Paras Verma, Ph. D

Deep-learning-based single-domain and multidomain protein structure prediction with D-I-TASSER @NatureBiotech １．D-I-TASSER is a hybrid deep-learning and physics-based pipeline that outperforms AlphaFold2 and AlphaFold3 in both single-domain and multidomain protein structure prediction. ２．Unlike AlphaFold’s end-to-end deep learning, D-I-TASSER integrates multiple deep learning restraints (distance/contact/hydrogen-bonding) with fragment-based simulations, enabling superior structural refinement, especially for difficult targets. ３．For 500 nonredundant hard protein domains, D-I-TASSER achieved an average TM-score of 0.870 — 5% higher than AlphaFold2 and significantly higher than all other versions of AlphaFold, including v3. ４．On hard domains where AlphaFold2 struggled (TM < 0.8), D-I-TASSER showed a dramatic improvement (average TM = 0.707 vs. 0.598), successfully folding 63 cases where AlphaFold2 failed. ５．A key innovation is the domain-splitting and reassembly protocol, allowing D-I-TASSER to handle large multidomain proteins with better interdomain orientation than AlphaFold2. ６．On 230 multidomain proteins, D-I-TASSER achieved higher full-chain TM-scores in 88% of cases. The average TM-score improvement was 12.9% over AlphaFold2. ７．For large multidomain proteins (e.g. PDB 7jtkB, 6irdC), D-I-TASSER accurately modeled both interdomain and intradomain distances, where AlphaFold2 failed due to shallow MSAs. ８．DeepMSA2, the new iterative MSA generation module, is crucial to D-I-TASSER’s success. It builds deeper alignments and boosts model quality significantly, especially for orphan proteins. ９．In CASP15 blind tests, D-I-TASSER created correct folds for 95% of domains and outperformed AlphaFold2 and other top groups by a wide margin, particularly in interdomain modeling. １０．Unlike static AlphaFold predictions, D-I-TASSER produces diverse conformations, effectively capturing multiple states like open/closed forms in spike protein complexes. １１．D-I-TASSER was used to model over 95% of the human proteome, creating 34,968 domain-level and 19,512 full-chain structures. Many of these complement AlphaFold’s database coverage. １２．Estimated TM-score evaluations (eTM) indicate that D-I-TASSER’s models are robust, even in the absence of experimental structures, providing confidence for genome-scale structural annotation. １３．Overall, D-I-TASSER sets a new standard by combining deep learning with classical simulation, achieving high accuracy across difficult, large, and multidomain targets. 💻Code: zhanggroup.org/D-I-TASSER/do… 📜Paper: nature.com/articles/s41587-0… #DeepLearning #ProteinStructure #StructuralBiology #AlphaFold #DITASSER #Bioinformatics #CASP15 #HumanProteome #Multidomain #MSA