Orchestra: Corroboration-Based Regulatory Candidate Discovery via Composed Bioinformatics MCP Agents

1 Introduction

• The paper directly tests whether independent agents’ agreement is more trustworthy, a claim rarely tested for composed multi-agent systems.RegNetAgents’ earlier validation tested candidate lists against chance, not whether agreement with CASCADE’s experimental evidence provides a meaningful filter.Source: §1 · p. 2
• The authors identify shared authorship and the lack of independent peer review as a threat to the independence claim, addressed by negative controls and an independent-ground-truth check.Those checks test the threat rather than assuming it away.Source: §1 · p. 2
• This is a validation report for Orchestra’s cross-system corroboration feature, not a biological discovery paper reporting experimentally validated regulator–gene relationships.Every experiment uses Orchestra’s real agentic entry point.Source: §1 · p. 3, §1.1 · p. 3

1.1 Contribution

• At least two CASCADE sources distinguish focal genes from negative controls, and the pattern strengthens in STAD (deduplicated OR = 6.03 vs. 2.59 for BRCA/COAD).Corroboration also adds value beyond MI edge weight in both panels (BRCA/COAD p = 0.0234; STAD p = 0.0001).Source: §1.1 · p. 3
• Tumor-acquired regulators are ranked by ARACNe mutual-information edge weight before the 10-candidate cap, replacing arbitrary alphabetical selection.The contribution addresses an open limitation in Orchestra’s original candidate selection.Source: §1.1 · p. 3
• Four robustness checks address pseudo-replication, reproduce RegNetAgents’ published results, use permutation testing, and test an independently curated cancer-gene list.The independent ground truth is the Sanger COSMIC Cancer Gene Census.Source: §1.1 · p. 3

2.1 Orchestra architecture and the validated code path

• The sequential comparison path first gets three regulator sets from RegNetAgents, then forwards selected sets to CASCADE; tumor-acquired regulators were newly added as an opt-in validation path.Experiments enabled validation and ranking while preserving default behavior for existing callers, and invoked the real workflow entry point.Source: §2.1 · p. 4

2.2 Candidate selection: ranking tumor-acquired regulators by ARACNe MI edge weight

• Because the 10-candidate cap truncates large tumor-acquired pools, Orchestra sorts candidates by descending ARACNe mutual-information edge weight before applying it.The returned pool is alphabetically ordered, an order that carries no confidence or importance information.Source: §2.2 · p. 4, Figure 1 · p. 5

2.3 Corroboration scoring: four independent evidence sources

• Each candidate receives a 0–4 corroboration count from LINCS knockdown, DepMap essentiality, super-enhancer status, and DoRothEA transcription-factor confidence.A source scores positive when its structured check meets that source’s criterion.Source: §2.3 · pp. 5–6

2.4 Focal gene panel and negative controls

• The BRCA/COAD analysis reuses RegNetAgents’ 23 focal gene–cancer-type queries and 20 pooled negative-control queries, analyzing only tumor-acquired regulators.The controls comprise housekeeping and tumor-expressed non-driver genes and test whether corroboration tracks generally well-studied candidate genes.Source: §2.4 · p. 6

2.5 Ground truth: OncoKB

• OncoKB cancer-gene labels were fetched independently from its public API; the Sanger COSMIC Cancer Gene Census supplies a second ground truth for robustness checks.OncoKB’s literature-curation bias is discussed as a limitation.Source: §2.5 · p. 6

2.6 Statistical framework

• Fisher’s exact tests compare corroborated and uncorroborated candidates at three thresholds, with Benjamini–Hochberg correction applied per panel across testable primary comparisons.The analysis also deduplicates focal candidates to check pseudo-replication and uses 1,000-draw permutations on the deduplicated set; other tests re-check the ≥2 finding individually.Source: §2.6 · pp. 6–7

2.7 A second, independent panel: STAD

• A separately constructed STAD panel began with 15 focal genes; four lacking GREmLN coverage were dropped before experiments, leaving 11 focal genes plus 10 negative controls.The remaining queries used STAD’s own ARACNe network and selected up to 10 candidates by MI edge weight.Source: §2.7 · p. 7

3.1 A single evidence source is not diagnostic; multi-source agreement is

• At ≥2/4, focal-gene candidates show OncoKB enrichment (p = 0.0028; BH-adjusted p = 0.0166), while negative controls do not (p = 0.0721).Table 1 reports the threshold comparisons for the MI-weight-ranked BRCA/COAD panel.Source: Table 1 · p. 8, §3.1 · p. 8, Figure 2 · p. 9
• At ≥1/4, neither focal genes (p = 0.1281) nor negative controls (p = 0.0658) reach significance; at ≥3/4, too few candidates make either group informative.Only 3 focal-gene and 2 negative-control candidates reach ≥3/4.Source: §3.1 · p. 8

3.2 MI edge weight itself predicts candidate quality, and shows no detected association with corroboration count

• Across 339 focal candidates, the top three MI-ranked candidates show 35.3% OncoKB membership versus 13.9% for the bottom three (p = 0.0003), with rank correlation ρ = −0.177 (p = 0.0011).Rank 1 denotes the highest MI edge weight.Source: §3.2 · p. 9
• MI weight and corroboration count show no detected association (ρ = −0.049, p = 0.3687), so the data are consistent with, but do not establish, independent signals.The scores derive from ARACNe co-expression edges and four experimental-assay checks, respectively; a null result could miss a weak correlation.Source: §3.2 · pp. 9–10

3.3 Robustness: deduplication

• Of 229 focal-gene candidate-rows, 211 are unique genes, while 16 genes recur across multiple focal-gene queries.TBC1D9 and ZFP91 occur three times each; 14 genes occur twice each.Source: §3.3 · p. 10
• After deduplication, the ≥2 association remains significant (OR = 2.59, p = 0.0091 uncorrected), versus raw-row OR = 2.89 (p = 0.0028; BH-adjusted p = 0.0166).The deduplicated result is based on one row per unique gene.Source: §3.3 · p. 10
• The modest change after removing repeated rows indicates that pseudo-replication is not driving the ≥2 finding.The deduplicated result remains significant.Source: §3.3 · p. 10

3.4 Robustness: reproducing RegNetAgents’ published results

• The 10-gene CTNNB1/BRCA tumor-acquired pool reproduced RegNetAgents’ published candidate set and OncoKB overlap exactly: 4 of 10 genes.The overlapping genes were YAP1, DDR2, IL6ST, and ARID3A.Source: §3.4 · p. 10

3.5 Robustness: permutation test

• Only 11 of 1,000 permutations matched or exceeded the 20.3-percentage-point OncoKB rate gap on 211 deduplicated genes (empirical p = 0.0110).The reshuffling held corroborated and uncorroborated group sizes fixed, independently checking the Fisher’s-exact result.Source: §3.5 · p. 10

3.6 STAD: replication in a third cancer type

• The STAD panel scored 210 candidates—110 focal and 100 controls—with no coverage-related failures.Ten tumor-acquired regulators per query were selected by descending ARACNe MI edge weight.Source: §3.6 · p. 10
• OR = 5.82, p = 0.0024 raw, for STAD focal genes at ≥2; deduplication strengthened this to OR = 6.03, p = 0.0021, while controls fell short (p = 0.0930).Figure 3 shows focal genes reaching significance in both cancer types, but not negative controls; candidate selection used descending MI edge weight.Source: §3.6 · p. 10, Figure 3 · p. 11
• The replicated ≥2 pattern is supported in BRCA, COAD, and STAD, but is not claimed to hold across all TCGA cancer types.STAD was newly constructed on a separate patient cohort.Source: §3.6 · pp. 11–12

3.7 Threshold justification

• A single positive source behaves differently across panels: it is not significant in either BRCA/COAD group, but after correction is significant for STAD focal genes only.In STAD, ≥1 is nominally significant for both focal genes and controls before correction.Source: §3.7 · p. 12
• ≥2 is the smallest count at which multi-source agreement is possible: a single positive source cannot agree with itself.The threshold is grounded in the meaning of corroboration, not chosen after observing which value worked.Source: §3.7 · p. 12
• Whether CASCADE agreement adds value beyond RegNetAgents’ topology signal is a separate cross-system question tested in Section 3.9.Source: §3.7 · p. 12

3.8 Robustness: an independent cancer-gene ground truth

• The check used the Sanger Institute’s 581-gene COSMIC Cancer Gene Census, curated separately from OncoKB and DoRothEA.Membership was retrieved via OncoKB’s public API, which republishes the Sanger flag, leaving a small residual dependency on that API.Source: §3.8 · p. 12
• Against the Census, BRCA/COAD focal genes reached OR = 3.44 (p = 0.0058; permutation empirical p = 0.0030), while negative controls did not (OR = 1.77, p = 0.2783).The observed rate gap was 17.6 percentage points.Source: §3.8 · p. 12
• In STAD, focal genes reached OR = 6.30 (p = 0.0036; permutation empirical p = 0.0050), while negative controls fell short (OR = 2.85, p = 0.1739).This matched the qualitative OncoKB pattern using a ground truth curated by an organization separate from OncoKB and DoRothEA.Source: §3.8 · p. 12

3.9 Does corroboration add value beyond the strongest single predictor?

• The primary test regresses OncoKB membership on MI-top-3 status and corroboration count; a stratified cross-check tests rank-4–10 candidates separately.The stratified test excludes candidates in their query’s MI-top-3, where MI edge weight already does the work.Source: §3.9 · p. 13
• In BRCA/COAD, the regression likelihood-ratio test was significant (p = 0.0234), while the stratified cross-check did not clear significance (permutation p = 0.0680).Among candidates outside their query’s MI-top-3, 29 corroborated candidates reached OncoKB at 31.0% versus 16.3% for the other 123 (OR = 2.32).Source: §3.9 · p. 13
• In STAD, outside the MI-top-3, corroborated candidates reached OncoKB at 72.7% versus 15.9% (OR = 14.13; Fisher p = 0.0003).The permutation test gave p < 0.001, with none of 1,000 relabelings reaching the observed gap.Source: §3.9 · p. 13
• Corroboration adds significant value beyond MI edge weight in both panels (BRCA/COAD likelihood-ratio p = 0.0234; STAD p = 0.0001).The paper identifies this as its clearest evidence for the central architectural claim.Source: §3.9 · p. 13
• Beyond DoRothEA, the other three sources add value in STAD (regression p = 0.0189) but not detectably in BRCA/COAD (p = 0.3958).The BRCA/COAD stratified cross-check was underpowered, and the authors leave open which CASCADE source drives corroboration.Source: §3.9 · p. 14
• In STAD, only six DoRothEA-negative candidates had at least two of the other sources, too few for the stratified test to resolve.The section distinguishes this narrower DoRothEA comparison from the central finding that corroboration adds value beyond MI edge weight in both panels.Source: §3.9 · pp. 13–14

4 Limitations

• RegNetAgents, CASCADE, and Orchestra share one author, and none had undergone independent peer review at the time of writing.The paper tests data and inference independence but makes no claim of developer independence.Source: §4 · p. 14
• Only LINCS directly tests whether a candidate regulates the queried focal gene; the other three CASCADE sources assess properties of the candidate alone.Those three sources can score a candidate highly without showing that it regulates the specific focal gene.Source: §4 · p. 14
• The 10-candidate cap excludes most tumor-acquired regulators in most queries, so excluded candidates are never evaluated.Raising or removing the cap was outside the paper’s scope.Source: §4 · p. 14
• The validated claim covers BRCA, COAD, and STAD, not all 14 cancer types supported by RegNetAgents and CASCADE.The paper treats each additional tested cancer type as incremental evidence, not a definitive universal result.Source: §4 · pp. 14–15
• OncoKB and DoRothEA both draw on published research attention, leaving a residual possibility of dependency despite negative controls and the Sanger CGC check.The Sanger CGC flag was retrieved through OncoKB’s API rather than COSMIC’s own portal.Source: §4 · p. 15
• A non-significant correlation test between MI edge weight and corroboration count shows no detected association, not statistical independence.The authors did not run an equivalence test or power analysis for this comparison.Source: §4 · p. 15

5 Data and Code Availability

• Orchestra is available under the MIT license, and its v1.3.3 release is archived on Zenodo.The tumor-acquired validation and ranking parameters are implemented in the workflow and MCP tool schema.Source: §5 · p. 15
• Reproduction requires RegNetAgents v1.2.5 or later and CASCADE, the two child MCP servers Orchestra composes.Earlier RegNetAgents versions fall back to unranked candidate selection rather than failing.Source: §5 · p. 15
• The repository includes scripts and full results for BRCA/COAD, STAD, and the Sanger COSMIC Cancer Gene Census cross-check.The listed scripts also test incremental value beyond MI edge weight and beyond DoRothEA.Source: §5 · pp. 15–16
• Regenerating candidate data requires configured RegNetAgents and CASCADE and rerunning collection scripts across all 64 queries.The gene-match guard reported zero mismatches across those queries; the CGC flag snapshot was retrieved through OncoKB’s public API.Source: §5 · pp. 15–16

6 AI Usage Disclosure

• Claude Code assisted with Orchestra development, experiments, and drafting; the human author reviewed and validated the AI-assisted work.The paper reports 374 unit tests with zero regressions and independent statistical checks using multiple methods.Source: §6 · p. 16

• No funding was received for this work.Source: §8 · p. 16