A new computational tool from the National University of Singapore combines artificial intelligence with physics-based simulations to predict protein structures more accurately than AlphaFold, the tool that won a Nobel Prize in 2024.
D-I-TASSER, published in Nature Biotechnology, achieved approximately 13% greater accuracy than AlphaFold2 and AlphaFold3 on complex multi-domain proteins - the large proteins with multiple connected parts that existing AI tools struggle to model reliably.
How It Works
Where AlphaFold relies primarily on deep learning, D-I-TASSER takes a hybrid approach. The tool slices large proteins into smaller tractable sections, predicts each fragment’s structure individually, then reassembles them using physics-based constraints to ensure the final 3D structure makes physical sense.
“For most proteins, we still do not know their 3D structures, and that remains a major blind spot in biology,” said Professor Zhang Yang, who led the development team. “When we can see a protein’s structure more clearly, we can better understand what goes wrong in disease.”
The physics component addresses a key limitation of pure machine learning approaches: they can produce structures that look plausible but violate fundamental physical principles. By constraining the assembly process with physical modeling, D-I-TASSER produces more coherent three-dimensional forms.
Why This Matters for Drug Discovery
A protein’s shape determines its function. Many diseases occur when proteins misfold or interact abnormally with other molecules. Accurate structure prediction accelerates drug design by revealing binding sites and interaction mechanisms that would otherwise require expensive and time-consuming experimental methods like X-ray crystallography or cryo-electron microscopy.
The tool has already generated reliable structural models for most human proteins, including many that were previously difficult to analyze with existing methods. Multi-domain proteins - which include many disease-relevant targets - have been particularly challenging because their connected parts move and interact in complex ways.
The Open-Source Landscape
D-I-TASSER joins a growing competitive field of protein structure prediction tools. Just days ago, ByteDance released Protenix-v1, an open-source model that also claims to outperform AlphaFold3. The race to build better structure prediction tools reflects both the scientific importance of the problem and the limitations of current approaches.
AlphaFold transformed structural biology when it was released in 2021, correctly predicting structures for most single-domain proteins. But complex proteins, RNA structures, and protein-protein interactions remain active research frontiers where significant improvements are still possible.
What Comes Next
The research team is extending D-I-TASSER’s framework to RNA structure prediction and protein-protein interaction modeling, with particular focus on antibody-antigen complexes relevant to drug and vaccine development. They also plan to model dynamic protein folding pathways - how proteins actually fold over time rather than just their final structures.
The tool is freely available for academic research through the Zhang Lab at NUS, continuing the open science tradition that has characterized much of the structural biology field.