Google and DeepMind spinoff Isomorphic Labs has developed an AI model called AlphaFold 3 that can, it's claimed, predict the structure of molecules more accurately than existing tools.

According to DeepMind's co-founder, AlphaFold 3 "can predict the structures and interactions of nearly all of life's molecules with state-of-the-art accuracy including proteins, DNA and RNA."

It's hoped the system can be used to design drugs and treatments faster and more accurately than current methods, and research suggests that may be the case. AlphaFold 3 attempts to predict the 3D structures of biomolecular complexes from descriptions of those complexes; this is potentially useful for doing things like developing drugs that are a better fit for a specific application.

"This breakthrough opens up exciting possibilities for drug discovery, allowing us to rationally develop therapeutics against targets that were previously difficult or deemed intractable to modulate," declared Isomorphic Labs, in a blog post.

According to an article in the science journal Nature, AlphaFold 3 is "capable of high accuracy prediction of complexes containing nearly all molecular types present in the Protein Data Bank" - a tool for visualizing molecular structures.

The developers claim their AI model offers a 50 percent improvement over existing prediction methods on the PoseBusters benchmark, and that in certain categories of molecular interactions, it has doubled prediction accuracy.

The AI model improves upon its predecessor, AlphaFold 2, which debuted in 2020 and has been cited in biomedical research work more than 20,000 times. Where AlphaFold 2 was designed to predict protein structure, AlphaFold 3 can be used to explore a broader set of molecular systems - multiple proteins, DNA, RNA, and small molecule ligands, for example.

It's not free from hallucination, though. The paper reveals that about 4.4 percent of the time, AlphaFold 3 will incorrectly predict "chirality" - when a structure is distinct from its mirror image and cannot be superimposed. The model's predictions also sometimes produce overlapping atoms, among other occasional shortcomings.

With the debut of AlphaFold 3, Google has released AlphaFold Server - a free platform for biomolecular exploration - as long as the work is non-commercial and abides by the terms of use. Training machine learning models with Server data is not allowed, for example.

AlphaFold Server lets you input data through its web interface for assorted biological molecule types, using a FASTA file. The AI model then processes the job and returns an overview of the structure as a 3D rendering.

Several researchers have complained that Google and Isomorphic Labs have not released relevant code or model weights, which would allow academics to better assess the work.

"Computational papers without code should not be accepted," argued Stephanie Wankowicz, a structural and computational biologist at UCSF's Fraser Lab, in a social media post. "Plus restricted access impedes the advancement of open science."

Google did not immediately respond to a query regarding whether it intends to provide code and model weights at some later time. ®

https://www.theregister.com//2024/05/09/google_deepmind_alphafold3_model/