Understanding how amino acid sequences translate into complex, three-dimensional structures is a problem that has been puzzling scientists for decades1. Unlocking this mystery is key to uncovering information about how specific structures allow proteins to carry out specialised functions, opening new frontiers in synthetic biology and personalised medicine. Since the introduction of machine learning into the field, we’ve seen a huge variety of models to assist with this task. In this feature, we discuss the tools that we feel have had, and continue to have, the greatest impact in the field of computational structure prediction.
Want to learn more? Remember to check out our Proteomics Playbook when it publishes later this year!
Protein structure prediction
AI has dramatically changed many aspects of research within the life sciences, and proteomics is no exception. The field of protein structure prediction and design has undergone a transformative evolution, largely driven by advancements in AI models. Proteins play crucial roles in virtually all biological processes, with their functions being intimately linked to their three-dimensional structures. Therefore, making accurate protein structure predictions is essential for understanding biological mechanisms and designing novel therapeutics, enzymes and biomaterials.
Historically, determining protein structures relied heavily on experimental techniques such as X-ray crystallography, nuclear magnetic resonance (NMR) spectroscopy and cryo-electron microscopy (cryo-EM). While these methods have provided invaluable insights, they are often time-consuming, expensive and limited by the complexity of protein structures. The advent of computational methods marked a significant step forward, offering the potential to predict protein structures in silico.
Among the many computational approaches that have recently been developed, AI models have emerged as particularly powerful tools. The introduction of deep learning techniques has revolutionised the field, enabling unprecedented accuracy in structure prediction. A landmark achievement in this domain is the development of AlphaFold by DeepMind, a model that predicts protein structures with atomic-level accuracy that rivals experimental methods. This breakthrough has opened new avenues for scientific exploration and practical applications.
Below, we will explore the major advancements in AI models in recent years, with particular emphasis on free and open-source tools that are driving scientific advancement forwards within proteomics.
AlphaFold: The gold-standard for structural modelling
No feature on protein structure modelling would be complete without featuring AlphaFold, the cornerstone computational tool that has had a profound impact on proteomics research. Developed by Google DeepMind, AlphaFold was first released back in 2018, quickly demonstrating its prowess by placing first in the 13th annual Critical Assessment of Structure Prediction (CASP13) competition. Google’s new tool decimated the competition at CASP13, accurately predicting structures for almost 60% of proteins, compared to only 7% for the second-place model in the same category2.
The secret to the model’s success was its innovative deep learning approach, which used a convoluted neural network trained on PDB structures to calculate the distance between pairs of residues. Using this information, the model can generate “distograms” using multiple sequence alignment (MSA) to predict the structure of the protein given only its amino acid sequence3. This information, alongside the probability of backbone torsion distribution, is then optimised using gradient descent to create the final structure prediction.
Two years later, AlphaFold came back better than ever with the release of AlphaFold24. The second iteration was entered into the CASP14 competition, again demonstrating the impressive accuracy of the redesigned model. In many cases, the computational models were competitive with experimental structures determined with established techniques, such as x-ray crystallography. AlphaFold2 proved to be a major step forwards for AI-based structural modelling, shown to predict structures with atomic accuracy – even when no similar structures are known.
So, what causes this massive boost in performance? AlphaFold2 improvements can be attributed to a massive overhaul of the model’s architecture. One paper comparing the two models notes that AlphaFold2 uses several databases to construct an MSA, which is helpful in determining which parts of the sequence may be prone to mutation5. Additionally, the model identifies proteins with a similar structure to create a template structure, or “pair representation”. Although this was not a novel strategy in protein modelling at the time, AlphaFold2’s main advantage was its dual neural network modules – Evoformer and structure module – that work to iteratively refine the structure using MSA and template information.
AlphaFold 2 improvements + database expansion
Building on this work, Google Deepmind pushed several major updates to further improve the model’s accuracy. In particular, the release of AlphaFold Multimer was a significant milestone, allowing users to predict protein complexes containing multiple chains6. In the latest release, AlphaFold2.3 included 30% more training data, the incorporation of recent and large protein structures into the training dataset, and more chains and MSA sequences than before for AlphaFold Multimer7.
Also in 2022, Google Deepmind partnered with EMBL’s European Bioinformatics Institute (EMBL-EBI) to create a database for AlphaFold predictions for nearly all known catalogued proteins8. This ambitious project, the AlphaFold Protein Structure Database, is an impressive collection of over 200 million protein structure predictions that are freely accessible to the scientific community. Covering the proteomes of 48 species – including human – the database has already proved to be an essential resource for proteomics work, with over 1.4 million users across 190 countries so far9.
After the eventual release of the AlphaFold2 source code, the proteomics community were free to build upon the model to improve its prediction capabilities for certain applications. From increasing prediction speed or adapting the model to generate complex antibody structures, an impressive variety of tools and models have been developed building on the AlphaFold2 model.
Want to learn more? Remember to download our Proteomics Playbook, releasing late 2024.
RoseTTAFold
Following in the wake of DeepMind’s AlphaFold2 release, Baek et al. took inspiration from the innovative model to explore network architectures using similar features for improved structure prediction. They developed a three-track network, RoseTTAFold10, which takes patterns in protein sequence (one dimensional, 1D), amino acid interactions (2D) and three-dimensional structure (3D) into account. Using all three dimensions, this information flows back and forth through the network, allowing it to collectively reason about relationships within and between sequences, distances and coordinates. RoseTTAFold achieved similar results to AlphaFold2 in the CASP14 competition, demonstrating that the model is applicable to a range of modelling problems, including predicting challenging x-ray crystallography structures.
But that’s not all. 2024 saw a major advancement with the publication of RoseTTAFold All-Atom – a next-generation prediction and design tool for assemblies containing proteins, nucleic acids, small molecules, metals and chemical modifications11. Based on the RoseTTAFold three-track structure (see Figure 1), All-Atom was trained using protein–small molecule, protein-metal and covalently modified protein complexes found in the PDB. The resulting model can handle full biological assemblies, described by first author Rohith Krishna as “like switching from black and white to a colour TV”12.
Figure 1: General biomolecular modelling with RoseTTAFold All-Atom (RFAA). RFAA takes input information about the molecular composition of the biomolecular assembly to be modelled, including protein amino acid and nucleic acid base sequences, metal ions, small molecule bonded structure, and covalent bonds between small molecules and proteins. Sourced from Krishna et al., 202313.
Notably, this advanced capability also resulted in a significant advancement in the group’s RFdiffusion protein design tool.
AlphaFold3
However, All-Atom is not the only model capable of advanced biomolecule structural prediction. This year, in collaboration with digital biology company Isomorphic Labs, Google DeepMind launched a groundbreaking update to their structure prediction model: AlphaFold314. Building on their previous work, AlphaFold3 moves beyond proteomic analysis to support the structural modelling of a range of biomolecules, including DNA, RNA and small molecule ligands – presenting a massive step forward for drug development. Compared to AlphaFold2, their newest model features an improved Evoformer module and the use of a diffusion network, which is similar to those used in AI image generation.
The model may be DeepMind’s most powerful tool yet, but its unusual release caused some controversy within the scientific community. Unlike its predecessor, its usage is far more limited with the model only being available to non-commercial researchers in the form of a webserver, where users can submit up to 10 jobs a day. Importantly, AlphaFold3 was published without its source code or weights provided, which is a regular expectation for all research published in Nature journals15. In an open letter, some researchers questioned the reproducibility of the work, stating, “This does not align with the principles of scientific progress” and that the lack of available code “compromises peer review”16.
In response, the authors of the work quickly course corrected, with DeepMind’s vice-president of research, Pushmeet Kohli, claiming that the code would be made available for academic use within six months15. However, this controversy has prompted other labs to begin work on an open-source AlphaFold3 alternative17, highlighting the need for open-source and accessible tools to drive scientific progress forwards.
OpenFold: Teaching an Old Dog New Tricks
The world of structural biology was transformed by the introduction of AlphaFold2, a breakthrough in predicting protein structures with unprecedented accuracy. However, the implementation of AlphaFold2 left a gap: the lack of accessible code and data necessary for training new models. This limitation has impeded the exploration of new applications, such as predicting protein–ligand complex structures, and understanding the model’s learning processes.
To overcome this problem, the OpenFold consortium was created; a non-profit organisation on a mission to develop free and open-source software tools to catalyse innovation in biological research and drug discovery. And to start, they set their sights on creating state-of-the-art AI-based protein modelling tools that can predict molecular structures with atomic accuracy.
Published in Nature Methods in May 2024, OpenFold was introduced as a fast, memory-efficient and fully trainable implementation of AlphaFold2, built from the ground up to match its predecessor’s accuracy. What sets OpenFold apart is its robustness in generalising, even when faced with a deliberately constrained and diverse training set. This model not only enhances our understanding of protein folding but also paves the way for tackling new challenges, making it a pivotal resource for the protein modelling community.
Want to know more about the use of AI tools in proteomics? Check out part two of this feature, where we’ll explore protein design, publishing soon. And don’t forget to check out our comprehensive Proteomics Playbook when it comes out later this year.
References
1. Chaaban, S., Ratkevičiūtė, G. & Lau, C. AI told you so: Navigating protein structure prediction in the era of machine learning. Biochem. 46, 7–12 (2024).
2. Sample, I. Google’s DeepMind predicts 3D shapes of proteins. The Guardian (2018).
3. Senior, A. W. et al. Improved protein structure prediction using potentials from deep learning. Nature 577, 706–710 (2020).
4. Jumper, J. et al. Highly accurate protein structure prediction with AlphaFold. Nature 596, 583–589 (2021).
5. Bertoline, L. M. F., Lima, A. N., Krieger, J. E. & Teixeira, S. K. Before and after AlphaFold2: An overview of protein structure prediction. Front Bioinform 3, 1120370 (2023).
6. Evans, R. et al. Protein complex prediction with AlphaFold-Multimer. bioRxiv 2021.10.04.463034 (2022) doi:10.1101/2021.10.04.463034.
7. Zidek, A. Technical_note_v2.3.0.Md. https://github.com/google-deepmind/alphafold/blob/main/docs/technical_note_v2.3.0.md (2022).
8. Hassabis, D. AlphaFold reveals the structure of the protein universe. Google DeepMind https://deepmind.google/discover/blog/alphafold-reveals-the-structure-of-the-protein-universe/?_gl=1*ch8il2*_up*MQ..*_ga*MTYwMDY4MzA0MC4xNzE2ODg2Mjc2*_ga_LS8HVHCNQ0*MTcxNjg4NjI3NS4xLjAuMTcxNjg4NjI3NS4wLjAuMA.. (2022).
9. A glimpse of the next generation of AlphaFold. Google DeepMind
10. Baek, M. et al. Accurate prediction of protein structures and interactions using a three-track neural network. Science 373, 871–876 (2021).
11. Krishna, R. et al. Generalized biomolecular modeling and design with RoseTTAFold All-Atom. Science 384, eadl2528 (2024).
12. Institute for Protein Design, University of Washington. Introducing All-Atom versions of RoseTTAFold and RFdiffusion. Institute for Protein Design https://www.ipd.uw.edu/2023/10/introducing-rosettafold-and-rfdiffusion-all-atom/ (2023).
13. Krishna, R. et al. Generalized biomolecular modeling and design with RoseTTAFold All-Atom. bioRxiv 2023.10.09.561603 (2023) doi:10.1101/2023.10.09.561603.
14. Abramson, J. et al. Accurate structure prediction of biomolecular interactions with AlphaFold 3. Nature (2024) doi:10.1038/s41586-024-07487-w.
15. AlphaFold3 — why did Nature publish it without its code? Nature Publishing Group UK http://dx.doi.org/10.1038/d41586-024-01463-0 (2024) doi:10.1038/d41586-024-01463-0.
16. Wankowicz, Stephanie Beltrao, Pedro Cravatt, Benjamin Dunbrack, Ronald Gitter, Anthony Lindorff-Larsen, Kresten Ovchinnikov, Sergey Polizzi, Nicholas Shoichet, Brian Fraser, James. AlphaFold3 transparency and reproducibility. https://zenodo.org/records/11206103 (2024) doi:10.5281/zenodo.11206103.
17. Callaway, E. Who will make AlphaFold3 open source? Scientists race to crack AI model. (2024) doi:10.1038/d41586-024-01555-x.
