I-TASSER Tutorial: Predict Structure and Read the C-score

To use I-TASSER, register on zhanggroup.org with an institutional email, paste a FASTA sequence of 10 to 1500 residues, and wait for up to five models. Use model 1 unless you have a biological reason not to. Its C-score (range -5 to 2, higher is better) gives an estimated TM-score; a TM-score above 0.5 indicates the correct fold.

SWISS-MODEL is quick when your protein has a close relative in the PDB, and AlphaFold has changed what students expect from a predicted structure. I-TASSER still earns a place in a thesis workflow: it builds models by threading your sequence through many templates and reassembling fragments, it reports a confidence score designed to estimate how far the model is from the true structure, and it adds function predictions (ligand-binding sites, EC numbers, GO terms) on the same results page. This guide from the StemSkills Lab team (10+ years in structural bioinformatics, drug design and molecular modelling) walks through registration, submission, every section of the output, and the mistakes that cost beginners a week of waiting.

What is I-TASSER and how does it build a model?

I-TASSER (Iterative Threading ASSEmbly Refinement) is a protein structure and function prediction server from the Zhang group, now hosted at zhanggroup.org/I-TASSER. It works in three stages:

1. Threading. The meta-server LOMETS runs several threading programs that align your sequence against structures in the PDB, and I-TASSER keeps the most significant template alignments.
2. Assembly. Fragments cut from those templates are reassembled by replica-exchange Monte Carlo simulations into thousands of full-length conformations, called decoys. Regions with no template are built from scratch.
3. Clustering and refinement. The SPICKER program clusters the decoys by structural similarity. Up to five models are reported, one per largest cluster, then refined at atomic level.

The method is described in the Zhang group’s protocol paper, Roy, Kucukural and Zhang, “I-TASSER: a unified platform for automated protein structure and function prediction”, Nature Protocols 5:725-738 (2010), and the standalone suite in Yang et al., Nature Methods (2015). The server’s home page reports that it has completed predictions for 834,030 proteins submitted by 215,148 users from 169 countries (counter read in November 2026), and that it was ranked as the top server for structure prediction across a series of CASP experiments under the name “Zhang-Server”.

When should you use I-TASSER instead of SWISS-MODEL or AlphaFold?

Check the faster routes first. If the AlphaFold Database already has a confident model of your UniProt entry, or SWISS-MODEL finds a template with good identity and coverage, those models usually come back much sooner. I-TASSER is worth the wait when:

• SWISS-MODEL finds only distant templates with low identity, and you need a second, independently built model to compare;
• you want structure-based function hints (binding-site residues, EC number, GO terms) alongside the model;
• your supervisor or a reviewer asks for a model built by a classical template-based method, or for a consensus between methods;
• you need to exclude close homologues on purpose, for example to test how well a method does without the obvious template.

For a longer comparison of predicted and experimental structures, see AlphaFold vs homology modelling vs experimental structures.

How do you register for I-TASSER?

Registration is mandatory, and it is where many students in India get stuck. The registration page states that the servers are for academic use only and that the system does not accept free or commercial email domains such as @gmail.com or @hotmail.com. Use your university or institute address.

1. Open the Registration link on the I-TASSER page and fill in your name, institution and institutional email.
2. A confirmation email with your password arrives automatically. The page warns that these emails often land in spam or junk, so check there first.
3. If nothing arrives an hour after registering and checking spam, the page asks you to contact the group with the exact email address you registered.

The same login works for D-I-TASSER, the group’s newer deep-learning version. Downloading the standalone package needs a separate registration.

How do you submit a sequence to I-TASSER?

1. Get a clean FASTA sequence. Use the canonical sequence from UniProt, trimmed to the construct you care about. Remove signal peptides or long disordered tails if they are not part of your question; they add residues the server has to model with little template support.
2. Paste it into the form. The sequence must be within 10 to 1500 residues. Enter your registered email and password, and optionally a job ID so you can recognise the result.
3. Decide on the options. Most students leave these empty. Option I lets you add contact or distance restraints, or name a specific template by PDB ID. Option II excludes templates, either above a sequence identity cutoff you type (for example 30%) or from a list of PDB IDs. Option III lets you assign secondary structure to specific residues.
4. Choose public or private. Results are public by default. If you make the job private, you receive a key to access it; the form notes that many users lose this key, so save it.
5. Submit once. The form asks you to submit a new job only after your previous one has completed.

The Queue page shows where your job stands. Waiting time depends on server load and protein length, so plan your thesis timeline with days, not minutes, in mind.

What does each section of the I-TASSER results page mean?

The results page is long. Read it top to bottom in this order:

• Predicted secondary structure and solvent accessibility. Per-residue helix, strand and coil, and how buried each residue is predicted to be.
• Predicted normalized B-factor. A per-residue estimate of how flexible or uncertain each region is. High values often mark loops and termini you should treat with caution.
• Top 10 threading templates. For each template: Iden1 (sequence identity in the aligned region), Iden2 (identity over the whole template chain), Cov (aligned residues divided by query length) and Norm. Z-score. The output page states that a normalized Z-score above 1 means a good alignment. Several templates with Z-scores above 1 and high coverage is a good sign; all of them below 1 means your model leans heavily on the assembly simulation.
• Top 5 final models. Each with its C-score. Only model 1 gets an estimated TM-score and RMSD.
• Proteins structurally close to the target. TM-align matches model 1 against the PDB and lists the 10 closest structures with their TM-score, RMSD, identity and coverage.
• Predicted function using COACH. Ligand-binding sites, EC numbers and GO terms, each with its own confidence score.

How do you read the C-score, TM-score and RMSD?

C-score is I-TASSER’s confidence in a model, calculated from the significance of the threading alignments and the convergence of the assembly simulations. The I-TASSER About page gives its range as -5 to 2, with higher values meaning higher confidence.

TM-score measures how similar two structures are on a 0 to 1 scale. It was introduced by Zhang and Skolnick (Proteins, 2004) because RMSD is dominated by local errors: a misplaced tail can inflate RMSD even when the overall fold is right. According to the About page, a TM-score above 0.5 indicates a model of correct topology.

The link between the two is empirical. On a benchmark of 500 non-homologous proteins, the Zhang group found a correlation of 0.91 between the C-score of model 1 and its TM-score to the native structure, and 0.75 with its RMSD. That correlation is what lets the server turn a C-score into an estimated TM-score and RMSD, and it is also why those estimates are given only for model 1: the correlation is weak for lower-ranked models.

A worked example from the server’s own sample output: model 1 has C-score -0.08, estimated TM-score 0.70 ± 0.12 and estimated RMSD 6.3 ± 3.9 Å. Read that as “the fold is very likely right (even the low end, 0.58, is above 0.5), but individual regions may be several ångströms off”. That matters for what you do next. A correct fold is enough to discuss domain arrangement or map mutations; docking into a binding pocket needs the pocket itself to be accurate, so check the template coverage and B-factor there before you trust it.

Which of the five models should you pick?

Model 1, in almost every case. The models are ranked by cluster size, not by C-score, so you will sometimes see model 3 or 4 with a higher C-score than model 1. The About page addresses this directly: the C-scores of models 2 to 5 are listed for reference only and should not be compared, and “the first model is on average the most reliable and should be considered” unless you have a specific reason.

Good reasons to look past model 1 include known biology: a disulfide that must form, catalytic residues that must sit together, or a domain arrangement confirmed by experiment. If you choose another model, say why in your methods. Whichever you pick, run it through a structure validation check (Ramachandran plot, clashes) before using it, and consider refining the model if validation flags problems. Fewer than five models is not an error: the server’s example output explains that a converged simulation can produce fewer clusters, which usually indicates good model quality.

What goes wrong most often, and how do you fix it?

• Registration rejected. A Gmail, Yahoo or Hotmail address will not be accepted. Register with your institutional email. If your institute has no email system, ask your supervisor to register and submit on your behalf.
• No password email. Check spam or junk first; the registration page names this as the usual cause. After an hour, contact the group with the exact address you used.
• Sequence refused. It is outside 10 to 1500 residues, or the FASTA contains non-amino-acid characters, numbers or a second sequence. Paste a single clean sequence. For a protein longer than 1500 residues, model domains separately; the group’s I-TASSER-MTD platform is designed for multi-domain proteins.
• Second job not accepted. The form asks for one job at a time. Wait for the first to finish.
• Lost access to a private job. Private results need the key you were given at submission. Keep jobs public unless the sequence is confidential, and save the key if it is.
• Confusing the two C-scores. The model C-score runs from -5 to 2. The ligand-binding and EC predictions in the function section use separate confidence scores that range from 0 to 1. Do not compare them.
• Low-confidence model. If model 1 has a strongly negative C-score and every threading template has a normalized Z-score below 1, the model mostly reflects the simulation, not templates. Compare it with an AlphaFold model, and check pLDDT and PAE for the regions you care about.

How should you report an I-TASSER model in your thesis?

State the server and the date you ran it, the sequence range modelled, any options you set (restraints, excluded templates), which model you chose and why, and its C-score with the estimated TM-score and RMSD. Name the top templates with their identity and coverage. The server page asks users to cite specific articles, and the list has changed as new versions (C-I-TASSER, I-TASSER-MTD, D-I-TASSER) were published, so copy the references shown on the page on the day you submit. Add your validation results, then the downstream steps: refinement, preparation for MD or docking.

Frequently asked questions

Is I-TASSER free to use?

Yes, for academic, non-commercial use. You need to register with an institutional email address; free email domains are not accepted.

What is a good C-score in I-TASSER?

Higher is better on a range of -5 to 2. Rather than a fixed cutoff, read the estimated TM-score the server gives for model 1: if it stays above 0.5 even after subtracting its uncertainty, the fold is very likely correct.

Why does a lower-ranked I-TASSER model have a higher C-score than model 1?

Models are ranked by cluster size, not C-score, and the C-score only correlates well with quality for model 1. The server’s own guidance is to use model 1 unless you have a biological reason to prefer another.

Is I-TASSER better than AlphaFold?

For most single-chain proteins, AlphaFold models are the stronger starting point today. I-TASSER remains useful as an independent second model, for its function annotations, and when you want control over templates. The Zhang group’s newer D-I-TASSER server combines deep learning with I-TASSER assembly.

Can I use an I-TASSER model for docking or MD?

Yes, after validation. Check that the binding site is covered by a good template and has low predicted B-factor, fix any Ramachandran outliers, then prepare the structure as you would any other model.

Structure prediction is one stage of our computational biology skills roadmap, which shows where I-TASSER fits between sequence analysis and docking or MD.