Skip to main content
Purna can predict the 3D structure of a protein from its amino acid sequence and save the result as an interactive structure artifact in your session. You can also predict wildtype and mutant structures to assess how a specific variant might affect protein conformation. Structure predictions are powered by NVIDIA Boltz2, a state-of-the-art deep learning model for biomolecular structure prediction. Results are rendered using Mol*, the same viewer used by RCSB PDB and the AlphaFold Database.

​
Predicting a structure

Provide an amino acid sequence and ask Purna to predict its structure:
  • “Predict the structure of this sequence: MKWVTFISLLFLFSSAYS…”
  • “What would the 3D structure of BRCA1 RING domain look like?”
  • “Predict the structure from this FASTA file” (attach a .fasta file)
Purna validates the input and submits the sequence to Boltz2. The chat shows a compact submission entry while the prediction runs. Use Compute to follow its progress. When the prediction finishes, open the generated CIF file from Files to inspect it in the built-in 3D viewer.

​
Input requirements

Multi-chain complexes are supported — provide a multi-sequence FASTA file or describe the chains in your message.

​
The 3D viewer

The structure viewer supports standard molecular visualization controls:
  • Rotate — Click and drag to rotate the structure
  • Zoom — Scroll to zoom in and out
  • Pan — Right-click and drag to pan

​
Confidence coloring

Structures are colored by pLDDT (predicted Local Distance Difference Test), the per-residue confidence metric:
Regions with low pLDDT (yellow/orange) are often intrinsically disordered or flexible loops. High-confidence regions (blue) are the most reliable for structural interpretation.

​
Comparing wildtype vs mutant

Ask Purna to compare two structures:
  • “Predict the structure of TP53 wildtype and the R248W mutant, then compare them”
  • “How does the p.V600E mutation affect the structure of BRAF kinase domain?”
  • “Compare the structure of this sequence with and without the G12D mutation”
Purna can submit separate predictions for the wildtype and mutant sequences. When both runs finish, open their generated structure files from Files and compare their confidence patterns and regions of interest.

​
Comparison workflow

  1. Purna validates the wildtype sequence and the requested variant.
  2. It submits a prediction for each sequence.
  3. You can follow both runs from Compute.
  4. The completed structure files appear in Files, labeled for the corresponding sequence.

​
Downloading results

Open the predicted .cif file from Files, then use the download action to save it locally. The mmCIF file can also be opened in PyMOL, ChimeraX, VMD, or Mol* standalone.

​
How it works

  1. Validation — Purna checks that your sequence uses standard amino acids and is within the length limit.
  2. Submission — The sequence is sent to NVIDIA Boltz2 via their health API.
  3. Prediction — Boltz2 runs the structure prediction (async, typically 1-3 minutes).
  4. Storage — The resulting mmCIF file is stored in cloud storage for access.
  5. Rendering — Open the saved CIF file from Files to render the structure in Mol* with pLDDT coloring.
Identical sequences submitted within 10 minutes return the cached prediction — no duplicate computation.
Structure predictions are computational estimates, not experimental results. Always check the pLDDT confidence coloring. For high-stakes interpretations, validate against experimental structures from PDB when available.

​
When to use structure prediction vs existing structures

For well-studied proteins, check PDB and AlphaFold first — experimental and pre-computed structures are available instantly. Use on-demand prediction for novel sequences, custom mutants, or multi-chain complexes not in existing databases.