Predict checkpoint-blockade response for a tumor from its biomarker profile

Turn a tumor’s molecular profile into an evidence-graded immune-checkpoint-inhibitor response assessment that names which biomarker is driving the call, so the axes disagree in the open rather than being averaged away.

   
Problem class Knowledge synthesis
Subject areas Immunology and Microbiology, Translational Medicine, Drug Repurposing and Discovery
Evidence level Proposed
Complexity One skill or MCP
Availability Institutional access
Compute Laptop

Problem

You have a tumor molecular profile — a TMB value from a panel or exome, MMR/MSI status, a PD-L1 IHC score, HLA genotype, and ideally bulk tumor RNA — and someone needs to know whether checkpoint blockade is worth pursuing, for a molecular tumor board, a trial-eligibility screen, or a retrospective cohort stratification. The single-biomarker answer is the trap. TMB-high and MSI-high both carry tumor-agnostic approvals, yet response within those groups is strikingly heterogeneous, and each biomarker fails differently: TMB reports neoantigen supply and says nothing about whether T cells are in the tumor, while a T-cell-inflamed expression signature reports the reverse and is nearly uncorrelated with TMB. Resistance genotypes (B2M loss, JAK1/JAK2 loss-of-function) can void a favorable score entirely.

What “solved” looks like is not a single number. It is a per-sample record that carries each biomarker axis separately with its cancer-type-specific threshold, names the resistance features found, grades how strong the clinical evidence is for that biomarker in this tumor type, and flags disagreement between axes as disagreement — plus enough provenance that the call can be re-derived when the underlying knowledgebases have moved on.

  1. Install the ToolUniverse MCP server, then the Immunotherapy Response Prediction skill. The skill is a reasoning layer over ToolUniverse tool calls, so the server must be registered first or every lookup fails. Follow the install paths on those pages. The skill ships disable-model-invocation: true, so you must ask for it by name.

  2. Assemble the inputs as a table, not as prose. One row per sample, with the columns the skill actually consumes: cancer type (as a specific diagnosis, not “carcinoma”), TMB in mutations/Mb with the assay it came from, MMR/MSI status and method, PD-L1 score with the antibody clone and scoring convention (TPS vs CPS vs IC — they are not interchangeable), HLA class I genotype, and any somatic variants in B2M, JAK1, JAK2, STK11, KEAP1, PTEN, CD274. Commit this as ici_inputs.csv. Missing fields are fine and should stay blank — the skill grades evidence per axis, and a blank is more honest than an imputed value.

  3. Capture the workflow as a reusable command, not a one-off chat. Write .claude/commands/ici-response.md so the same phrasing runs over every sample:

    Use the tooluniverse-immunotherapy-response-prediction skill on each
row of ici_inputs.csv. For every sample:
  1. Classify TMB against the threshold for that specific cancer
     type, and record the threshold used and its source.
  2. Assess MMR/MSI and PD-L1 with the scoring convention named
     in the input; do not convert between TPS/CPS/IC.
  3. Evaluate the listed mutations as sensitizing or resistance
     features, citing the CIViC/FDA evidence behind each call.
  4. Profile the immune microenvironment from the expression
     table if one is supplied; otherwise mark that axis absent.
  5. Emit the 0-100 integrated score AND the per-axis component
     scores with their T1-T4 evidence grades - never the
     integrated score alone.
  6. Flag any sample where two axes disagree in direction.
Append one row per sample to ici_biomarkers.csv with columns for
each axis, its evidence grade, and the reason string. Write the
narrative report to reports/<sample_id>.md, citing only values
present in ici_biomarkers.csv.

  4. Record provenance, because none of this is byte-reproducible. The skill queries live knowledgebases (CIViC, Open Targets, FDA labels, Human Protein Atlas, ClinicalTrials.gov, PubMed) whose contents change weekly. Have the run emit provenance.json capturing the ToolUniverse version, the skill’s commit SHA, the UTC date of the query, the release date or accession of each knowledgebase consulted, the sha256 of ici_inputs.csv, and the model id. Commit ici_inputs.csv, the command file, ici_biomarkers.csv, reports/, and provenance.json together.

  5. Read the disagreements first. The rows worth a human’s time are the ones where axes conflict — TMB-high with a cold microenvironment, or PD-L1-positive with B2M loss. Treat the integrated 0–100 score as an index into the per-axis table, not as the answer. This output is a research aid; it is not a clinical decision tool and must not be used as one without independent review.

Why this assembly

Rung 2, and it stops there. Plain Claude Code (rung 1) can look up a TMB threshold, but the work here is not lookup — it is applying cancer-type-specific thresholds, keeping the axes separate, and attaching evidence grades from the right knowledgebase to each call. That’s a long, ordered protocol with many chances to quietly substitute a pan-cancer threshold for a tumor-specific one, which is exactly the error the skill’s eleven-phase structure exists to prevent. One skill plus its MCP server covers every lookup the protocol needs — Open Targets, CIViC, FDA pharmacogenomics, HPA, IEDB, Enrichr, trials, PubMed — so a multi-tool harness would only re-implement what the skill already routes.

Availability

Institutional access. The software is free: ToolUniverse and its skills are Apache-2.0, and every knowledgebase behind them is public. The gate is the data. Patient TMB, MSI, PD-L1, HLA, and expression values are clinical molecular-profiling results and sit inside your IRB, consent, and data-governance regime — and the MCP server sends query terms to external APIs, so confirm with your governance office what may leave the institution before running this on identifiable samples. De-identified cohort-level runs (TCGA, IMvigor210) have no such constraint and are Fully open in practice.

Compute requirements

Laptop. Every step is an API call; there is no local model and no heavy computation. Wall-clock is dominated by the number of knowledgebase round-trips the skill’s eleven phases make — expect a few minutes per sample, so a 200-sample cohort is an overnight batch rather than an interactive session, and you should rate-limit and checkpoint ici_biomarkers.csv incrementally so a mid-run failure doesn’t cost the whole cohort.

Evidence

Proposed. No documented attempt of this exact assembly — the ToolUniverse skill driving a checkpoint-response assessment — is known, and the skill’s integrated 0–100 score has not been benchmarked against clinical outcomes by anyone. Treat the score as a structured summary of published biomarker logic, not as a validated predictor.

The individual axes it composes are well validated. The T-cell-inflamed gene expression profile was developed across 220 pembrolizumab-treated patients in 9 cancers and independently confirmed against PD-L1 IHC in 96 head-and-neck patients (Ayers et al., JCI 2017). The central design choice of this recipe — keep the axes separate — comes from the pan-tumor analysis of >300 samples across 22 tumor types and four KEYNOTE trials showing TMB and the T-cell-inflamed GEP are independently predictive and only weakly correlated, capturing distinct neoantigenicity and T-cell-activation biology, with joint stratification revealing resistance patterns neither axis sees alone (Cristescu et al., Science 2018).

Two findings justify the caution built into steps 2 and 5. In 5,621 metastatic breast cancers, TMB-high alone did not correlate with immune infiltrate or immune-response gene signatures; only concurrent PD-L1-positivity or dMMR/MSI-high identified T-cell-inflamed tumors, and B2M mutation and CD274 amplification carried additional signal (Sammons et al., Front. Oncol. 2023). Within dMMR/MSI-high tumors — a group with tumor-agnostic approval — response remains heterogeneous, and a 20-gene immune signature separated true responders at 55.6% vs 32.8% response in IMvigor210 (Kim et al., Cancers 2025). A single-biomarker call, and an integrated score that hides which axis moved it, both fail on exactly these cases.

Alternatives considered

  • Annotate tumor somatic variants with clinical actionability evidence (rung 2, narrower). Reach for that when the question is “what therapy does this variant match?” across all drug classes. It reads the same somatic VCF but answers a targeted-therapy question; it does not assemble the immune-axis biomarkers.
  • Compute the signatures yourself from the expression matrix (rung 1). If you already have a bulk RNA matrix and want the T-cell-inflamed GEP or an MCP-counter deconvolution as numbers for a cohort analysis, score them directly with the functional enrichment or TF/pathway activity paths. That is cheaper and fully reproducible. This recipe is for the case where you also need the clinical-evidence grading and resistance-genotype review attached.
  • A trained multimodal ICI-response model. Nothing Claude-installable exists today; if you need a calibrated probability rather than a graded assessment, you are building it yourself from cohort data.

See also

Sources

Tried this recipe?

Share feedback — what worked, what didn’t, what you’d change. The form opens with this recipe pre-selected and a link back to this page.