Protein-RNA-Ion: PDB 8AW3 structure. [Image courtesy of Google DeepMind/alphafoldserver.com/]
The adoption of a diffusion-based approach allows AlphaFold 3 to generate accurate 3D models of biomolecular complexes by predicting the raw coordinates of individual atoms, a significant departure from its predecessor’s protein-centric architecture.
Capturing the dynamism of biology
Demis Hassabis
In a press briefing, Demis Hassabis, CEO of DeepMind Technologies, explained the evolution of AlphaFold over the years and the significance of AlphaFold 3: “With AlphaFold 2, the big milestone moment in structural biology was in essence you can think of it as our solution to the static picture of proteins, going from predicting the amino acid sequence to the 3D structure of the protein,” he said. Structural biologists have embraced AlphaFold 2, using it, for instance, to resolving the structures of large protein complexes like the nuclear pore complex and the tuberculosis Mce1 protein. AlphaFold 2 has “unlocked all kinds of amazing research,” Hassabis said. “It has been cited over 20,000 times now and is being used by thousands of researchers around the world.
AlphaFold 3 represents the “next step” toward using AI to understand and model biology. “We know that biology and biological phenomena are dynamic,” Hassabis added. “You have to understand how properties of biology emerge through the interactions between different molecules in the cell. You can think about AlphaFold 3 as our first big step towards that.”
Beyond proteins
AlphaFold 3 is able to model proteins interacting not only with other proteins but also with other biomolecules, including DNA and RNA strands. Another important advantage of AlphaFold 3 is its ability to model protein interaction with ligands, which is incredibly important for drug discovery.
Alongside this debut, Google DeepMind is launching the AlphaFold Server, a user-friendly tool that allows non-commercial researchers to harness the capabilities of AlphaFold 3. With just a few clicks, biologists can generate large and complex protein structures that have previously been challenging to predict quickly and accurately.
AlphaFold can predict an array of biological interactions and structures
Unlike its predecessor, AlphaFold 2, which relied on an architecture optimized for predicting the structure of individual proteins, AlphaFold 3 (AF3) uses a diffusion model that predicts raw atom coordinates. This shift towards a diffusion-based architecture allows AF3 to model an array of biomolecular interactions, including proteins, nucleic acids, small molecules, ions and modified residues with high accuracy. A paper introducing the model, titled “Accurate structure prediction of biomolecular 2 interactions with AlphaFold 3,” notes that it can accurately predict “complexes containing nearly all molecular types present in the Protein Data Bank,” a widely used repository of 3D structural data for large biological molecules.
The PDB was also a core resource for training and evaluating AlphaFold 3 and its predecessors. A team of researchers using the first iteration of the model placed first in the 13th Critical Assessment of Structure Prediction (CASP) competition in December 2018. Two years later, a team using AlphaFold 2 did the same at CASP 14. A 2021 paper in Nature describes the “redesigned version” of the neural network-based model.AlphaFold 3’s predictions are highly accurate. A Nature paper notes that they often outperform those of specialized tools for protein-ligand and protein-nucleic acid interactions.
Diffusion-Based Deep Learning Architecture: Unlike its predecessors, AlphaFold 3 employs a diffusion-based architecture that models the generation of atomic coordinates directly. This generative approach allows for highly accurate modeling across different types of biomolecular interactions without the need for excessive specialization for different molecular components.
AlphaFold 3 tackles hallucinations with cross-distillation and enhanced training
One of the key challenges in computational structure prediction is the issue of hallucinations, where models generate plausible-looking structures that do not accurately represent reality. To address this problem, AlphaFold 3 has implemented cross-distillation with predicted structures from its predecessor, AlphaFold-Multimer v2. “Cross-distillation with predicted structures from AlphaFold-Multimer v2 […] was used to mitigate hallucination,” the authors note in their paper. This technique allows AlphaFold 3 to learn from the successes and failures of its predecessor, ultimately leading to more accurate predictions. Additionally, the model has been trained on data that represent disordered regions as extended loops, enabling it to better capture the flexibility and variability of these regions. By incorporating these training enhancements, AlphaFold 3 takes a significant step forward in generating reliable and biologically meaningful structure predictions.
AlphaFold 3 offers accurate protein-ligand interaction predictions
AlphaFold 3’s remarkable ability to predict the interactions between proteins and small molecules holds immense promise for revolutionizing drug discovery. The paper highlights that “AF3 shows far greater accuracy on protein-ligand interactions than state-of-the-art docking tools,” suggesting that the model can reliably predict the binding sites and optimal shapes for potential drug molecules. This capability could significantly streamline the drug design process, as the authors note that “AF3 is able to accurately predict the binding modes of small molecules to proteins, which has potential applications in drug discovery.” By providing accurate insights into protein-ligand interactions, AlphaFold 3 could drastically reduce the time and cost associated with experimental methods, allowing researchers to focus their efforts on the most promising drug candidates.
AlphaFold 3 supports open science
DeepMind has made AlphaFold 3’s methodology, code and protein structure predictions freely accessible to the scientific community. This open science initiative aims to accelerate scientific research through access to high-quality protein structure predictions. The authors emphasize the transformative impact this move has already had on the scientific community, enabling researchers from diverse backgrounds and institutions to benefit from AlphaFold 3’s breakthrough predictions.
DeepMind had already freely published the predicted structures of more than 200 million proteins known to science made using earlier versions of AlphaFold.
By fostering collaboration and knowledge sharing, DeepMind’s open science approach has the potential to accelerate scientific discoveries and pharmaceutical development worldwide. This initiative ensures that the groundbreaking predictions of AlphaFold 3 are not limited to a select few but rather contribute to the collective advancement of scientific progress.
As Max Jaderberg, Chief AI Scientist at Isomorphic Labs, stated, “We really believe that ultimately, understanding more about this will translate into much more effective drugs in the clinic and into the hands of patients.”
Filed Under: Drug Discovery and Development, Omics/sequencing
