Inception of AlphaFold 3: Shining light from structure prediction to de novo design of biomolecules

Recently, Josh et al. investigate a new structure prediction tool, AlphaFold 3 model, which has successfully predicted the structure and interactions of all living molecules with unprecedented accuracy.¹ This is a breakthrough in the development of artificial intelligence (AI) tools (Figure 1).

AlphaFold 3 is able to achieve that scaling up the predictive power of a single deep learning model by an evolution of the AlphaFold 2 Evoformer and Structure Module. This prediction function involves complexes containing a more extensive range of biomolecules, such as proteins, nucleic acids, small molecules, ions, complexes modifying protein residues, and antibody-antigen interactions. It meets the gap of current AI technology for structure and interaction prediction, and its accuracy significantly exceeds that of existing specific interaction types prediction tools.² This proves that high accuracy modelling across biomolecular space is possible, promising to address the core challenge of molecular biology, which is to understand and finally regulate the complex atomic interactions in biological systems.

AlphaFold 3 introduces Diffusion Model as its core machine learning architecture, a model that has been successful in the field of AI image generation. Compared with the previous version of AlphaFold model, the diffusion model of AlphaFold 3 can directly generate the 3D coordinates of each atom, no longer relying on the structural modules of amino acid framework and side chain dihedral angles. This approach allows the model to be more flexible and intuitive in constructing 3D structures of proteins and their interacting partners. The principle of diffusion model is similar to the process of gradually get rid of the noise. And this model is able to provide confidence scores for its predictions, which helps to improve the predictive accuracy and confidence of the model. The design of the diffusion model considering the calculation efficiency and scalability, making AlphaFold 3 biomolecular systems that can handle the larger and more complex, this is particularly important for drug design and biological engineering, etc.

We can conclude that the progress of the AlphaFold system: (I) accurate model architecture and more powerful training algorithms, (II) faster and more efficient, and (III) expanded utility and universality.

AlphaFold 3 is, after all, a prediction tool, and there will still be some model limitations in terms of accuracy. In particular, those proteins with complex structural or dynamic properties may challenge the predictive power of AlphaFold 3. These prediction errors may be manifested in chiral violation rate, the case of atomic collisions, and spurious structural order during the prediction. The prediction results of AlphaFold 3 are highly dependent on the quality and quantity of the training data, which may influence the accuracy of predictions due to limitations or incompleteness of the data set. The prediction results are usually the static structure of the molecule, which makes it impossible to obtain a dynamic process of structural transformation between molecular interactions. Therefore, this may lead to uninterpretable results in AlphaFold 3 prediction, that is, the principle and mechanism behind the prediction result cannot be provided. This may cause errors in the exploration of the interaction of some newly discovered molecules, which is not a negligible existence in determining the experimental direction, and also a problem about applicability and credibility that needs to be solved in the future AlphaFold systems. It is necessary for us to ensure scientific research through both the experiment and calculation method.

Scientific research on using AlphaFold 3 production needs to be verified by experiments. Designing small molecules/ligands/ions based on targets that are difficult to target in diseases by AlphaFold 3, then researchers can predict their binding efficiency and then perform experimental proof, which may obtain a breakthrough in the treatment of malignant diseases.³ It also enables more precise risk assessments and targeted treatment recommendations for disease diagnosis. We look forward to using AlphaFold 3 or an array of tools to predict nucleic acid molecules that are directed to the target. A clip editor for gene editing has been demonstrated in AlphaFold 2 system.⁴ Meanwhile, great mysteries remain about noncoding RNAs. Researchers can also try to input RNA sequence references into AlphaFold 3, to explore the relationship between RNA sequences and secondary/tertiary structures, and even de novo design RNA vaccine sequences and predict efficiency through AI.⁵ This may be a way to save time and effort costs, and to get a reasonable simulation of the conjecture. RNA vaccine design and interaction prediction, including the relationship between sequence position and off-target side effects, to optimize the efficiency of RNA therapies. This may open up some new research directions, such as exploring new associations between protein structure and function and developing new drug/vaccine design strategies.

Taken together, AlphaFold 3 has more significant advantages over existing AI models in the de novo design area. AlphaFold 3 can more accurately predict the binding mode of antibodies/vaccine to their target proteins and the 3D structure, because they often contain specific proteins or their fragments that can trigger immune responses. Further, it helps scientists to find potential drug targets quickly and design new (unknown structure protein) vaccine, to shorten the drug research and development cycle, reduce costs. They may have better immunogenicity and collaborative drugs to improve treatment effect. During research, AlphaFold 3 could help researchers select or design more stable protein variants, these variations in the process of storage and transportation are less likely to degeneration, in the human body can keep active at the same time, thus improve the overall effect and safety of the functional biomolecules.

For the following application development of AlphaFold 3, this is an exciting and worth exploring topic. For proteins with known functions, researchers can design and customize protein structures with effective specific function by AlphaFold system. It promotes research in life sciences including protein interaction and cell signaling processes, and also provides novel ideas and targets for disease treatment. The de novo design of peptide/enzyme facilitates researchers to explore the influence of protein sequence differences on amino acid characteristics and protein conformation, and obtain customized peptide/enzyme design or even improve the efficacy of existing drugs. These directions have great potential in the fields of biomedicine and biomaterials.

AI technological breakthroughs following by AlphaFold 3 may stimulate more research and technological innovation in similar areas, to promote the progress of life science, development science, medicine, drug research and other fields. This will have a profound impact on public health and the pharmaceutical research industry in basic and clinical research, and promote the development of precision medicine. However, for the emergence of new intelligent tools, we still need to maintain a rational attitude, neither deny the progress of science and technology, but also discuss the shortcomings. AlphaFold 3 still has a certain degree of lag, including not dynamic simulation, not completely accurate and comprehensive, and not innovative prediction. It indeed helps us to greatly reduce the trial and error cost during the experiment, which is indeed a non-negligible advantage. We are delighted to accept this state-of-the-art AI tool, AlphaFold 3, and still looking forward to using it to for joint innovation with human intelligence.

Yue Hou and Yongye Huang provided the conception, funding support, revision and supervision. Sihui Zhang conducted the literature research and wrote the initial manuscript and drew the figure. Sihui Zhang and Yongye Huang are responsible for revision and proofreading of the manuscript. All authors have read and approved to publish the article.

The authors declare no conflict of interest.

Not applicable.