Written and maintained by CASRAI Editorial Board
Last updated
I-TASSER (Iterative Threading ASSEmbly Refinement) is a Zhang Group web server and standalone package that predicts a protein’s three-dimensional structure from its amino-acid sequence and then infers function from that structure. This guide is for researchers who have one protein sequence and need to run it, read the confidence scores and decide whether the model is usable. Facts below come from the Zhang Group server and output-guide pages, read 2026-11-08.
What I-TASSER is used for in research
The server describes I-TASSER as a hierarchical approach to protein structure and structure-based function prediction. It first identifies structural templates with the LOMETS threading programs, assembles full-length models from template fragments by iterative simulation, and then infers function by re-threading the top model through the BioLiP protein-function database.
Typical research uses are building a working structural model of a protein with no experimental structure, and generating hypotheses about enzyme class, Gene Ontology terms and ligand-binding sites. The output is a prediction, not an experimental structure, and should be reported as one.
Install or access (official source only)
I-TASSER is run through the Zhang Group server at zhanggroup.org/I-TASSER. Registration (email and password) is required before submitting a job. The same page offers a standalone package (version 5.2 at the time of reading) through its download page.
Licence limit that decides whether you can use it: the server page states, “The server is only for non-commercial use.” Industry users should read the licence terms on the official download page before relying on either route.
Step-by-step: submit a sequence to the I-TASSER server
- Prepare the sequence in FASTA format. The server accepts sequences of 10 to 1,500 residues; longer proteins must be split into domains first.
- Register on the I-TASSER page if you do not already have an account, then enter your email address and password on the submission form.
- Paste or upload the FASTA sequence. The job ID field is optional and is only a label for your own records.
- Choose whether the job is public or private. A private job is accessed with a key.
- Optionally set the advanced options. Option I supplies contact or distance restraints, or names a template (by PDB ID or uploaded structure) with or without an alignment. Option II excludes templates above a sequence-identity cutoff (the page gives 60% as an example) or excludes specific PDB IDs listed in an uploaded file. Option III supplies secondary-structure assignments for chosen residues.
- Submit the job. Submit a new job only after the previous one has completed.
Excluding close homologs (Option II) is the documented way to test how well the method performs on your protein when no near-identical template should be available.
Output and how to read it
A completed job returns up to five full-length models ranked by cluster density, a confidence score for each, a predicted TM-score and RMSD for model 1, predicted secondary structure and solvent accessibility, the top ten LOMETS threading templates, the top ten closest PDB structures, and predicted EC numbers, GO terms and ligand-binding sites with confidence values.
How to read C-score and TM-score
| Score | What it measures | How to read it (output guide) |
|---|---|---|
| C-score | Confidence in a model, calculated from the significance of the template alignments and the convergence of the assembly simulation | Typically ranges from -5 to 2; a higher value means higher confidence. It correlates with the TM-score (0.91) and RMSD (0.75) of model 1. |
| TM-score | Structural similarity between the model and the native structure, independent of protein length | Above 0.5 indicates a model with generally the correct topology; below 0.17 indicates a result no better than random. |
| RMSD | Average distance between model and native atoms | Sensitive to local errors, so the guide recommends reading it together with TM-score. A predicted RMSD is given for model 1 only. |
Per the output guide, model 1 is on average the most reliable, and a result with fewer than five clusters suggests the simulation converged and the model quality is good. For function, the guide directs users to the COACH-based predictions rather than inferring function from model quality alone.
Troubleshooting
- Submission is rejected: check that the sequence is FASTA and between 10 and 1,500 residues, and that you are logged in with a registered email.
- A second job will not start: the server asks for a new job only after the previous one has completed.
- Low C-score: treat the model as low confidence. Use the Option I and Option III inputs only if you have experimental or computational evidence for restraints or secondary structure.
- Cannot use the server for company work: the server is limited to non-commercial use; see the alternatives below.
Free alternatives and related tools
| Tool | Notes |
|---|---|
| Phyre2 / Phyre2.2 | Imperial College protein structure prediction server. The Imperial College server offers Phyre2.2, cited as Powell HR et al., J Mol Biol 437:168960 (2025). |
| ColabFold | Protein structure prediction tool, covered in a separate CASRAI guide. |
| SWISS-MODEL | Homology-modelling server, covered in a separate CASRAI guide. |
| PyRosetta | Python interface to the Rosetta modelling suite, covered in a separate CASRAI guide. |
For precomputed models of known proteins, see the AlphaFold Protein Structure Database guide. For the crowdsourced puzzle approach to structure folding, see Foldit.
FAQ
What does I-TASSER stand for?
Iterative Threading ASSEmbly Refinement, the name of the Zhang Group method for protein structure and function prediction.
Is I-TASSER free?
The server page states it is for non-commercial use only. Commercial users should consult the licence information on the official site.
How long a protein can I submit?
Between 10 and 1,500 residues in FASTA format.
Which model should I use?
Model 1 is on average the most reliable according to the output guide; check its C-score and predicted TM-score before use.
What is Phyre2 and how does it compare?
Phyre2 is a separate structure prediction server from Imperial College London. The Imperial College server offers Phyre2.2 (Powell HR et al., J Mol Biol 437:168960, 2025); consult that site for current inputs and limits.
How do I cite I-TASSER?
The server page lists three references; see Sources.
Sources
- Zhang Group, I-TASSER server: zhanggroup.org/I-TASSER (read 2026-11-08).
- Zhang Group, I-TASSER output guide: zhanggroup.org/I-TASSER/about.html (read 2026-11-08).
- Zheng W et al. D-I-TASSER. Nature Biotechnology (2025), doi:10.1038/s41587-025-02654-4.
- Zhou X et al. I-TASSER-MTD. Nature Protocols 17:2326-2353 (2022).
- Zheng W et al. Cell Reports Methods 1:100014 (2021).
- Imperial College, Phyre2: sbg.bio.ic.ac.uk/~phyre2 (read 2026-11-08).
