Across muscle diseases, the obvious insult explains surprisingly little of the weakness that follows. In dystrophin-deficient mice, physical tearing of the muscle-fiber membrane accounts for at most 21% of the strength loss, which itself reaches 70% — the rest flows through downstream chemistry that no one has mapped Mechanical factors tune the sensitivity of mdx muscle to eccentric strength loss and its protection by antioxidant and calcium modulators. In cystic fibrosis exacerbations, quadriceps strength drops 16.7% on admission with no correlation to inflammation markers or to how little patients move, and it recovers within a month Effect of acute exacerbations on skeletal muscle strength and physical activity in cystic fibrosis. The missing causation runs through intermediaries — calcium leaks, oxidative stress, microvascular injury — that are unmeasured in every relevant dataset.
Expert judgment about these chains exists, but it is locked inside qualitative consensus reports. The Delphi method — a structured way to pool expert opinion through anonymous ratings over several rounds, with the group's aggregate scores fed back between rounds — reliably turns opinion into checklists and priority lists, but its output cannot drive a computational model. Meanwhile, algorithms that build causal networks from data need exactly the kind of edge-level judgments experts hold, but they assume nothing important is left unmeasured, which is false here by construction. This proposal argues that the useful output of a Delphi panel on muscle weakness is not its consensus but its disagreement: the per-edge distribution of expert belief, used as an uncertainty set over causal structures within which we compute bounds rather than point estimates.
Core Claim
When a causal question runs through unmeasured variables, a panel's settled consensus is the wrong summary and a panel's structured dissent is the right one. We will run a Delphi process in which physiologists — aided by large-language-model agents playing specialist personas, following the multi-agent design of An LLM-based Delphi Study to Predict GenAI Evolution — are queried not about mechanisms in general but about specific ancestral claims ("does intracellular calcium dysregulation act upstream of mitochondrial oxidative phosphorylation suppression?"). Each edge ends the process with a probability distribution over "absent," "A causes B," and "B causes A." That distribution defines a set of plausible causal networks, and over that set we bound the contribution of each downstream pathway to force loss using the differentiable search machinery of Your Assumed DAG is Wrong And Here's How To Deal With it. If the panel's dissent is informative, the resulting intervals will contain the known quantitative results from four independent muscle-weakness studies; if the panel collapses to a single consensus graph, the bounds will fail.
Research Program
The work builds three connected pieces. The first is an edge-level Delphi protocol. Classical Delphi rounds ask experts to rate statements and converge toward the median; clinical genetics shows both the power and the limits of this style — the ClinGen panel in Expert Panel Curation of 31 Genes in Relation to Limb Girdle Muscular Dystrophy produced durable gene-disease classifications, but only for monogenic causes, and its outputs are static labels rather than predictive structure. Our protocol borrows its anonymity, iteration, and statistical feedback (content-validity-style thresholds for retaining or discarding edges, as used in Feasibility, comprehensibility and acceptability of the VISION-Cog, a novel tool to assess cognitive impairment in visually impaired older adults: a cross-sectional pilot study in Singapore), but rewrites the target: every query is a directed edge between two named physiological variables, and convergence is measured rather than enforced. Expert-In-The-Loop Causal Discovery: Iterative Model Refinement Using Expert Knowledge showed that experts are reliable at orienting single edges (accuracy around two-thirds suffices) even when they cannot specify whole graphs; we extend that insight from a single oracle to a dissenting panel, and from an algorithm that assumes no hidden causes to one that expects them.
The second piece is empirical anchoring. The published quantitative findings become hard constraints rather than things to rediscover: membrane damage's share of force loss is capped near 21% Mechanical factors...; inflammatory-load and inactivity edges must be near zero and respiratory-muscle edges spared in the exacerbation setting Effect of acute exacerbations...; the fibrotic module's relation to strength must be negative Transcriptomic analysis reveals abnormal muscle repair and remodeling in survivors of critical illness with sustained weakness. These anchors are held out of the panel's view for some validation tasks and injected as constraints for others, which lets us test whether consensus plus literature reproduces known physiology.
The third piece is bound computation over the dissent-defined set. Writing $\mathcal{G}$ for the set of directed acyclic graphs — networks of causal arrows with no circular paths — whose edge assignments stay within the panel's distribution, and $\tau(G)$ for the contribution of a given pathway to force loss under graph $G$, we compute
$$\left[\min_{G\in\mathcal{G}}\hat{\tau}(G),\ \max_{G\in\mathcal{G}}\hat{\tau}(G)\right],$$
the guaranteed range of the pathway's effect given everything the panel could not agree on. This directly reuses the Lagrangian relaxation over large graph spaces of Your Assumed DAG is Wrong..., which bounded treatment effects over up to $10^7$ candidate graphs, but feeds it panel dissent instead of a single analyst's assumptions, and adds latent variables — the panel nominates unmeasured factors (cytokine microenvironments, calcium sparks, microvascular flow) as explicit nodes so that confounding is represented rather than assumed away. This is the point of departure from Causal Modeling of Adverse Pregnancy Outcomes via Adaptive LLM Proposals, whose CLARA loop also couples language-model proposals with data scoring but returns one best graph under a no-hidden-causes assumption; here the deliverable is an interval whose width is an honest measure of expert uncertainty.
Proposed Experiments
Experiment one verifies the machinery on the standard synthetic network. We run the full pipeline — persona agents, edge queries, dissent aggregation, bound optimization — on the ALARM benchmark and compare against CLARA's reported numbers: a 32% improvement in Structural Intervention Distance over one-shot language-model prompting and 55% over purely data-driven algorithms Causal Modeling of Adverse Pregnancy Outcomes via Adaptive LLM Proposals. Matching or approaching those gaps confirms the plumbing before touching physiology.
Experiment two is the decisive test on muscle. Variables and anchors come from four studies: the mdx eccentric-contraction dataset (length change, velocity, work, membrane-dye uptake, force loss, drug rescue) Mechanical factors...; the cystic fibrosis cohort (quadriceps and respiratory strength, inflammatory markers, accelerometer activity) Effect of acute exacerbations...; the intensive-care biopsy cohort with its mitochondrial and fibrotic modules Transcriptomic analysis...; and the four-wave autopsy series linking endotheliitis and enteroviral protein to myocardial injury Comparative Study of the Myocardium of Patients from Four COVID-19 Waves. The primary validation is leave-one-study-out: the panel and the anchoring constraints never see the hamster infection results, and we ask whether the DAG set built from the other studies predicts the tissue-discordance pattern — severe diaphragm atrophy (a 48% drop in fiber area) despite no detectable virus, against heavy quadriceps viral load with intact morphology and a 23-fold MuRF1 response Characterization of SARS‐CoV‐2 Entry Genes in Skeletal Muscle and Impacts of In Vitro Versus In Vivo Infection. Success means the held-out effects fall inside the computed intervals and the predicted direction of the diaphragm-versus-limb asymmetry is correct; failure means the intervals are too narrow (miscalibrated dissent) or so wide they carry no information. There is no reported result found on Lacuna for any causal-discovery or Delphi-based analysis of these muscle datasets, so this experiment establishes the first reference numbers for the domain.
Experiment three studies the panel itself. We run human-only, agent-only, and hybrid panels across five rounds and test the field's known worry — that Delphi groups drift toward confident shared error — by checking whether shrinking disagreement predicts correctness on the held-out anchors. If dissent decays without accuracy rising, forced convergence is actively harmful and the case for preserving the distribution is made empirically, not just theoretically. The agent panels follow the persona and redundancy-filtering design that held thematic consistency at 100% across twelve runs in An LLM-based Delphi Study..., with retrieval augmentation to cover its documented knowledge-cutoff gap.
Baselines
The do-nothing reference is what the field already has: undirected co-expression modules treated as mechanism lists. The intensive-care study's mitochondrial module correlates with strength at $R=0.64$ and its fibrotic module at $R=-0.65$ Transcriptomic analysis..., and these correlations are the current best empirical summary of persistent weakness. As a baseline it sets the floor: it names candidate pathways but assigns no direction, distinguishes no driver from response, and yields no interval for any intervention, so any method that produces a directional bound containing the known sign of the fibrotic relationship beats it on the questions that matter.
The strongest existing competitor is CLARA Causal Modeling of Adverse Pregnancy Outcomes via Adaptive LLM Proposals, the closest published system: it generates candidate graphs with a language model, scores them against data with a decision-tree Bayesian information criterion, and re-injects the winners, reporting a 32% Structural Intervention Distance improvement over one-shot prompting and 55% over data-driven search on ALARM, and recovery of all 31 expert-validated edges on the nuMoM2b obstetrics cohort. We will run CLARA on exactly our muscle variable sets, with the same literature corpus supplied to its language model, the same compute, and the same tuning budget, and compare on held-out anchor coverage: whether its chosen graph's effect estimates, with bootstrap intervals, contain the held-out physiological effects. Beating CLARA means strictly higher coverage of held-out anchors at comparable interval width — and we expect it to fail specifically on the discordance predictions, because its single-graph output forces a guess on questions (tissue asymmetry, latent circulating factors) where guessing is unjustified.
The third baseline is the EXPERTINLOOP algorithm with a language-model oracle Expert-In-The-Loop Causal Discovery..., which queries an expert only to orient edges and provably recovers structure when the oracle is right about two-thirds of the time, outperforming PC, Hill-Climb, and GES. We run it on the same variables and oracle budget; it competes directly on edge orientation accuracy against our panel distribution, but like CLARA it assumes no unmeasured common causes, which in this domain is the assumption most likely to be false.
The key ablation collapses our own pipeline to its consensus: replace the per-edge distribution with the panel median, feed the single majority graph to the same bound machinery, and recompute. This isolates dissent preservation as the ingredient, and the prediction from Your Assumed DAG is Wrong... — that a wrongly assumed single graph biases effect estimates badly — is that coverage on held-out anchors drops while intervals look spuriously tight.
Why not just…?
Why not just run standard causal-discovery algorithms on the biopsy and cohort data? The datasets are tiny and noisy — 32 samples from 14 patients in the intensive-care cohort — and the corpus gives direct evidence of how such methods behave there: CLARA documented purely data-driven algorithms fluctuating wildly on noisy clinical data and improved on them by 55% on ALARM Causal Modeling of Adverse Pregnancy Outcomes via Adaptive LLM Proposals. Worse, every one of these algorithms assumes no unmeasured confounders, and the confounders here (calcium, cytokines, microvascular flow) are unmeasurable in the available data. Our experiments will settle the residual question fairly, since data-driven search is included inside the CLARA comparison.
Why not just prompt a large language model once to lay out the causal pathways? Single-shot language-model reasoning is documented to conflate correlation with causation and to be sensitive to phrasing — CLARA needed an adaptive loop to gain 32% in distance over one-shot prompting — and single-round evaluation carried a positivity bias in the multi-agent Delphi study An LLM-based Delphi Study.... A one-shot prompt also returns no uncertainty distribution, so it cannot produce a bound at all; the deliverable of this proposal is precisely the thing one-shot output cannot express. The agent-only arm of Experiment three measures how much of the panel's value the personas alone capture, which is the honest version of this alternative.
Why not just convene a conventional expert panel and publish its consensus diagram? This is the real threat, because it is cheaper and the field's default. Two things argue against it and one experiment decides it. First, consensus panels converge on shared misconceptions when unconstrained by data, a failure mode the clinical-genetics curation literature manages only by enforcing semi-quantitative evidence scoring Expert Panel Curation of 31 Genes..., which we adopt. Second, a single consensus graph is exactly the object shown to bias causal estimates when the true structure differs even slightly, which is why bound coverage of 97–100% required searching over contested graph spaces rather than picking one Your Assumed DAG is Wrong.... The consensus-collapse ablation is the direct test: if the median graph covers the held-out anchors as well as the full distribution, the expensive part of this proposal is unnecessary and we will say so.
Why not just use the co-expression modules already published? The intensive-care study already found the modules and their correlations Transcriptomic analysis...; nothing here rediscovers them. But a co-expression module is symmetric: it cannot say whether mitochondrial failure drives weakness or something upstream causes both, and the in vitro-versus-in vivo inversion of the same atrophy genes — MuRF1 down 64% in culture against a 23-fold rise in the intact animal Characterization of SARS‐CoV‐2 Entry Genes... — is a live demonstration that correlational structure misleads exactly where causal ordering is contested.
Decisive Checks
The proposal stands or falls on three numbers. Leave-one-study-out coverage: the hamster tissue-discordance effects (48% diaphragm fiber-area loss, 23-fold MuRF1 induction) must fall inside intervals computed without them, or the panel's dissent carries no signal. Ablation gap: the consensus-collapse graph must show measurably worse coverage than the full distribution, or preserving disagreement is decoration. Width honesty: intervals must be narrower than the trivial "anything from zero to total loss" bound that the 21%-versus-70% gap in Mechanical factors... already licenses, or the machinery adds nothing to what one careful experiment reported. Meeting all three would give muscle pathology its first quantified, disagreement-aware causal map of the pathways that dominate weakness, and give the Delphi method a computational output it has never had.
References
- Mechanical factors tune the sensitivity of mdx muscle to eccentric strength loss and its protection by antioxidant and calcium modulators
- Effect of acute exacerbations on skeletal muscle strength and physical activity in cystic fibrosis
- An LLM-based Delphi Study to Predict GenAI Evolution
- Your Assumed DAG is Wrong And Here’s How To Deal With It
- Expert Panel Curation of 31 Genes in Relation to Limb Girdle Muscular Dystrophy
- Feasibility, comprehensibility and acceptability of the VISION-Cog, a novel tool to assess cognitive impairment in visually impaired older adults: a cross-sectional pilot study in Singapore
- Expert-In-The-Loop Causal Discovery: Iterative Model Refinement Using Expert Knowledge
- Transcriptomic analysis reveals abnormal muscle repair and remodeling in survivors of critical illness with sustained weakness
- Causal Modeling of Adverse Pregnancy Outcomes via Adaptive LLM Proposals
- Comparative Study of the Myocardium of Patients from Four COVID-19 Waves
- Characterization of SARS‐CoV‐2 Entry Genes in Skeletal Muscle and Impacts of In Vitro Versus In Vivo Infection