Intratumoral CpG · Melanoma · ORR estimation

Melanoma CpG Response Estimator (CpG ORR Analog)

Enter a product's parameters. The tool matches it against the real trial arms and estimates ORR by weighted analogy — not a fitted model. Correlations, weights and significance all recompute live from whichever arms are active.

!Descriptive, hypothesis-generating only.
The reference set is a handful of arms across a few molecules — statistically it cannot support a real estimator. This is an analog matcher: it formalizes "your product looks most like arm X, expect roughly what X got." The fragility meter shows exactly how much one arm can swing the answer. Read every estimate as a range anchored to named comparators, never a point fact.

This is an analog matcher for objective response rate (ORR) in intratumoral CpG melanoma trials. You give it a product's parameters; it finds the most similar real trial arms and blends their reported ORRs. It is deliberately not a fitted statistical model — with this little data, a fitted equation would invent precision that isn't there.

01 What it actually does

Think of it as formalizing the sentence a domain expert says when they read the trial table: "your product looks most like arm X, so expect roughly what X got." Every estimate is a range anchored to named real arms — never a standalone number. The tool always shows you which arms it leaned on, how much each factor mattered, and how fragile the answer is.

02 The evidence base

A PubMed + ClinicalTrials.gov sweep for every intratumoral CpG / TLR9-agonist arm in melanoma that reported a RECIST ORR. That universe is small and now essentially complete:

14 intratumoral arms 4 molecules 13 report RECIST ORR every arm → PMID + NCT

The 4 molecules — nelitolimod (SD-101), tilsotolimod, vidutolimod (CMP-001), and cavrotolimod — are the complete IT-CpG-melanoma-with-RECIST-ORR landscape. Every arm's source is in the Arms & sources tab; you can toggle any arm on/off and the whole tool recomputes.

03 How the estimate is computed

  1. Normalize each factor to its observed 0–1 range across the active arms.
  2. Weight each factor by |r with ORR| × reliability (how many arms back it), computed live from the data. So the single dominant driver — PD-1-naïve status (r ≈ 0.87) — pulls hardest, while weak or thin factors barely register.
  3. Distance to each arm = weighted root-mean-square of the per-factor gaps:
    D = √( Σ wᶠ·(nᶠ − aᶠ)² ⁄ Σ wᶠ )
  4. Similarity via a Gaussian kernel: s = e^(−(D ⁄ 0.34)²).
  5. Estimate = the similarity-weighted mean of the arms' ORR — closest arms count most.
  6. Band = the similarity-weighted spread, widened for the small-n floor.

04 What keeps it honest

Fragility meter (leave-one-out): drops each arm in turn and re-estimates, so you see the full swing one arm can cause. Population lock: restrict to naïve-only or refractory-only arms, since PD-1 status is the one real driver. Raw anchors: the actual ORRs of your two closest arms, shown with no smoothing. Guards: zero dose / concentration / volume / lesions → "nothing administered, ORR ≈ 0"; out-of-range inputs get an extrapolating flag (capped by default, or uncapped by choice). Exclusions: ΔORR is left out (it's derived from ORR — circular); neoadjuvant vidutolimod reports pathologic MPR, so it can match as an analog but never anchors a RECIST number.

05 The hard limitation — read this

14 arms, 4 molecules. This cannot statistically support an estimator. Correlations at this sample size rarely reach significance, and a single arm can swing any coefficient — which is exactly what the fragility meter quantifies. Treat every output as hypothesis-generating, a structured way to reason by analogy from real comparators — not a response estimate. It is not medical or investment advice, and nothing here is peer-reviewed.

06 The tabs

EstimateEnter parameters → weighted ORR estimate, closest analogs, fragility, sensitivity sweep, and a rationale export.
FactorsWhat each input means and how strongly it tracks ORR — with a live breakdown of your product vs. its closest arm.
Arms & sourcesAll 14 arms with PMID + NCT; toggle any on/off, or add your own.
Source dataThe full arm-level data block behind everything.
CorrelationsA live Pearson r / valid-n / significance heatmap over the active arms.

The definitive landscape answer

PubMed + ClinicalTrials.gov sweep

The search was exhaustive. Bottom line: your 4 molecules are essentially the complete universe of IT-CpG-in-melanoma-with-a-RECIST-ORR. Specifically:

  • One 5th molecule exists — agatolimod / PF-3512676 (a class-B CpG) was injected intralesionally in melanoma, but reported only local-lesion regression (1 CR of 5), no RECIST ORR → can't anchor a number. It's left out; it could be added as a reference-only row (like the neoadjuvant arm) if you want it visible.
  • Everything else (lefitolimod, IMO-2055, DV281, etc.) was systemic and/or non-melanoma → out of scope by the IT-only rule.

14 intratumoral arms · 4 molecules · 13 report RECIST ORR — see the Arms & sources tab for every arm and its citation.

Product parameters

i

Estimated ORR

i
weighted analog estimate
–%
range —
—
ORR axis · reference arms plotted, estimated band shaded

Nearest analogs

similarity → weight on estimate

Sensitivity sweep

hold all else, vary one factor
factor

Compare & export

snapshots of configs

Why — factor weights driving this match

weight = |r with ORR| × reliability (live)
Factorr with ORRvalid nweightyour value → nearest arm gap
How the estimate is computed

Each factor is normalized to its observed 0–1 range across active arms. Product-to-arm distance is a weighted RMS over all factors both sides report: D = √( Σ wᶠ·(nᶠ−aᶠ)² ⁄ Σ wᶠ ), with wᶠ = |r(factor, ORR)| × reliability computed live from the active arms — so adding arms genuinely re-weights the model. Similarity is a Gaussian kernel s = e^(−(D∕0.34)²); the estimate is the similarity-weighted mean of the ORR-reporting arms, and the band is the similarity-weighted spread widened for the small-n floor.

Fragility meter: leave-one-out — drops each ORR-reporting arm in turn and recomputes, so you see the full swing one arm can cause. Guards: zero dose/conc/volume → "no active drug, ORR ≈ 0"; ΔORR and Median DoR excluded; neoadjuvant MPR never anchors a RECIST number.

How each point is used — r, n and weight update live as you toggle arms. Distance to every arm is a weighted average of the per-factor gaps below. A factor's pull = |r with ORR| × reliability. Sign of r is the direction it correlates with response (not proof of cause at small n). The prose caveats describe the original 6-arm baseline; the numbers on each card are live. Cards ordered by weight.

Live breakdown — your product vs. closest analog

—
Factorweightyour valuearm valuegap (0–1)share of distance
The reference set — every arm is intratumoral (IT alone or IT+SC); toggle any on/off and the whole tool recomputes. The original 6 arms plus 8 newly-researched arms (each sourced to a PMID/DOI below). A PubMed/ClinicalTrials.gov sweep confirmed these 4 molecules are essentially the complete IT-CpG-melanoma-with-RECIST-ORR landscape; purely-systemic CpG trials (e.g. CPG-7909 SC) were excluded by route. new arms were vetted from the literature; blanks are genuine (undisclosed), never zeros. Add your own arm at the bottom. Nothing here is peer-reviewed advice — verify each source before relying on it.

Add your own arm

enters the model immediately
Full arm-level data block (active arms). Every number the estimator uses. Blanks are genuine — cavrotolimod's conc/volume undisclosed; neoadjuvant vidutolimod reported pathologic MPR, so its RECIST ORR/ΔORR are blank; several newly-added arms disclosed only ORR + N.
Arm-level data block
CpG class key (immune profile)
ClassProfileExample
A (D-type)Strong type-I interferon (IFN-α) from pDCs; weak B-cell activationVidutolimod (CMP-001)
B (K-type)Strong B-cell / NF-κB activation; weaker interferonAgatolimod (CpG-7909)
CBoth profiles — strong interferon and strong B-cell activationSD-101, tilsotolimod
Unclassed (SNA)Spherical nucleic acid; class not stated by sponsorCavrotolimod (AST-008)

r values recompute live in the Correlations tab; the source workbook's Figures 2 & 5 correspond to the original 6-arm subset.

Detailed trial registry — 16 arms (from "Melanoma IT (local)")
The full verified registry behind the analysis set — every arm with route-purity verification, carrier/depot mechanism, exact dosing, sponsor, status, clinical outcome, verdict, grade and source. Click any arm to expand all fields. ⚠ marks arms with a subcutaneous / mixed-route component.
Pearson r, computed live from the active arms. Pairwise-complete (only arms reporting both factors count). Hover a cell for r, valid n and significance; click it for the scatterplot. Blue = positive, red = negative, depth = |r|. At small n almost nothing clears significance — hypothesis-generating.
i zoom 100%
fill = r (−1…+1) corner = valid n (greener→more arms) ring + bold = p<0.05 n/a

Read-out — original 6-arm baseline

  • PD-1-naïve status is the strongest ORR correlate (r +0.90) — the only one both strong and mechanistically sensible.
  • Added benefit shrinks in bigger, later trials: ΔORR vs Patients N −0.96, ΔORR vs Phase −0.82 — "efficacy evaporates in Phase 3".
  • Patients N vs raw ORR −0.64 is an artifact — big trials were the refractory Phase 3 ones. N is a reliability weight, not a driver.
  • Raw ORR rises with doses (+0.80) / weeks (+0.71), falls with volume (−0.86) / conc (−0.79) — mostly population proxies.
  • These figures are the 6-arm baseline; the live map above changes as you add or toggle arms.

Method: r = cov(x,y) ⁄ (σₓσᵧ) over pairwise-complete arms; significance via two-tailed t-test, t = r√((n−2)∕(1−r²)). Constant or n<3 pairs shown as n/a.