The I-TASSER Suite: protein structure and function prediction
By Jianyi Yang, Renxiang Yan, Ambrish Roy, Dong Xu, Jonathan Poisson, Yang Zhang. 2014
5,315 citations recorded by the source corpus.
The lowest free-energy conformations are identified by structure clustering. A second round of assembly simulation is conducted, starting from the centroid models, to remove steric clashes and refine global topology. Final atomic structure models are constructed from the low-energy conformations by a two-step atomic-level energy minimization approach. The correctness of the global model is assessed by the confidence score, which is based on the significance of threading alignments and the density of structure clustering; the residue-level local quality of the structural models and B factor of the target protein are evaluated by a newly developed method, ResQ, built on the variation of modeling simulations and the uncertainty of homologous alignments through support vector regression training. For function annotation, the structure models with the highest confidence scores are matched against the BioLiP5 database of ligand-protein interactions to detect homologous function templates. Functional insights on ligand-binding site (LBS), Enzyme Commission (EC) and Gene Ontology (GO) are deduced from the functional templates. We developed three complementary algorithms (COFACTOR, TM-SITE and S-SITE) to enhance function inferences, the consensus of which is derived by COACH4 using support vector machines. Detailed instructions for installation, implementation and result interpretation of the Suite can be found in the Supplementary Methods and Supplementary Tables 1 and 2. The I-TASSER Suite pipeline was tested in recent communitywide structure and function prediction experiments, including CASP10 (ref. 1) and CAMEO2. Overall, I-TASSER generated the correct fold with a template modeling score (TM-score) >0.5 for 10 out of 36 “New Fold” (NF) targets in the CASP10, which have no homologous templates in the Protein Data Bank (PDB). Of the 110 template-based modeling targets, 92 had a TM-score >0.5, and 89 had the templates drawn closer to the native with an average r.m.s. deviation improvement of 1.05 Å in the same threadingaligned regions6. In CAMEO, COACH generated LBS predictions for 4,271 targets with an average accuracy 0.86, which was 20% higher than that of the second-best method in the experiment. Here we illustrate I-TASSER Suite–based structure and function modeling using six examples (Fig. 1b–g) from the communitywide blind tests1,2. R0006 and R0007 are two NF targets from CASP10, and I-TASSER constructed models of correct fold with a TM-score of 0.62 for both targets (Fig. 1b,c). An illustration of local quality estimation by ResQ is shown for T0652, which has an average error 0.75 Å compared to the actual deviation of the model from the native (Fig. 1h). The four LBS prediction examples (Fig. 1d–g) are from CASP10 (ref. 1) and CAMEO2; COACH generated ligand models all with a ligand r.m.s. deviation below 2 Å. COACH also correctly assigned the threeand fourdigit EC numbers to the enzyme targets C0050 and C0046 (Supplementary Table 3). In summary, we developed a stand-alone I-TASSER Suite that can be used for off-line protein structure and function prediction. The I-TASSER Suite: protein structure and function prediction
Subfields: Molecular Biology, Computational Biology & Bioinformatics
How researchers see this paper
The researchers whose trusted networks rank it highest, strongest advocate first. Every researcher gets a different order.
In Molecular Biology
In Computational Biology & Bioinformatics
Citations in the corpus
- Multi-State Modeling of G-protein Coupled Receptors at Experimental Accuracy (2021, neutral): “Among them, GPCR-I-TASSER 23 , RosettaGPCR 24 , and GPCRdb 25, 26 have resulted in structure databases of the whole GPCR proteome via computational modeling.”
- CopulaNet: Learning residue co-evolution directly from multiple sequence alignment for protein structure prediction (2020, neutral): “Thus, the computational approaches for predicting protein structures purely from residue sequences are highly desirable [4, 5].”
- I-TASSER server: new development for protein structure and function predictions (2015, neutral): “The communitywide blind CASP experiments have shown that many automated servers, such as I-TASSER (1,2), Rosetta (3) and HHpred (4), can now generate structural models with accuracy comparable to the best human-expert guided modeling (5–7).”
- End-to-end differentiable learning of protein structure (2018, background): “, 2015), but protein structure pipelines continue to resemble the ways in which computers tackled vision and speech prior to deep learning, by having many human-engineered stages, each independently optimized (Xu and Zhang, 2012; Yang et al., 2015) (Figure 1).”
- mlDEEPre: Multi-Functional Enzyme Function Prediction With Hierarchical Multi-Label Deep Learning (2019, methodology): “…task, a number of methods have been proposed to deal with the problem, based on structure similarity (Dobson and Doig, 2005; Roy et al., 2012; Yang et al., 2015), sequence similarity (Tian et al., 2004; Quester and Schomburg, 2011), or machine learning techniques (des Jardins et al., 1997;…”
- Detecting distant-homology protein structures by aligning deep neural-network based contact maps (2019, methodology): “5 as identified by TM-align [37] structural alignment search through the I-TASSER template library [5].”
- Deep-learning contact-map guided protein structure prediction in CASP13 (2019, methodology): “10-12 On the other hand, “ QUARK ” , which was introduced in CASP9, has been based on the QUARK 13,14 program, which was designed to fold proteins by assembling continuously-distributed fragments from unrelated protein structures under the guidance of a knowledge-based energy function.”
- Universal Transforming Geometric Network (2019, methodology): “by RGN and UTGN models in the FM category for CASP12. 4 DISCUSSION Before RGN was introduced, the protein prediction competition was dominated by complex models that fuse together multiple pipelines (Yang et al. (2014)). They tend to incorporate biological priors, like co-evolutionary information an”