Skip to main content

From target to PDB: design a protein binder end to end

This tutorial walks through a complete binder discovery workflow using UbiMCP and UbiTools together. You will start with a disease name, identify a tractable target, pull its structure from PDB, design a protein binder against it, and predict the binder's 3D structure — all by prompting Claude with copy-pasteable messages.

By the end you will have:

  • A ranked list of targets for a disease of your choice
  • A PDB file for the target's binding domain
  • A set of de novo binder sequences designed against that domain
  • A predicted 3D structure of the top binder

What you need before starting:

The tutorial uses IL-6 receptor signaling as the worked example. Swap in any target you care about.

Step 1: Find a target​

Open Claude Desktop. Start a new conversation and paste this prompt exactly:

I'm researching therapeutic targets for rheumatoid arthritis.
Use UbiMCP to search Open Targets for the top five targets associated
with rheumatoid arthritis, ranked by genetic association score.
For each target, give me the gene name, Ensembl ID, and a one-sentence
description of why it's considered a target.

Claude will call execute_tool on the opentargets server and return something like:

1. IL6R — ENSG00000160712
   The IL-6 receptor is the primary signaling node for IL-6, a cytokine
   central to RA inflammation. Tocilizumab, an approved anti-IL-6R antibody,
   validates it as a tractable target.

2. JAK1 — ENSG00000162434
   ...

Pick your target. For this tutorial we proceed with IL6R.

Step 2: Get the binding domain structure​

Now you need a PDB structure for the extracellular domain of IL6R — specifically the domain that antibodies bind. Paste:

Using UbiMCP, search the PDB for structures of human IL-6 receptor
(gene: IL6R, UniProt: P08887) that include the extracellular D1-D2 domain
and have resolution better than 2.5 angstroms. List the top three hits
with PDB ID, resolution, and whether a ligand or antibody is co-crystallized.

You will get a table like:

PDB ID  Resolution  Co-crystallized ligand
------  ----------  ----------------------
4ZXW    1.9 Å       Sarilumab Fab (antibody)
4CNI    2.4 Å       IL-6 + gp130 complex
3L15    2.2 Å       Tocilizumab Fab (antibody)

Structure 4ZXW is a good choice: high resolution, no large complex that would obscure the binding surface. Note the PDB ID — you will use it in the next step.

Step 3: Identify the epitope residues​

Before designing a binder you need to know where to target. Paste:

Using UbiMCP, retrieve the structure 4ZXW from PDB and identify the
interface residues on IL6R that contact the antibody. List the residue
numbers and names in a compact format I can pass to a design tool.

Claude will parse the PDB contacts and return something like:

IL6R epitope residues (chain A, within 4.5 Å of antibody):
Arg47, Thr49, Lys53, Asp56, Glu57, Leu58, Arg179, Ser180, Gln182,
Asn183, Phe229, Glu233

As a residue number list: [47, 49, 53, 56, 57, 58, 179, 180, 182, 183, 229, 233]

Keep that residue list. If the response does not include a plain number list, ask:

Give me just the residue numbers as a Python list, no chain letters.

Step 4: Design binder sequences​

Now switch to UbiTools. Paste:

Using UbiTools, design 10 protein binders targeting IL6R using the
structure 4ZXW. The epitope residues to target on chain A are:
[47, 49, 53, 56, 57, 58, 179, 180, 182, 183, 229, 233]

Use design_antibody_cdrs to generate CDR loops that dock against this
epitope. Return the top 3 designs ranked by predicted binding score,
showing the full heavy chain sequence for each.

This call is compute-intensive and will take two to five minutes. Claude will show progress updates as UbiTools runs dyMEAN on the backend.

When it finishes you will get sequences like:

Design 1 (score: -8.4 kcal/mol)
Heavy chain:
EVQLVESGGGLVQPGGSLRLSCAASGFTFSSYAMSWVRQAPGKGLEWVSAISGSGGSTYYADSVKGRFTISRDNSKNTLYLQMNSLRAEDTAVYYCAKDYYYYGMDVWGQGTTVTVSS

Design 2 (score: -7.9 kcal/mol)
...

Design 3 (score: -7.6 kcal/mol)
...

Copy the heavy chain sequence for Design 1.

Step 5: Predict the binder structure​

Now predict the 3D structure of your best design so you can inspect it. Paste:

Using UbiTools, predict the 3D structure of this antibody heavy chain
sequence using predict_antibody_structure. Return the PDB-format structure.

Sequence:
EVQLVESGGGLVQPGGSLRLSCAASGFTFSSYAMSWVRQAPGKGLEWVSAISGSGGSTYYADSVKGRFTISRDNSKNTLYLQMNSLRAEDTAVYYCAKDYYYYGMDVWGQGTTVTVSS

Replace the sequence above with your actual Design 1 sequence.

UbiTools will run ImmuneBuilder or ABodyBuilder3 (whichever produces higher confidence for this chain type) and return a PDB block:

ATOM      1  N   GLU A   1     ...
ATOM      2  CA  GLU A   1     ...
...
END

Ask Claude to save this to a file:

Save that PDB content to a file called il6r_binder_design1.pdb

Claude will write the file to your Desktop (or current directory). You can open it in PyMOL, ChimeraX, or any structure viewer.

Step 6: Sanity check the design​

Before wrapping up, do a quick developability check. Paste:

Using UbiTools, run assess_developability on this heavy chain sequence.
Flag any aggregation, viscosity, or humanness concerns.

Sequence:
EVQLVESGGGLVQPGGSLRLSCAASGFTFSSYAMSWVRQAPGKGLEWVSAISGSGGSTYYADSVKGRFTISRDNSKNTLYLQMNSLRAEDTAVYYCAKDYYYYGMDVWGQGTTVTVSS

A clean result looks like:

Humanness score: 82% (OASis, threshold: 80% — PASS)
Predicted viscosity at 100 mg/mL: low (DeepViscosity score: 0.12)
Aggregation hotspots: none above threshold
Overall: developable

If any flag fails, ask Claude to suggest mutations:

The humanness score is below threshold. Suggest three point mutations
that would improve humanness without disrupting the CDR-H3 loop.

What you built​

In this tutorial you:

  1. Used UbiMCP to find and rank targets from Open Targets
  2. Retrieved a high-resolution PDB structure through the PDB server
  3. Extracted interface residues to define the binding epitope
  4. Designed antibody CDR sequences against that epitope using dyMEAN via UbiTools
  5. Predicted the 3D structure of the top design using ImmuneBuilder/ABodyBuilder3
  6. Ran a developability check to catch manufacturability issues early

The whole workflow — literature to designed binder with a structure — took a single Claude conversation and under ten minutes of wall-clock time.

Next steps​

  • Refine with affinity optimization: use optimize_antibody_binding to introduce additional mutations guided by the complex structure.
  • Search for similar known antibodies: use search_antibodies against the Observed Antibody Space to see if anything similar exists in the literature.
  • Try a nanobody: the same workflow works for VHH — omit light_chain and pass your sequence as nanobody instead.
  • Scale up: batch_number_sequences and the REST API let you run hundreds of sequences in parallel outside of the LLM conversation.