Hospitals that each hold a few hundred tumor profiles would like to find disease genes together without pooling patient records. Federated learning makes this possible: a shared model travels to each site, and only updates come back. To make the guarantee formal, those updates must also carry differential privacy noise, meaning random noise calibrated so that no single patient's record measurably changes what leaves the site. Gene expression data breaks this recipe. With roughly 18,000 genes and a few hundred patients, isotropic noise added to every coefficient drowns the signal. The one federated, private tumor classifier on TCGA breast cancer in our file worked only after a correlation filter cut 17,814 genes down to 9 (Differentially Private, Federated Learning on Gene Expression Data for Tumour Classification). That choice gives up the goal of finding pathway-level biomarkers.
Our proposal is to use the public gene interaction network a second time. In centralized analysis, the graph-guided fused lasso, a penalty that pushes coefficients of genes linked in a protein-interaction network to be exactly equal, already recovers coherent modules and weak coordinated signals that the plain lasso misses (Curtis et al.; Wan et al.). We claim the same penalty, applied by the server to each round's noisy aggregate, also works as a denoiser for the privacy noise. If the true coefficients are piecewise constant on the graph, the noise that survives should scale with the number of gene modules rather than the number of genes. This denoising costs no extra privacy budget, because it is post-processing of an already-privatized quantity.
Core Claim
Simply "federating" the graph-fused lasso is not by itself a research contribution, and we say so up front. The network is public (BioGRID) and the penalty does not touch patient data. The server can therefore compute the hard, nonseparable proximal step itself, using smoothing proximal gradient (Xi Chen et al.) or an exact graph-fused solver, while clients only send loss gradients. The obstacle described for decentralized settings disappears once the penalty step lives on the server.
The open question is a different one: whether the graph prior can make record-level differential privacy affordable when the number of genes greatly exceeds the number of patients. We separate two roles for the graph. The first is a statistical penalty in the objective. The second is a per-round denoising strength tied to the known noise level. Clients $k=1,\dots,K$ hold losses $f_k$, and the target is
$$\min_\beta \sum_k f_k(\beta) + \lambda\|\beta\|_1 + \gamma \sum_{(j,l)\in E} w_{jl}\,|\beta_j - s_{jl}\beta_l|,$$
where $w_{jl}$ are public edge weights and $s_{jl}\in\{\pm1\}$ are edge signs. At round $t$ the server receives a securely summed, noise-perturbed gradient $\hat g^t$ and sets
$$\beta^{t+1} = \arg\min_b \tfrac12\|b-(\beta^t-\eta_t \hat g^t)\|^2 + \eta_t\lambda\|b\|_1 + (\eta_t\gamma+\tau_t)\sum_{(j,l)\in E} w_{jl}|b_j - s_{jl} b_l|.$$
Here $\tau_t \propto \eta_t\sigma_t$ is extra fusion scaled to the noise standard deviation $\sigma_t$ that the step injects. It decays with the step size, so it denoises early rounds without permanently biasing the solution. The decaying-step schedule follows the analysis in Liu and Zhou, which shows that polynomially decaying steps remove the variance floor of local descent under heterogeneity.
A second, concrete issue is that the original fused-lasso signs $s_{jl}$ come from sample correlations (Curtis et al.), which is a privacy leak. We estimate them privately, only on the $|E|$ network edges and not over all gene pairs, through secure aggregation with a small, accounted share of the privacy budget. We compare this against signs taken from public co-expression data.
Relation to Prior Work
The closest work is Exploring Private Federated Learning with Laplacian Smoothing. It applies the quadratic smoother $(I+\sigma L)^{-1}$ to noisy aggregated gradients, with $L$ built over neural-network parameter indices, and gains up to 5% MNIST accuracy at $\epsilon=6$. That paper supplies the post-processing idea. We differ in three ways: we use a biological feature graph, we use an absolute-value fusion that produces exact ties and exact zeros (so selection happens inside training), and we target sparse models where genes far outnumber patients.
Federated feature selection so far ignores feature structure. Barik and Honorio recover lasso support in one shot by majority vote over local marginal screens, and SeiFS securely ranks features one at a time. FMFS adds structure, but only as disjoint surface patches in a federated group lasso, and without formal privacy.
On the centralized side, Structured Integrative Learning pools cohorts under an overlapping graph penalty but assumes the data can be pooled. Its robustness to 20% edge deletion motivates our graph-corruption tests.
Two papers constrain the design. PPML-Omics shows that shuffling can rescue utility for deep models; we treat this as complementary amplification rather than as a competitor. Flimma shows that exact secure aggregation of sufficient statistics beats meta-analysis under label imbalance, which tells us that local-fit-then-combine baselines should lose.
Research Program
The work has four parts. We first implement the server-side fused proximal step with record-level Gaussian noise added at clients before secure summation, and track privacy with a standard composition accountant. We then derive, for squared loss, a bound on the noise-induced error of the denoised iterate in terms of the number of edges where the true $\beta$ changes, and test empirically whether that bound tracks behavior. Third, we study the private estimation of edge signs. Finally, we run controls that ask whether the gains come from the biology of the network or from any sparse smoothing. For that last question we replace BioGRID with a degree-matched random graph and with networks that have 20% of edges deleted or 20% of edges added spuriously.
Baselines
The naive reference is the published two-stage pipeline: correlation-based pre-filtering followed by differentially private federated training. It reports 92.1% tumor-versus-normal accuracy at $\epsilon=1$, $\delta\approx10^{-6}$, against a 96.6% non-private baseline on TCGA-BRCA with 590 samples (Differentially Private, Federated Learning on Gene Expression Data for Tumour Classification). We rerun it on our splits rather than borrowing the number, and we also report it on harder subtype tasks. Beating it means matching its accuracy at equal $\epsilon$ while also selecting a multi-gene, connected module.
The centralized non-private upper bound is Logit-Lapnet, which recovered ESR1, PGR and ERBB2 for triple-negative breast cancer with more than 63% of the 262 selected genes interconnected (Wan et al.). We also run a centralized graph-fused solver; for reference, smoothing proximal gradient reports 29.23% error on the 295-tumor, 8,141-gene dataset (Xi Chen et al.).
The strongest private competitor is DP-Fed-LS (Liang et al.) given the same BioGRID Laplacian, so that it differs from our method only in using quadratic rather than absolute-value fusion. No reported result on gene expression was found on Lacuna for it, so we report our own runs. Other private baselines are private federated lasso (identical pipeline with $\gamma=\tau_t=0$), the one-shot majority-vote support recovery of Barik and Honorio, and meta-analysis of locally fitted graph-regularized models.
The key ablation sets $\tau_t=0$, keeping the graph only in the objective. This isolates the noise-calibrated denoising term.
For fairness, all methods share the same splits, the same total $\epsilon$ (including the budget spent on edge signs), the same number of rounds and bytes sent, and the same hyperparameter grid. Tuning uses simulated data and one held-out public cohort, because tuning on private sites would itself leak information. Winning means higher module recovery and held-out accuracy than DP-Fed-LS at $\epsilon\le 3$, together with a gap to the centralized non-private bound that stays roughly flat as the number of genes grows.
Why not just…?
Why not just apply DP-Fed-LS with the protein-interaction Laplacian? This is the real threat. Quadratic smoothing may capture most of the denoising, since it is also free post-processing. Our argument is that quadratic smoothing blurs module boundaries and never produces exact zeros or ties, so selection still needs a separate step. The head-to-head comparison at equal $\epsilon$ settles the question. If quadratic smoothing matches our method on both prediction and module recovery, we will report that the absolute-value fusion is unnecessary.
Why not just pre-filter to a handful of genes and train privately? The 9-gene result shows this works for a coarse binary task. Marginal filters, however, miss genes whose individual effects are weak but which act together. Curtis et al. found 10 novel hotspots only after fusing over modules, and the plain lasso missed PGR in Wan et al.. Our planted-weak-module simulations and the PGR recovery check measure this directly.
Why not just securely aggregate sufficient statistics, as Flimma does, and solve centrally? For squared loss, this is exact and protects the intermediate quantities (Flimma). It offers no formal guarantee on the released model, though. Making $X^\top X$ differentially private requires noise on roughly $p^2$ entries, which is about $3\times10^8$ for TCGA-BRCA. We include this approach as a non-private-output reference, along with a version that adds noise to the statistics.
Why not just combine local models by meta-analysis or majority vote? Flimma shows meta-analysis F1 falling to 0.93–0.95 under label imbalance. Local fits with $n_k\ll p$ are also unstable. We test both approaches on label-imbalanced site splits.
Proposed Experiments
We use four datasets. TCGA-BRCA (590 to 850 samples, 17,814 genes) is split into 5 to 10 simulated hospitals, with both balanced and label-imbalanced partitions; tasks are tumor versus normal and triple-negative versus other subtypes. The 8,141-gene breast cancer dataset is used with MSigDB pathways. ADNI multi-stage data is used following SIL. Finally, simulations on the real BioGRID graph plant modules with known boundaries.
Metrics are held-out AUC and accuracy, F1 and Jaccard overlap of selected genes against the planted or centralized support, the fraction of selected genes that are interconnected, recovery of ESR1, PGR and ERBB2, and membership-inference advantage. Each is reported at $\epsilon\in\{1,3,6\}$.
Three checks decide the outcome. First, at fixed $\epsilon$, the excess error of our method over the centralized fit should stay roughly constant as the number of genes grows from 2,000 to 18,000 when the number of modules is held fixed, while that of private lasso grows. Second, our method should recover PGR at $\epsilon=1$ where private lasso does not. Third, replacing BioGRID with a degree-matched random graph should remove the gain; if it does not, the benefit is generic smoothing and not use of biological knowledge, and we will say so.
References
- Differentially Private, Federated Learning on Gene Expression Data for Tumour Classification
- Structured association analysis leads to insight into Saccharomyces cerevisiaegene regulation by finding multiple contributing eQTL hotspots associated with functional gene modules
- Molecular pathway identification using biological network-regularized logistic models
- Smoothing proximal gradient method for general structured sparse regression
- Analysis of regularized federated learning
- Exploring Private Federated Learning with Laplacian Smoothing
- Recovering Exact Support in Federated lasso without Optimization
- Secure and Lightweight Feature Selection for Horizontal Federated Learning
- Federated Morphometry Feature Selection for Hippocampal Morphometry Associated Beta-Amyloid and Tau Pathology
- Integrative learning of structured high-dimensional data from multiple datasets
- PPML-Omics: a Privacy-Preserving federated Machine Learning method protects patients’ privacy in omic data
- Flimma: a federated and privacy-aware tool for differential gene expression analysis