Why the 2024 Nobel Prize in (AI for) Chemistry Matters So Much
In summary: because the research and development behind this Nobel Prize unlocked not only new biology but also new computer science, all directly leading to advances for mankind along various fronts.
The 2024 Nobel Prize in Chemistry, awarded for groundbreaking advances in protein structure prediction and design using AI, reflects the transformative moment we (scientists at the interface between chemistry and biology) are going through. And it is even more notable after another Nobel Prize being awarded, just a day before, to those who developed the very basis of AI.
At its core, the ability to reliably predict protein structures with AI is reshaping both basic and applied biological research. Besides, the work by DeepMind and by Prof. Baker's Institute for Protein Design established pillars for many others, a huge community indeed, to build upon. You got a glimpse into it in my summary above, but let's dive into this as I connect you to other blog posts where I develop the subtopics deeper.
Unlocking Fundamental Biological Research
Determining protein structures experimentally is a painstakingly slow process. Slow, time-consuming, expensive, and full of frustrations as many proteins can't even be produced in forms suitable to investigation. Experimental methods, such as X-ray crystallography or cryo-electron microscopy, required significant resources, time, and sometimes failed altogether.
That's why we scientists have dreamed for decades about the existence of computer programs capable of replacing all the costly and difficult experimental work by predictions in the form of 3D models of the molecules as well of their dynamics and interactions. And actually we keep dreaming this! After decades of little or no progress as tracked specifically by a dedicated consortium called CASP, Deemind broke the game with AlphaFold 2:
The Critical Assessment of Structure Prediction (CASP): Over a quarter century tracking the state…
These AI-based predictions of molecular structures have changed the game in what's called Structural Biology (biology studied at the fundamental atomic level), by allowing scientists to determine (or well, rather, to predict but at high confidence!) these structures quickly and without even requiring extremely powerful computers.
Moreover, the breakthrough has unlocked entire fields of biology that were previously held back by the experimental bottleneck.
Revolutionary Applications
On a practical level, the ability to predict protein structures reliably opened up new avenues in drug discovery and enzyme engineering, revolutionizing the pharma and biotech industries, as well as healthcare and biotechnological manufacture. Researchers can now design new medicines, industrial enzymes, and other biotech applications much faster, leading to innovations that can address pressing challenges in health, energy, and the environment.
More so with the latest AlphaFold, version 3, and with software for structure prediction from Professor (and now Nobel Prize winner) Baker's lab, both of which can handle and model the 3D structures not only of proteins but also of their complexes with many other kinds of biologically relevant molecules:
"Sparks of Chemical Intuition"—and Gross Limitations!—in AlphaFold 3
AlphaFold and Other AI Tools for Molecular Structure Go Beyond Proteins
Beyond Structure Prediction: Protein Design!
One major aspect of this Nobel Prize, that doesn't touch directly the DeepMind people although they do also work on this, is the leap AI has made in the design of entirely new proteins from scratch. A field led by far by Prof. Baker, who achieved impressive results even without AI, AI-powered methods have taken it to new heights. The ability to design novel proteins has far-reaching implications, from creating more effective treatments to building proteins with functions that don't exist in nature at all.
In this way we can design medicaments, enzymes, and other molecules that are far from any potential solution existing in nature.
You can learn more about protein design in general here:
The Era of Machine Learning for Protein Design, Summarized in Four Key Methods
Unlocking New AI
One of the most significant impacts of DeepMind's work with AlphaFold 2, as outlined in their groundbreaking Nature paper, was how it unlocked new mathematical frameworks, algorithms, and concepts that have since been adopted and adapted by the broader AI and scientific community, over and over.
Among several elements, probably these were key:
- The core of AlphaFold 2's success is its Evoformer module, a then-novel neural network architecture that significantly advances how evolutionary data (multiple sequence alignments, MSAs) is used. Traditionally, MSAs helped scientists identify evolutionary relationships but "offline", that is not as part of the same neural network predicting the protein structures. This was helpful, but had drawbacks that were solved by DeepMind's innovation. In fact, AlphaFold 2's Evoformer iteratively refines predictions by exchanging information between sequence and structural representations throughout the network. This feedback loop enables AlphaFold 2 to reach highly accurate structural models quickly and efficiently, even for complex proteins, exploiting the information contained in the MSAs in a radically different way that, above all, is much more robust.
- The other critical innovation was the use of attention mechanisms within the neural network, allowing AlphaFold 2 to model proteins as spatial graphs where amino acids are represented as nodes. This graph-based approach enables the system to reason about the physical and evolutionary interactions within the protein structure in ways that were previously impossible. And in particular, the attention mechanism allowed it to grasp structural relationships between sets of amino acids (protein building blocks) that are far in a protein's sequence but close in space.
The innovative architecture behind AlphaFold 2, particularly its use of attention mechanisms and neural networks, set a new standard for computational biology. These advancements not only revolutionized protein structure prediction but also provided a blueprint for tackling complex problems in other fields, such as genomics, drug discovery, and even climate modeling.
And of major importance, by making their code and approaches openly available, DeepMind enabled a wave of innovation, inspiring other computer scientists to apply similar techniques to develop state-of-the-art AI models in a variety of disciplines, further extending the reach of their now Nobel-winning work. Unfortunately, the same didn't happen in their subsequent model, AlphaFold 3; but fortunately, others are getting there:
First Winners Emerge in the "Race" to Open-Source AlphaFold 3
www.lucianoabriata.com I write about everything that lies in my broad sphere of interests: nature, Science, technology, programming, etc. Subscribe to get my new stories by email. To consult about small jobs check my services page here. You can contact me here. You can tip me here.
