Domain-aware priors stabilize, not merely enable, vertical federated learning in data-scarce coral multi-omics

TL;DR

The results motivate design principles for VFL in extreme P>>N regimes, emphasizing domain-informed dimensionality reduction, stability-focused evaluation, and interpretable feature selection for scarce biological data.

Abstract

Vertical federated learning (VFL) enables multi-laboratory collaboration on distributed multi-omics datasets without sharing raw data, but exhibits severe instability under extreme data scarcity (P >> N) when applied generically. Here, we investigate how domain-aware design choices; specifically gradient saliency-guided feature selection with biologically motivated priors; affect the stability, interpretability, and failure modes of VFL architectures in small-sample coral stress classification (N = 13 samples, P = 90,579 features across transcriptomics, proteomics, metabolomics, and microbiome data). We benchmark REEF (Robust Expert Encoder Federation), a domain-aware VFL framework, against two baselines on the Montipora capitata thermal stress dataset: (i) a standard NVFlare-based VFL and (ii) LASER, a state-of-the-art label-aware VFL method. REEF achieves an AUROC of 0.776 +/- 0.039 after reducing dimensionality by 98.6% (90,579 to 1,300 features), substantially outperforming NVFlare VFL at chance level (AUROC 0.500 +/- 0.125, p = 0.0106, Cohen's d = 2.265) and numerically exceeding LASER (AUROC 0.557 +/- 0.191, p = 0.0995, Cohen's d = 1.068), with 3-5-fold variance reduction. An equal-weights ablation confirms that biological priors specifically contribute stability: removing priors yields statistically indistinguishable mean AUROC (p = 0.405) but 2.3x higher variance (CV 0.110 vs 0.050). Negative control experiments using permuted labels produce AUROC near or below chance (0.357 for REEF, 0.238 for NVFlare), consistent with the absence of gross data leakage. These results motivate design principles for VFL in extreme P >> N regimes, emphasizing domain-informed dimensionality reduction, stability-focused evaluation, and interpretable feature selection for scarce biological data.

Figures (6)

• Figure 1: Comparative AUROC performance across VFL architectures. Bar chart with error bars (mean $\pm$ SD, five random seeds). REEF (AUROC $0.776 \pm 0.039$) substantially outperforms NVFlare VFL ($0.500 \pm 0.125$) and LASER ($0.557 \pm 0.191$). The equal-weights ablation ($0.814 \pm 0.090$) achieves statistically indistinguishable mean AUROC ($p=0.405$) but 2.3$\times$ higher variance than REEF, confirming that biological priors specifically provide the stability advantage rather than dimensionality reduction alone.
• Figure 2: Statistical significance analysis. Comparison of NVFlare and LASER against the REEF baseline. NVFlare shows statistically significant underperformance ($p = 0.0106$, Cohen's $d = 2.265$). The LASER comparison yields a large effect size (Cohen's $d = 1.068$) but does not reach significance ($p = 0.0995$) due to LASER's high inter-seed variance.
• Figure 3: Training stability analysis across random seeds. (A) Strip plot showing individual per-seed AUROC values (filled circles) and mean bars for all four conditions. REEF (dark blue) shows a compact distribution centered at 0.776 (SD=0.039). The equal-weights ablation (orange) achieves comparable mean performance but with greater spread (SD=0.090 vs 0.039), confirming the 2.3$\times$ variance inflation when biological priors are removed. LASER (light blue) and NVFlare (grey) exhibit high inter-seed variability. Dashed line indicates chance level (AUROC=0.5). (B) Per-seed trajectory showing the ablation's high inter-seed volatility (range 0.714--0.952) compared to REEF's tighter band (0.714--0.833).
• Figure 4: Omic layer saliency under equal-weights ablation (no biological priors). Mean gradient importance per omic layer averaged across all LOOCV folds and five random seeds. Proteomics dominates with $\sim$20$\times$ higher importance than transcriptomics, despite receiving equal feature budgets and uniform embedding weights. Error bars show standard deviation across seeds.
• Figure 5: Impact of domain-aware gradient saliency feature selection. The raw multi-omics feature space (90,579 dimensions) is reduced by approximately 98.6% to a robust subset of 1,300 features prior to federated training. Note the logarithmic scale on the y-axis, emphasizing the magnitude of the "sieve" effect that restores a viable signal-to-noise ratio.