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

arXiv:2609.05496 · q-bio.MN, cs.MA · Submitted 2026-08-27 · Read on arXiv

Listen

Transcript

Introduction to the show: ident: Genomics Radio. Generated commentary on the latest computational biology and genomics papers.

Ines: We're looking at the paper: "Orchestra: Corroboration-Based Regulatory Candidate Discovery via Composed Bioinformatics MCP Agents".

Marcus: The gist The central architectural claim is that requiring RegNetAgents’ network-topology evidence and CASCADE’s independent experimental evidence to agree on a candidate regulator yields a more trustworthy candidate than either…

Ines: First, who's behind it and why it matters.

Paper summary: Marcus: So we just talked about the core idea of this paper, which is that combining RegNetAgents and CASCADE gives you better results than either system alone. Now let's talk about how they set up this experiment and what they actually tested on it.

Ines: They start by detailing their candidate selection mechanism where every experiment selects the ten regulators with the highest ARACNe mutual information edge weight, which is RegNetAgents’ own per-edge network-confidence signal.

Yuki: That selection method is important because the tumor acquired pool RegNetAgents compares returns a range from eight to eighty-four candidates per query, so that ten candidate cap drops most of it in almost every query.

Marcus: Right, so they sort the pool by descending weight before applying that cap. If the map isn't there, they fall back to the unranked order which didn't happen in any experiment reported here.

Ines: Then they assess corroboration by summing up positive sources, where four lightweight tools run in under a second per query with all four sources checked, giving a more direct measurement of the underlying evidence than keyword-matching CASCADE’s generated narrative text.

Yuki: They assess corroborated versus uncorroborated OncoKB enrichment with Fisher’s exact test at three thresholds: greater than or equal to one source, greater than or equal to two sources, and greater than or equal to three sources.

Marcus: And the results show that at one source, neither focal genes nor negative controls reach significance, but the pattern emerges cleanly at two sources where focal-gene significance appears where it was absent at one source.

Ines: They checked robustness on this finding with a permutation test and an exact reproduction crosscheck against RegNetAgents’ own published results.

Yuki: And they validated the pattern in a third cancer type, STAD, showing an odds ratio of five point eight two and a Benjamini–Hochberg adjusted p-value of zero point zero zero four seven there.

Marcus: So, when you put all that together from this paper on Orchestra: Corroboration-Based Regulatory Candidate Discovery via Composed Bioinformatics MCP Agents, the big finding is that the system requires at least two sources to predict driver status reliably.

Conclusion: Ines: Looking at the title and authors of this paper, Orchestra: Corroboration-Based Regulatory Candidate Discovery via Composed Bioinformatics MCP Agents by Jose A. Bird, it really highlights how important it is for AI systems to be built in a way that they have multiple independent ways of checking their work.

Marcus: It’s about moving beyond just having one strong predictor and instead demanding agreement across different methodologies, which is exactly what this paper shows happens in practice with these two agents.

Yuki: The implication for the wider field is that when we are looking at complex biological data like regulatory networks, we can't rely on a single inference path to draw conclusions about driver genes.

Ines: So, for anyone working in this area, it suggests that building systems where independent agents check each other’s work is a way to build more trustworthy discovery tools.

Marcus: It moves the focus from just finding the strongest signal within one system to validating that signal using different types of evidence.

Yuki: And as we look forward, this work shows that testing these multi-agent compositions isn't just theoretical; it’s a practical way to test if complex biological hypotheses hold up across different data modalities.

Bird AI Solutions

q-bio.MN, cs.MA

Submitted: 2026-08-27

Updated: 2026-10-16

Comments: 20 pages, 3 figures. v4: driver reference set changed from OncoKB to IntOGen; data re-collected on RegNetAgents v1.3.0 / CASCADE v1.5.0 with updated panels; all results, tables and figures updated

Code: https://github.com/jab57/Orchestra

License: http://creativecommons.org/licenses/by/4.0/

Importance score: 80/100

The gist: The gist The central architectural claim is that requiring RegNetAgents’ network-topology evidence and CASCADE’s independent experimental evidence to agree on a candidate regulator yields a more

Key concepts

RegNetAgents: This agent infers the regulatory network topology within biological data. It analyzes how genes interact within a system to predict potential regulatory relationships, essentially mapping out the connections between different parts of a gene regulation network.
CASCADE: This agent provides independent functional evidence for candidate regulators. It gathers data from various experimental sources, such as knockdown experiments and transcription factor confidence scores, offering direct proof about a regulator's biological activity.
Mutual Information (MI) Edge Weight: This is a metric used by RegNetAgents to select the top candidates. It measures how much information one gene provides about another within the network structure. Candidates with higher MI edge weights are considered more strongly connected and thus prioritized.
Corroboration Scoring: This process assesses a candidate's trustworthiness by checking agreement across multiple independent evidence sources (up to four). A high corroboration count, especially when at least two sources agree, indicates a much stronger biological signal than any single source alone.

Summary

The gist The central architectural claim is that requiring RegNetAgents’ network-topology evidence and CASCADE’s independent experimental evidence to agree on a candidate regulator yields a more trustworthy candidate than either system alone.

Orchestra's Architecture and Composition

Orchestra composes two independently-built bioinformatics MCP servers – RegNetAgents [Bird, 2026c], which infers regulatory network topology, and CASCADE [Bird, 2026a], which provides independent experimental evidence (LINCS knockdown, DepMap essentiality, superenhancer status, DoRothEA transcription-factor confidence) – into a single multi-agent LangGraph workflow [LangChain, Inc., 2024] exposed via the Model Context Protocol [Anthropic, 2024]. Orchestra’s central architectural claim is that requiring RegNetAgents’ network-topology evidence and CASCADE’s independent experimental evidence to agree on a candidate regulator yields a more trustworthy candidate than either system alone. This claim has not previously been tested directly: RegNetAgents’ own validation asked only whether its candidate lists beat chance, never whether cross-system corroboration adds precision beyond either system individually.

Candidate Selection Mechanism

The paper details a principled candidate-selection mechanism where every experiment selects the ten regulators with the highest ARACNe mutual-information (MI) edge weight, RegNetAgents’ own per-edge network-confidence signal, rather than an arbitrary subset. This selection is crucial because Candidate selection matters to this test, so every experiment here selects the ten regulators with the highest ARACNe mutual-information (MI) edge weight, RegNetAgents’ own per-edge network-confidence signal The tumor-acquired pool RegNetAgents’ compare network contexts returns ranges from 8 to 84 candidates per query, so the ten-candidate cap drops most of it in almost every query (all STAD queries, 88% of BRCA/COAD queries) When rank tumor acquired=True, Orchestra sorts the pool by descending weight before applying the cap; if the map is absent it falls back to the unranked order, which did not occur in any experiment reported here.

Corroboration Scoring and Significance

A candidate’s corroboration count is the sum of positive sources (0–4), where four lightweight, structured-field tools run in well under a second per query with all four sources checked, and give a more direct measurement of the underlying evidence than keyword-matching CASCADE’s generated narrative text. The analysis assesses corroborated-vs-uncorroborated OncoKB enrichment with Fisher’s exact test at three thresholds (≥ 1, ≥ 2, ≥ 3 of 4 sources), and the results show that At ≥ 1, neither focal genes (p = 0.1281) nor negative controls (p = 0.0658) reach significance. The pattern that matters emerges cleanly at ≥ 2, where focal-gene significance appears where it was absent at ≥ 1 (p = 0.0028, BH-adjusted p = 0.0166).

Robustness and Replication

The study employs four robustness checks on the core ≥ 2 finding, including "deduplication against pseudo-replication, a permutation test independent of the parametric test’s assumptions, an exact reproduction crosscheck against RegNetAgents’ own published results, and the independent ground-truth check described above". The pattern replicates and strengthens in a third cancer type, STAD (OR = 5.82, Benjamini–Hochberg-adjusted p = 0.0047), while negative controls again fall short of significance. Furthermore, the results are validated against a second, independently curated ground truth sharing no curating organization with OncoKB or DoRothEA (Sanger COSMIC Cancer Gene Census) in both panels.

Value of Corroboration Beyond Single Predictors

The paper tests whether corroboration adds value beyond MI edge weight itself, finding that "Corroboration adds significant value beyond MI edge weight in both panels – this is the clearest evidence in this paper for Orchestra’s central architectural claim: cross-system corroboration captures real signal that the single strongest predictor available, from either system, does not. This finding is further supported by a stratified cross-check in STAD where the effect is large: of the 74 candidates outside their query’s MI-top-3, the 11 corroborated ones reach OncoKB at 72.7% versus 15.9% for the other 63 (OR = 14.13)".

Limitations and Nuances

The limitations acknowledge that DoRothEA appears to account for it alone in BRCA/COAD but not in STAD, and we do not resolve this. The paper concludes that That corroboration adds value beyond DoRothEA alone holds only in STAD. The validated claim is scoped to BRCA, COAD, and STAD – three cancer types, two independently constructed panels.

Data and Code Availability

Orchestra is available at https://github.com/jab57/Orchestra under the MIT license. Reproducing any result in this paper requires both child MCP servers Orchestra composes (Section 2.1): RegNetAgents [Bird, 2026c], available at https://github.com/jab57/RegNetAgents (v1.2.5 or later specifically for the single-call MI-weight mechanism of Section 2.2; on earlier versions Orchestra falls back to unranked candidate selection rather than failing), and CASCADE [Bird, 2026a], available at https://github.com/jab57/CASCADE, which supplies the four corroboration evidence sources of Section 2.3.

AI Usage Disclosure

Development of Orchestra – the MCP server, the composition workflow, and the corroborationscoring feature validated here – was assisted by Claude Code (Anthropic), an AI coding tool, under direct human supervision. All code was tested (374 unit tests, zero regressions) and all statistical results were independently verified via multiple methods (Fisher’s exact test, permutation testing, deduplication analysis, ground-truth cross-checking) before being reported.

Acknowledgements

This work builds directly on RegNetAgents [Bird, 2026c] and CASCADE [Bird, 2026a], and reuses RegNetAgents’ own published focal-gene and negative-control panels for direct comparability with their reported results. We acknowledge the GREmLN development team at the Chan Zuckerberg Initiative AI for the pre-trained gene embeddings and population-averaged regulatory networks underlying RegNetAgents; the TCGA Research Network and the aracne.networks compilation [Lim and Califano, 2018] for the tumor-state ARACNe networks; the OncoKB team for the curated cancer-gene database used as the primary ground truth; and the COSMIC Cancer Gene Census team at the Wellcome Sanger Institute [Sondka et al., 2018] for the independent ground truth used in Section 3.8.

Funding

No funding was received for this work.

Competing Interests

The author declares no competing interests.

--- Page 1 ---

Orchestra's Architecture and Composition

Improvements for AI systems

Improved candidate selection mechanism: Implement a principled candidate-selection mechanism by ordering regulators by ARACNe MI edge weight before the 10-candidate cap, resolving the limitation of accepting an arbitrary alphabetical output order.
Enhanced corroboration threshold logic: Shift from a simple presence check to requiring genuine multi-source agreement, as demonstrated by testing that requiring genuine multi-source agreement (≥ 2) separates focal genes from negative controls in both panels, before and after multiple-testing correction.
Value assessment of cross-system corroboration: Systematically test whether the cross-system check adds value beyond it specifically by using a logistic regression likelihood-ratio test, confirming that corroboration adds significant value in both tested panels (BRCA/COAD p = 0.0234; STAD p = 0.0001).
Robustness and validation framework: Integrate four specific robustness checks into the pipeline: "deduplication against pseudo-replication, a permutation test independent of the parametric test’s assumptions, an exact reproduction crosscheck against RegNetAgents’ own published results, and the independent ground-truth check described above."
Dynamic threshold justification: Employ a principled threshold setting by deriving it from what corroboration means—the corroboration count measures agreement among CASCADE’s four independent sources, and agreement, by definition, requires at least two of them to be positive—rather than choosing a value after seeing which one worked.
Advanced predictive modeling: Develop a model that assesses whether corroboration adds value beyond the strongest single predictor by using a regression test that compares corroboration count (0–4) with MI-top-3, showing that for STAD, this measure is significantly significant while in BRCA/COAD it is not.

Abstract

Orchestra composes two independently built bioinformatics MCP servers -- RegNetAgents, which infers regulatory network topology, and CASCADE, which provides independent functional evidence (LINCS knockdown, DepMap essentiality, super-enhancer status, DoRothEA transcription-factor confidence) -- into one multi-agent LangGraph workflow. Its central claim is that requiring the two systems to agree on a candidate regulator yields a more trustworthy candidate than either alone. We test this on the TCGA tumor-acquired regulator tier, selecting each query's ten regulators by ARACNe mutual-information edge weight. Across BRCA/COAD focal genes and matched negative controls, requiring at least 2 of 4 evidence sources predicts known-driver status (IntOGen) among focal genes (odds ratio 3.15, Benjamini-Hochberg-adjusted p=0.016; 2.92 after deduplication) but shows no association among negative controls (p=0.70); a single source does not in BRCA/COAD. In a second panel (STAD) the pattern reproduces and is stronger (OR 6.71, adjusted p=0.0047), again with no association in controls. Corroboration adds value beyond MI edge weight in both panels (BRCA/COAD p=0.039; STAD p=0.0001), though modestly in BRCA/COAD. Deduplication, a permutation test and a reproduction of RegNetAgents' published results agree; a second, separately derived driver list (Vogelstein et al. 2013) is directionally consistent but too sparse to reach significance. All experiments invoke Orchestra's real agentic entry point.