Proteins are fundamental to life. They catalyse chemical reactions, facilitate cell signalling, and provide structural support throughout cells. The human body contains around 20,000 different proteins, and while scientists have identified their amino acid sequences, the three-dimensional (3D) structures of most proteins remain unknown. This is a problem, because it is only through the precise folding of these amino acids that functionality is conferred.

In recent years, the field of structural biology has been revolutionised by computational tools that can accurately predict protein structure from amino acid sequences. The gold standard is AlphaFold2, a deep learning-based tool developed by Google DeepMind. It analyses patterns in related protein sequences from different organisms and predicts how the protein’s amino acids should be arranged in a 3D space. Since its release in 2020, AlphaFold2 has remained the benchmark for large-scale protein structure prediction. DeepMind released AlphaFold3 in 2024 with moderate improvements over AlphaFold2 in certain tasks.

At present, most computational methods perform best when there is abundant evolutionary information available for a protein and when the protein consists of only a single structural unit, or domain. However, the majority of proteins are made up of two or more domains.

A new hybrid tool: D-I-TASSER

Seeking a tool that can predict the 3D shape of human, multi-domain proteins, a team led by Professor Zhang Yang from NUS Computing and NUS Biochemistry developed D-I-TASSER (Deep learning-based Iterative Threading ASSEmbly Refinement). This hybrid tool combines deep learning and traditional physics-based modelling techniques. Their work was published in Nature Biotechnology.

Specifically, D-I-TASSER models the structure of a protein by performing folding simulations guided by AI-derived restraints, which provide information about which parts of the protein are likely to be close together and the distances between them. For multi-domain proteins, it breaks the protein into individual domains, models them separately, and then reassembles them into a 3D structure using advanced physics-based simulation techniques.

Tests on single-domain proteins showed that D-I-TASSER not only outperformed earlier versions of itself (I-TASSER ^[1] and C-I-TASSER ^[2] ), but also surpassed AlphaFold2 in accuracy. For multi-domain proteins, D-I-TASSER outperformed AlphaFold2 with 13% better accuracy in predicting whole protein structures, and 3% better on domains.

“D-I-TASSER’s strengths lies in its ability to combine multiple sources of structural restraints from both diverse AI models and template-based threading alignments. This multi-source strategy is further strengthened by the use of physics-based Monte Carlo simulations, which fine-tune the AI-generated structures to achieve greater atomic-level precision,” explained Prof Zhang.

The researchers also compared the coverage of D-I-TASSER to AlphaFold2 on the human proteome. AlphaFold2 could model nearly all human proteins (98.5%) as single chains, even those up to 2,700 amino acids long. D-I-TASSER focused on proteins up to 1,500 amino acids but modelled them in finer detail, covering about 95% of the proteome. While AlphaFold2 offered slightly broader proteome coverage, D-I-TASSER delivered higher overall structural accuracy, especially in multi-domain proteins.

Better understand human biology

These results can be attributed to D-I-TASSER’s domain-splitting approach, which allows for more effective modelling of complex, multi-domain proteins. D-I-TASSER also addresses gaps left by AlphaFold2, such as limited sequence information or high structural complexity.

With an ability to better predict the 3D structures of human multi-domain proteins, the work of D-I-TASSER is expected to help scientists better understand human biology and advance medical research, from disease studies to novel drug development.