AI and the Protein Problem – An Unsolved Mystery in Drug Development

Developed in 2020 and awarded the Nobel Prize in Chemistry in 2024, AlphaFold revolutionized science by making it possible to predict protein structures. Would this development usher in a new era of drug discovery?

• The short answer: No. At least, not yet.
• The longer, more comprehensive answer: Protein behavior remains one of the greatest obstacles in drug development.

AlphaFold as the Subject of the 2024 Nobel Prize in Chemistry 

What Is AlphaFold?

AlphaFold is an AI program that can predict proteins' three-dimensional structures with high accuracy from their amino acid sequences. This technology has enabled revolutionary advances in biology and drug development.

Scientists who played key roles in developing AlphaFold received the 2024 Nobel Prize in Chemistry. Demis Hassabis and John Jumper of Google DeepMind won the prize for developing AlphaFold, an AI tool that predicts three-dimensional protein structures. The same prize was also awarded to David Baker of the University of Washington, known for his work in computational protein design.

Proteins: The Lock-and-Key Problem

A drug target, such as a mutation, is like a lock. The drug designed to open or disable that lock is the key. However, the key may fail because the lock is molecularly flexible.

Proteins are essentially like molecular springs.

Your key bends the lock or changes its shape. If you do not account for these dynamics, your key fails.

Another complication is that proteins do not operate in isolation. Their interactions with other proteins, ribonucleic acid (RNA) , and DNA can affect how they bind to molecules and the shapes they adopt.

AlphaFold's Contributions and Limitations

AlphaFold and its subsequent versions (AlphaFold Multimer, AlphaFold 3) can predict proteins' interactions with other molecules. However, these models still have significant weaknesses.

According to Dr. James Zou of Stanford University, “Predicting these dynamics and proteins' multiple conformations remains very challenging for AI models” . Dr. Monteiro da Silva offers another perspective: “We need to incorporate more physics to make these models more useful.”

While pursuing his doctorate at Brown University, Monteiro da Silva spent three years running physics-based simulations. These simulations sought to understand why proteins carrying mutations resist drugs . The results showed how the different shapes a protein can adopt prevent drug binding. However, the process was very slow: modeling four mutations took three years.

Dynamic Protein Analysis with AI

AlphaFold's ability to predict protein structures made it possible to identify pockets where drugs could bind. However, this is only a starting point.

Identifying a pocket is an important step, but designing a drug for that pocket or determining what it binds remains very difficult.

To address this challenge, Monteiro da Silva and his team manipulated protein amino acid sequences to develop a model that predicts “protein ensembles” . These ensembles represent the most common shapes a protein can adopt under particular conditions.

Monteiro da Silva explains, “If we can predict these ensembles quickly, we can screen mutations and develop drugs that do not encounter resistance.” The team tested this approach on a leukemia-causing kinase called ABL1. They made an important advance by screening 100 mutations in just one month.

PIONEER: Mapping Protein Interactions

The protein problem is not limited to a single protein; one protein can interact with hundreds of others. These interactions are critical to understanding how diseases develop.

Dr. Feixiong Cheng and his team at Cleveland Clinic combined AlphaFold's protein structure predictions with next-generation sequencing technologies that identify mutations in the human genome to develop PIONEER , a model that predicts the three-dimensional structures and binding sites of protein-protein interactions.

Dr. Cheng explains, “We can tell not only that one protein binds to another, but also where it binds” . Using PIONEER, the researchers tested the effects of 3000 mutations on 7000 protein pairs. These findings will be used to develop new treatments for lung and endometrial cancer.

The Future of Drug Discovery: More Physics and Dynamics

Although AI is opening new doors in drug development, it has not yet produced a complete breakthrough because of proteins' complex dynamics. However, models such as AlphaFold give scientists more information about how proteins function. This provides clues to designing drugs more effectively.

As Dr. Monteiro da Silva puts it, “If we can screen 100 mutations in a day, we will have taken a huge step forward in this field.” These advances could accelerate drug development and promise an exciting future for science.

Recurrent mutations are being investigated using models such as AlphaFold Multimer and AlphaFold 3 to model protein interactions. However, this approach remains slow at scale. Compared with Cleveland Clinic's PIONEER model, AlphaFold Multimer is reportedly 5000 times slower. Nevertheless, AlphaFold's approach may model some binding sites that PIONEER does not reveal.

Modeling the interaction network will require trying many different approaches, each with its own limitations.

PIONEER is a very good step, but other methods that add further detail to this approach should also be used.

We Could Not Have Reached This Point Without ChatGPT

AI and protein modeling research is not limited to physics and biology. Large language models (LLMs) add a new dimension to this field. Dr. James Zou of Stanford University used GPT-4 to conduct a study of the protein language model evolutionary scale modeling (ESM-2) . This model can predict protein structures from protein sequences.

Zou and his team first trained GPT-4 on a dataset consisting of thousands of scientific papers and studies describing mutations' biophysical properties, functions, and relationships to disease. They then integrated this knowledge into ESM-2, improving its ability to predict “which mutations could have larger or smaller effects” . Zou suggests that the same approach could also be used for models such as AlphaFold.

“Large language models contain far more of the information in texts about mutations' functions and biophysical properties” , says Zou. “But you cannot feed this information directly into AlphaFold. Language models play an important complementary role here.”

AlphaFold's Mystery: The Dynamics Behind Its Predictions

How AlphaFold makes its predictions is still not fully understood. Monteiro da Silva says, “AlphaFold is somehow learning protein dynamics phenomenologically” . Scientists are working to understand this process and develop more accurate models. For now, however, AI-based methods still need support from physical simulations.

The Way Forward: Collaboration Between Physics and AI

Combining physical models with AI tools offers great potential for understanding protein interactions and accelerating drug development. Both large language models and protein structure prediction tools such as AlphaFold can speed progress in this field. However, these approaches still need to be used together, and scientists must continue improving the tools.

AI's growing role is opening a new era in science's understanding of protein dynamics. Bringing physics and language models together could make drug discovery and protein analysis much faster and more effective.