Hype Around AlphaFold 3: Results vs Limitations

(This post is part of the latest Substack newsletter, "Where Tech Meets Bio," which you can check to also read weekly tech+bio highlights and a deep dive "4 Biotech Startups Developing Breakthrough Drug Modalities.")

So, the biggest news of the week was that Deep Mind and Isomorphic Labs , two of Google’s core AI subsidiaries, have launched AlphaFold 3, a brand-new AI model that builds on the success of their previous model AlphaFold 2, and can predicts molecular structures with even higher accuracy, and isn’t only restricted to proteins.

In fact, AlphaFold 3 manages proteins and their interactions with ligands, ions, DNA, RNA, and more. Let’s summarize some of the actual improvements (as per disclosed information):

Achievements of AlphaFold 3

1. Broader Molecular Predictions: Unlike AlphaFold 2, which primarily focused on predicting protein structures, AlphaFold 3 expands its scope to predict interactions between proteins and a wide array of other molecules, including DNA, RNA, and small molecule ligands. This comprehensive capability allows for a more complete understanding of cellular processes at the molecular level.
2. Improved Accuracy: AlphaFold 3 significantly enhances prediction accuracy. For protein interactions with other molecule types, the accuracy has improved by at least 50% over existing methods. For certain critical types of molecular interactions, the accuracy has even doubled. This leap in precision marks a substantial improvement in the tool's utility for scientific research and pharmaceutical applications.
3. Innovative Architecture and Processing: The model incorporates an evolved version of the Evoformer, a deep learning architecture that was fundamental to the success of AlphaFold 2. Additionally, AlphaFold 3 employs a diffusion network to assemble its predictions, similar to those used in AI image generators. This method begins with a cloud of atoms and iteratively refines their arrangement to achieve a highly accurate final structure.
4. Utility in Drug Discovery: AlphaFold 3's enhanced capability to accurately predict how drugs and other therapeutic molecules interact with proteins opens up new avenues in drug design. This is particularly vital for understanding complex interactions in the human body and designing new drugs that can effectively target specific proteins or molecular pathways.

Comparative Advancements Over AlphaFold 2

• Scope of Molecular Interaction: AlphaFold 2 was a revolutionary tool for predicting the structure of proteins in isolation or in simple complexes. AlphaFold 3 extends this to a full spectrum of biomolecules, enabling a holistic view of cellular machinery and interactions. This broader scope is critical for understanding complex biological processes and diseases.
• Computational Efficiency and Accessibility: With the launch of the AlphaFold Server, AlphaFold 3 is made accessible to a wider scientific community, facilitating easy and free access to its capabilities. This server enables researchers, regardless of their computational resources or expertise in machine learning, to model complex molecular structures and interactions.
• Practical Applications and Collaborations: While AlphaFold 2 laid the groundwork for understanding protein structures, AlphaFold 3 is directly applied in real-world scenarios, particularly in drug design. Collaborations with pharmaceutical companies, such as those undertaken by Isomorphic Labs, are utilizing AlphaFold 3 to tackle real-world drug design challenges, potentially leading to new treatments for diseases.

Limitations, concerns

Now, I don’t need to explain how cool the new Alpha Fold 3 is, because pretty much every mainstream media outlet has already done it with fanfares.

Let’s talk about some nuances instead, outlined in broad strokes in this MIT Tech Review article:

1. Accuracy and Reliability: While AlphaFold 3 shows improvements in certain areas, its accuracy varies significantly depending on the type of interaction being modeled, with success rates ranging from 40% to over 80%. For some specific interactions, like protein-RNA, the model is noted to be quite inaccurate.
2. Risk of Hallucination: The use of diffusion techniques, while innovative, introduces the risk of the model hallucinating — generating plausible but non-existent structures. Even though more training data has been added to mitigate this risk, it hasn't been completely eliminated.
3. Restricted Access: Unlike AlphaFold 2, which had its code released as open-source, AlphaFold 3 will not have its full code publicly available. Instead, Google DeepMind will provide access through the AlphaFold Server, which limits experimentation to non-commercial purposes and restricts the types of molecules that can be studied. This restricts broader utilization and independent verification or modification by the research community.
4. Impact of Access Limitations: The decision to not release the full code and to impose usage restrictions could hinder the widespread adoption and innovative application of AlphaFold 3 in the scientific community. It could also limit the model's potential impact, particularly in environments where commercial use could drive rapid advancements and applications.

Structure is not all

Beyond possible nuances with AlphaFold 3 itself, it is important to remember, that solving structural aspects, no matter how precise, is just a small bit of a drug discovery puzzle.

As Derek Lowe put it nicely:

“Structure is not everything. It's very useful, very good to have, and it will accelerate a lot of really useful research. But it does not take you directly to a drug, nor to a better idea about a target for a drug, nor to a better chance of passing toxicity tests, nor to a better chance of surviving oral dosing and the bloodstream and the liver.”

By no means am I a skeptic, though; I am totally impressed by the news and congratulate everyone behind the new achievement with the launch of AlphaFold 3.

Anyway, you can try and experiment at AlphaFold Server; a good place to start is to watch the demo.

Also, here is a nice experimentation thread on Twitter where Jan Kosinski explores various angles of AlphaFold 3 and what it can (and can't) do correctly now.

(This post is part of the latest Substack newsletter, "Where Tech Meets Bio," which you can check to also read weekly tech+bio highlights and a deep dive "4 Biotech Startups Developing Breakthrough Drug Modalities.")