Protein Structure Prediction Tutorial

This tutorial walks you through predicting the 3D structure of a small protein from scratch using the BV-BRC Protein Structure Prediction Service. We’ll use crambin — a 46-amino-acid plant seed protein and one of the classic “hello-world” structures in computational biology. Its small size means every tool you choose returns a result in minutes, which makes it ideal for learning.

What you’ll do

1. Sign in to BV-BRC.

2. Upload a FASTA file with the crambin sequence to your workspace.

3. Submit a prediction job using the Auto tool selector.

4. Wait for the job to complete.

5. Open the result, interpret the confidence metrics, and download the structure.

Prerequisites

• A free BV-BRC account (register here)

• A web browser

You do not need a GPU, an MSA, or any local software. Everything happens on the BV-BRC servers.

Running locally? If you’re running a local copy of the BV-BRC web UI, open http://localhost:3000/app/PredictStructure instead of the production URL below. The form, the workspace, and the result viewer behave identically — the job runs on the real BV-BRC backend either way.

Background — what we’re predicting

Crambin is a hydrophobic 46-residue protein from the Crambe abyssinica seed. Its experimental structure (PDB ID 1CRN) has been solved at near-atomic resolution, which means we have ground truth to compare against. The sequence:

TTCCPSIVARSNFNVCRLPGTPEALCATYTGCIIIPGATCPGDYAN

It contains three disulfide bonds (Cys3–Cys40, Cys4–Cys32, Cys16–Cys26) and forms a compact α-helix / β-sheet fold. If a tool gets crambin wrong, something is seriously broken; if it gets it right, that tells you very little — crambin is in every training set. We use it here because it’s small and fast, not because it’s a hard test.

For a deeper introduction to what the prediction is actually doing, see the Protein Folding Primer (companion doc).

Step 1 — Sign in

Open https://alpha.bv-brc.org in your browser. Click Sign In in the top-right corner.

Sign in with your BV-BRC credentials.

Step 2 — Upload the crambin FASTA

Save the following as crambin.fasta on your computer:

>1CRN|Crambin|46aa
TTCCPSIVARSNFNVCRLPGTPEALCATYTGCIIIPGATCPGDYAN

In BV-BRC:

1. Click Workspaces → My Workspace in the top nav.

2. Navigate into the folder you want to use (or create a new one called tutorial-crambin).

3. Click Upload in the green Action Bar on the right.

4. Choose FASTA Sequence File, select crambin.fasta, and click Start Upload.

The file appears in the folder once upload completes.

Step 3 — Open the Protein Structure Prediction service

From the main menu, choose Services → Protein Tools → Protein Structure Prediction.

Or go directly to: https://alpha.bv-brc.org/app/PredictStructure

You should see the input form:

The fields are documented in detail in the Quick Reference. For this tutorial we’ll fill in only the basics.

Step 4 — Fill in the form

Field	Value
Prediction Tool	Auto
Protein	click the workspace selector, choose crambin.fasta
DNA / RNA	leave empty
Ligands	leave empty
SMILES	leave empty
MSA File	leave empty
Output Folder	click the folder selector, choose tutorial-crambin
Job Name	type crambin-auto (the field is pre-filled with PredictStructure-<timestamp> — overwrite it)

A submission creates a job with the name you enter. The job appears in your workspace as an object at <Output Folder>/<Job Name> — for this tutorial, tutorial-crambin/crambin-auto. That object holds all job-related info: parameters, status, logs, and the prediction results.

The Result location bar directly below the Job Name field previews this path as you type. The form runs live validation; the Submit button enables once you have an Output Folder, a Job Name, at least one biomolecular input, and — for Boltz / OpenFold / Chai — an MSA file.

Why Auto? With only a protein and no MSA, the auto-selector picks ESMFold — it’s the only engine that runs without an MSA on a single sequence (see the tool selector decision tree). If you pick boltz or chai without uploading an MSA, the submission will fail with a policy error.

Step 5 — Submit

Click Submit at the bottom of the form. A confirmation toast appears in the lower-left and the new job shows up in My Jobs (top-right user menu).

Step 6 — Wait

Crambin via ESMFold on a GPU runs in roughly 15–60 seconds plus queue time. Refresh the Jobs list, or open the job to see its live status:

Status	Meaning
queued	Waiting for a worker
in_progress	Running
completed	Output is in your workspace
failed	Something went wrong; click into the job to see stderr

Step 7 — Open the result

Click the completed job. The workspace object (tutorial-crambin/crambin-auto) opens with this layout (full reference: Quick Reference → Output Results):

crambin-auto/
├── results.json
├── report.html             ← interactive 3D viewer
├── inputs/
│   └── query.fasta
├── predictions/
│   ├── rank_1.pdb          ← the predicted structure
│   └── rank_1.cif
├── reports/
│   ├── confidence.json
│   ├── plddt.csv
│   └── pae.png
└── metadata/
    ├── tool.json
    └── runtime.json

View the structure

Click report.html and choose View from the Action Bar. A 3D viewer opens (3Dmol.js) showing crambin folded into its characteristic small α/β fold, with three disulfide bridges visible if you toggle the sticks representation for cysteines.

Read the confidence

Open reports/confidence.json (or look at the panel beneath the 3D viewer in report.html):

{
  "tool": "esmfold",
  "plddt_mean": 87.4,
  "plddt_min": 71.2,
  "ptm": 0.81,
  "model_confidence": 0.83
}

Interpreting these:

• plddt_mean = 87.4 — high confidence. A value above 80 on a 46-residue monomer says ESMFold is confident across the whole chain.

• plddt_min = 71.2 — even the worst residue is in the “confident” band (>70). No disordered tails.

• ptm = 0.81 — the overall fold is almost certainly right.

If you compare the rank-1 PDB to the experimental 1CRN structure (e.g. via TM-align), you should see a TM-score above 0.9 — ESMFold reproduces crambin essentially perfectly. As noted above, this isn’t a stress test; it confirms the pipeline works end-to-end.

Download the structure

From the Action Bar, choose Download on predictions/rank_1.pdb to save it locally. Open it in PyMOL, ChimeraX, Mol*, or any PDB viewer.

What to try next

Variation	What to change	What you’ll learn
Same protein, different tool	Pick boltz and upload crambin.a3m (a pre-computed MSA — examples ship with the test data)	How a richer MSA affects diffusion-tool confidence
Multi-chain complex	Submit a multi-record FASTA (chains A and B in one file). Crambin doesn’t naturally dimerize — try a known dimer like the Cro repressor instead	Multimer prediction and ipTM interpretation
Protein + ligand	Add ligand: ATP and pick boltz (with an MSA)	Co-folding with a small molecule, PoseBusters-style
Compare engines side-by-side	Run the CWL workflow multi-tool-comparison.cwl from the PredictStructureApp repo	Quantitative head-to-head on your own targets

Troubleshooting

No inputs supplied

You hit Submit with the Protein, DNA, RNA, ligand, and SMILES fields all empty. Provide at least one.

Invalid ligand CCD code 'ABCD'

CCD codes are 1–3 alphanumeric characters. Use ATP, not ATPase.

Boltz / OpenFold / Chai require an MSA

Set MSA Source to Precomputed MSA from Workspace and upload an .a3m, .sto, or .pqt file, or set it to Use MSA Server or Service to have BV-BRC compute the MSA with ColabFold. Auto and ESMFold don’t need an MSA at all.

Job is queued for a long time

GPU partitions are shared. ESMFold jobs have low resource requests and usually start within minutes; Boltz / OpenFold / Chai may wait longer. AlphaFold 2 jobs can wait significantly longer because they hold a GPU for an hour or more.

Result viewer doesn’t render

report.html loads 3Dmol.js from https://3dmol.org/build/3Dmol-min.js (a CDN, not an embedded script). If the viewer appears blank:

• Network can’t reach 3dmol.org — check connectivity, corporate proxy / firewall rules, or a captive portal.

• Ad-blocker or content-script extension is blocking the CDN — try a private/incognito window or disable extensions for the workspace domain.

• Strict Content-Security-Policy in the browser is rejecting the script — open the developer console (Cmd-Opt-J / Ctrl-Shift-J) and look for Refused to load the script or Blocked by CSP errors.

The underlying structure file (predictions/rank_1.pdb) still downloads and opens in PyMOL, ChimeraX, or any local viewer if you’d rather skip the embedded viewer entirely.

See also

• Quick Reference — every form field documented

• Recipes — patterns for multi-chain complexes, ligands, MSAs, reruns

• Protein Structure Prediction Service — open the form