. In this way, we can understand the model’s predictions and have more confidence in what it is telling us.”
