New AlphaFold update means we can predict bigger more complex proteins with ease

AF_unmasked, a new version of AlphaFold, enhances prediction accuracy for large protein complexes by integrating experimental data.

A new advancement to the artificial intelligence model AlphaFold, known as AF_unmasked, is transforming protein structure prediction by tackling the long-standing challenge of accurately modelling large protein assemblies. This development, led by researchers from Linköping University and Stockholm University, marks a significant leap in predicting multimeric complexes—structures composed of multiple proteins—which are crucial in biological processes and disease mechanisms.

Improving AlphaFold’s reach in complex modelling

Since its initial release by DeepMind in 2020, AlphaFold has dramatically improved single-protein structure prediction. However, large multimeric complexes, especially those without evolutionary templates, posed a limitation for AlphaFold, as the model’s predictions were often computationally expensive and lacked reliable accuracy for these complex formations. AF_unmasked addresses this limitation by incorporating experimental data directly into the prediction process, enhancing AlphaFold’s accuracy and efficiency for multimeric proteins.

Lead researcher Claudio Mirabello and his team developed AF_unmasked to integrate experimentally obtained structural templates, such as those from cryo-electron microscopy (cryo-EM) and X-ray crystallography, into the AlphaFold model without needing extensive re-training. This integration allows AF_unmasked to produce highly reliable models for complex structures of up to 10,000 amino acids, improving both the speed and quality of predictions, especially in cases where limited evolutionary information is available.

Enhanced accuracy through structural inpainting

One of the main innovations of AF_unmasked is its ability to perform “structural inpainting.” By combining experimental templates with AlphaFold’s learned energy potential, the model fills in missing regions within experimental structures, enhancing its flexibility in predicting protein dynamics and interface interactions. The study demonstrated that even imperfect experimental templates, containing incomplete structural data, could be refined by AF_unmasked. For instance, in trials involving challenging CASP15 targets, the model generated near-perfect predictions with a DockQ score above 0.8, a benchmark of high structural accuracy.

AF_unmasked was tested across various use cases, including both well-characterised multimeric complexes and those with high flexibility or missing regions. The research team found that in scenarios where traditional AlphaFold struggled, such as with large, flexible structures, AF_unmasked’s inpainting capabilities successfully resolved missing sections and accurately modelled regions with conformational variability. This functionality is especially valuable for understanding dynamic protein interactions, as seen in proteins like ClpB and neurofibromin, whose large, flexible domains are critical to their biological roles.

Applications in biomedical research

This innovation opens new possibilities for exploring protein complexes involved in cellular processes, offering researchers a practical tool to model larger, more complex assemblies with high accuracy. For biomedical research, AF_unmasked’s ability to integrate experimental data can facilitate studies on protein-protein interactions, drug target identification, and the effects of genetic mutations on protein structure. Proteins like neurofibromin, linked to tumour suppression, and Rubisco, a key enzyme in carbon fixation, are among the many where AF_unmasked can significantly enhance structure-function insights.

In addition, AF_unmasked’s capability to predict mutation impacts on protein interfaces promises valuable applications in genetic research. For instance, the model accurately predicted structural rearrangements in isocyanide hydratase following a single-point mutation, highlighting its utility in exploring disease mechanisms at the molecular level.