Wei Zheng, Qiqige Wuyun, Yang Li, Quancheng Liu, Xiaogen Zhou, Chunxiang Peng, Yiheng Zhu, L. Freddolino et al.
D-I-TASSER is a hybrid method that combines deep learning predictions with traditional physics-based folding simulations, achieving superior performance over AlphaFold2/3 in both single-domain and multidomain protein structure prediction.
Deep learning-based methods (e.g., AlphaFold) have achieved great success in protein structure prediction, but limitations exist in automated modeling of multidomain proteins and accuracy for some domains. Additionally, pure deep learning methods may overlook physical principles, sparking debate on the necessity of traditional force field simulations.
D-I-TASSER integrates multi-source deep learning potentials (e.g., contact prediction, distance distributions) into iterative threading fragment assembly simulations. It introduces a domain splitting and assembly protocol for automated modeling of multidomain proteins. This method combines the predictive power of deep learning with the sampling and energy optimization of physical simulations.
Benchmark tests and CASP15 experiments show D-I-TASSER outperforms AlphaFold2 and AlphaFold3 on both single-domain and multidomain proteins. Large-scale folding experiments demonstrate it can fold 81% of protein domains and 73% of full-chain sequences in the human proteome, with results highly complementary to AlphaFold2 models. This highlights the effectiveness of integrating deep learning with classical physics-based folding simulations for high-accuracy genome-wide protein structure and function predictions.