What's next after AlphaFold2 on protein structure prediction?

AlexNet won the ImageNet competition in 2012 and started the AI era. AlphaFold2 won the CASP14 in 2020 (AlphaFold in CASP13 was not as impressive as AlphaFold2). Was it the AlexNet moment in Biology? 

Many people believe so, as we see AlphaFold2 has made significant impacts in biology and medicine. A review published earlier this year nicely listed many impactful applications of AlphaFold2. In our newsletters, we mainly focus on the most recent developments. For example, researchers use AlphaFold and other AI tools to predict the evolution of viruses and design vaccines that could help us prepare for the next pandemic. DeepMind also used AlphaFold2 to predict the effect of missense mutations, which we have shared in our first newsletter. Below are more highlights of recent studies.

New tools are developed to help better utilize predicted protein structures. The large amount of protein structures brings a new challenge — how to quickly identify structurally similar sequences. In a recent paper on Nature Biotechnology, a workflow was developed to address the challenge. First, structural similarity can be measured (or trained) using the embeddings from a protein large language model (LLM). Then, one can search for proteins based on embedding similarity. Once structurally similar proteins are identified, we can structurally align the proteins. The alignment algorithm is similar to sequence alignment, with the only difference being the penalty parameters come from LLM embeddings.

Researchers try to improve the protein structure prediction algorithm. For example, AlphaFold2 uses multiple sequence alignment (MSA) to capture the evolutionary information, but MSA is quite computationally expensive. We may use a protein LLM to learn the same information, so that structure can be predicted from a single protein sequence. This was what Meta did in their ESMFold. Last week (October 10, 2023), BioMap and Baidu published their protein structure prediction algorithm on Nature Machine Intelligence. They combined LLM and AlphaFold2’s Evoformer algorithm, and achieved comparable accuracy with AlphaFold2.

(BTW, on October 10, 2023, BioMap announced a groundbreaking strategic collaboration with Sanofi to co-develop cutting-edge AI modules for biotherapeutic drug discovery leveraging BioMap’s AI platform. Sanofi pays BioMap $10M upfront with a total potential deal value of over $1 billion.)

AlphaFold2 isn’t the end of experimental structural biology. On the contrary, experimental structural biology becomes better with AlphaFold2 and AI. For instance, scientists from Fudan University developed an AI algorithm to assist cryo-EM to better capture the structural dynamics. (And there was an earlier example published in Science to get the structure of cytoplasmic ring of nuclear pore complex by integrative cryo-EM and AlphaFold.)

Many challenges still exist in protein structure prediction, including conformational dynamics, multimer, antibody, protein binding (especially antibody/TCR-antigen binding), orphan protein, mutation effect on stability and binding affinity, non-protein molecular structure and their interaction with protein, protein design (an emerging and promising area), etc. Stay tuned! We will share more when new impactful studies are published.

In 2022, 10 years later after AlexNet, ChatGPT started to revolutionize the whole society. Are we going to see the ChatGPT moment in Biology in 2030?

P.S. Demis Hassabis and John Jumper won the 2023 Lasker Award for the invention of AlphaFold.

##########

If you find the newsletter helpful, please consider:

• 🔊 Sharing the newsletter with other people

• 👍 Upvoting on Product Hunt

• 📧 Sending any feedback, suggestions, and questions by directly replying to this email or writing reviews on Product Hunt

• 🙏 Supporting us with a cup of coffee.

Thanks, and see you next time!