No Site Left Worse Off: Pareto-Certified Coalitions for Multi-Site RNA-seq Analysis

Hospitals that each hold a few dozen RNA sequencing samples cannot fit reliable expression models alone. RNA sequencing (RNA-seq) counts how many copies of each gene's transcript a sample contains. Those counts are noisy in a gene-specific way, so estimating that noise needs many samples. Current federated approaches handle this by putting every site into one shared model. Examples are privacy-preserving federated training in PPML-Omics and the federated batch correction of Fed-ComBat. But pooling mismatched cohorts can do real harm. In Bayesian multi-domain learning for cancer subtype discovery, competing transfer methods fell to about 51% accuracy when a poorly related cancer type was used as auxiliary data, while the paper's selective-sharing model reached 78.10%. A small hospital asked to join a consortium therefore faces a real risk that joining makes its own analysis worse.

We propose to let sites form their own analysis groups. They use the merge-and-split rules from wireless coalition games: groups fuse or break apart through local moves, and a move is allowed only if nobody in it loses and someone gains. Each site scores a candidate group by one number: how well the group's shared count model predicts that site's own held-out samples. The new ingredient is a statistical test on this rule. A merge happens only if every member's improvement is confirmed beyond sampling noise. The consortium can then promise, with stated confidence, that no participating hospital ends up worse than if it had worked alone. No existing federated transcriptomics method offers that guarantee.

Core Claim

The claim applies when the product of collaboration is a shared statistical model rather than a personalized predictor. Examples are a common batch correction, shared gene-wise noise estimates for differential expression, and a common subtype classifier. In that setting, sites should form disjoint groups. Each site's payoff should be its held-out Negative Binomial likelihood: a count model whose variance can exceed its mean, as RNA-seq requires. The merge rule should be a Pareto test with confidence bounds. We expect this to achieve two things. First, it should match or beat the all-sites model on average. Second, it should strictly dominate the all-sites model on the worst-off site.

The second property is what a consortium agreement needs, and it is where we expect utility-summing or similarity-based grouping to fail. Those rules happily sacrifice a small site to raise the average.

Research Program

Direction one supplies the machinery, direction two supplies the regime and the payoff. At any time the $N$ sites form a partition, meaning a set of non-overlapping groups. Each group $S$ jointly fits one count model with three parts: site-specific library-size and batch offsets, shared gene-wise dispersions (the extra-Poisson noise level of each gene), and shared subtype-specific mean expression. The fit is federated. Only per-gene sufficient statistics and gradients leave a site, never raw counts.

Site $i$'s payoff for group $S$ is

$$u_i(S) = \frac{1}{|V_i|}\sum_{x \in V_i} \log \mathrm{NB}\big(x \mid \hat\mu_S(x), \hat\phi_S\big),$$

where $V_i$ is site $i$'s private held-out sample set, $\hat\mu_S$ are the expected counts under the model fit by $S$, and $\hat\phi_S$ are its estimated gene dispersions. Using the count likelihood rather than accuracy or squared error addresses a problem noted in the research file. pFedSV values partners with generic validation losses, which ignore overdispersion. The Negative Binomial model follows BMDL, which argues against log-transforming counts before modeling them.

A merge of groups $S$ and $T$ is executed only if

$$\mathrm{LCB}_{1-\alpha}\big[u_i(S\cup T)-u_i(S_i)\big] \ge 0 \ \ \forall i \in S\cup T, \quad \text{and} \quad \mathrm{LCB}_{1-\alpha}\big[u_j(S\cup T)-u_j(S_j)\big] > 0 \ \text{for some } j.$$

Here $S_i$ is site $i$'s current group. $\mathrm{LCB}_{1-\alpha}$ is a paired bootstrap lower confidence bound over the samples in $V_i$. We apply a Bonferroni correction across members, meaning $\alpha$ is divided by the number of members being tested. Splits obey the same test in reverse.

The deterministic Pareto order is what guarantees convergence in Coalitional Games for Distributed Eavesdroppers Cooperation and Distributed Coalition Formation Games for Secure Wireless Transmission. That guarantee assumes payoffs are known exactly. Refitting models introduces estimation noise that could reopen cycles. We therefore cache payoff evaluations and keep a record of partitions already visited, and we will prove convergence for the certified rule.

The file names estimation variance in marginal-value computations as an open problem in the coalition literature. The confidence-bound rule is our answer to it in this setting. Following the no-grand-coalition results of the UAV coalition study and the hierarchical federated incentive paper, we expect stable partitions of several mid-sized groups rather than one consortium.

The payoffs are non-transferable, meaning a hospital's predictive quality cannot be paid out to another hospital. This matches the clinical setting better than the transferable-profit game of the hierarchical incentive paper.

Baselines

The naive reference points are solo analysis and the grand coalition. In solo analysis, each site fits the same Negative Binomial model on its own data. The grand coalition is one federated model over all sites. It is run both with the same count model and with a Fed-ComBat-style shared harmonization adapted to counts, following Fed-ComBat, which forces every site into one harmonization space. The grand coalition is the design every existing federated omics system uses, so beating it on the worst-off site is the central test. Beating it means a higher minimum-site held-out likelihood with no drop in mean likelihood beyond noise.

The strongest existing method for avoiding negative transfer on this data type is centralized BMDL. It learns which latent gene modules each cohort shares, using binary on/off switches per module. It reports 78.10% subtype accuracy when head and neck cancer data assists lung cancer classification, against as low as 51% for competing transfer methods. It also shares 98 of 100 factors between two lung sequencing platforms but only 25 between lung and head and neck cancer. We run it with pooled raw data. It is therefore an upper reference that ignores privacy, not a fair federated competitor. Coming within a few points of it while never pooling data would count as success. The grouping it implies, platform duplicates together and unrelated tumors apart, also serves as an external check on the partitions our method finds.

The strongest federated partner-selection method is pFedSV. Each client estimates Shapley values, a game-theoretic measure of each partner's average marginal contribution, by Monte Carlo sampling, then keeps partners whose contribution is positive. It reports 61.18% on CIFAR-10 under a Dirichlet label split with $\alpha=0.1$, against 53.83% for FedFomo. We found no reported result on Lacuna for pFedSV on RNA-seq data. We will run it with the same Negative Binomial model, the same held-out sets, and a matched budget of federated fits.

Two ablations isolate the key ingredient. The first uses the same payoff and merge moves but a utilitarian rule: merge if the summed payoff rises. The second uses the Pareto rule on point estimates without confidence bounds. The proposal wins only if the certified rule has fewer sites made worse than solo than both ablations. A third ablation swaps the Negative Binomial payoff for accuracy, to test the likelihood choice.

Why not just…?

Why not just cluster sites by expression similarity and fit one model per cluster? This is the most serious threat. Hierarchical clustering on each site's averaged expression profile needs no game and may recover cancer-type groupings well. It is symmetric, though, and blind to sample size. A large site and a small site with similar profiles look interchangeable, even if the large site's batch artifacts would swamp the small one. The clustering also needs a threshold and gives no per-site guarantee. We include it with its threshold tuned on held-out data. If it matches the certified rule on both mean and worst-site payoff, we will report that the game adds nothing beyond its guarantee.

Why not just use personalized weighting like pFedSV instead of disjoint groups? Per-site weights may predict slightly better for each site individually. However, they produce no shared artifact. There is no common batch correction and no common set of dispersion estimates on which a joint differential expression study can be run. pFedSV itself does not form stable disjoint partitions, and its partner choices need not be mutual. We compare on per-site likelihood, and separately on the reproducibility of differentially expressed gene lists within groups. Grouping should win on the second measure even if weighting wins narrowly on the first.

Why not just harmonize everyone with Fed-ComBat and pool? Harmonization removes location and scale shifts. It cannot fix genuine biological mismatch, which BMDL's 25-of-100 shared factors between lung and head and neck cancer show is large. The grand-coalition baseline tests this directly, with harmonization applied.

Proposed Experiments

The main testbed builds simulated consortia from The Cancer Genome Atlas (TCGA) bulk RNA-seq, the benchmark used by both PPML-Omics and BMDL. We split cohorts into 10 to 40 synthetic sites that vary in three ways: cancer type (including the lung and head and neck pairing from BMDL), sequencing platform (the two lung platforms), and site size (25 to several hundred samples, echoing BMDL's 25-sample regime). Into chosen sites we inject known batch distortions of controlled strength.

Metrics are:

  • per-site held-out Negative Binomial log-likelihood;
  • subtype classification accuracy;
  • the fraction of sites worse than solo;
  • the worst-site gap to solo;
  • agreement of the recovered partition with the planted structure, measured by the Adjusted Rand Index (a score of agreement between two groupings, 1 for identical);
  • the number of federated fits needed to converge.

A second testbed treats the four single-cell cohorts used by PPML-Omics (Yan, Pollen, Camp-Liver, Hrvatin) as sites. These are distinct tissues, so the correct outcome is mostly refusal to merge, and clustering quality should not fall below solo. For context, PPML-Omics reports an Adjusted Rand Index of 0.859 on Hrvatin.

Three checks decide whether the idea works. First, under strong injected batch effects, the certified rule should keep the fraction of sites worse than solo near the nominal error rate $\alpha$, while the utilitarian and uncertified ablations and the grand coalition should exceed it. Second, recovered groups should pair the two lung platforms and separate unrelated cancers, mirroring BMDL's sharing pattern without pooling data. Third, mean payoff should not lag the best non-guaranteed baseline by more than bootstrap noise. If similarity clustering passes all three checks equally, the contribution reduces to the statistical no-harm guarantee, and we will present it as that.