📊 statistics

FL-Sailer: Efficient and Privacy-Preserving Federated Learning for Scalable Single-Cell Epigenetic Data Analysis via Adaptive Sampling

FL-Sailer is a novel federated learning framework that overcomes the ultra-high dimensionality, sparsity, and heterogeneity of single-cell ATAC-seq data through adaptive leverage score sampling and an invariant VAE architecture, enabling privacy-preserving, scalable, and superior collaborative epigenomic analysis across institutions.

Original authors: Guangyi Zhang, Yi Dai, Yiyun He, Junhao Liu

Published 2026-05-07

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Guangyi Zhang, Yi Dai, Yiyun He, Junhao Liu

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are trying to solve a massive jigsaw puzzle, but the pieces are scattered across five different locked rooms. Each room belongs to a different hospital, and strict privacy laws forbid anyone from taking the puzzle pieces out of their room. You need to build the picture together, but you can't move the pieces.

This is the challenge scientists face with single-cell epigenetic data (specifically scATAC-seq). This data is like a super-detailed map of how our genes are turned on or off in millions of individual cells. It's incredibly valuable for understanding diseases, but it's also:

Huge: It contains millions of "pieces" (features) per person.
Sparse: Most of the pieces are blank (95% empty space).
Private: Hospitals can't share the raw data due to patient privacy laws.

Enter FL-Sailer, a new method proposed in this paper that acts like a clever "remote puzzle solver." Here is how it works, using simple analogies:

1. The Problem: Too Heavy to Carry

If you tried to send the entire puzzle (the raw data) from one hospital to a central server to be assembled, the file would be so massive it would clog the internet, and it would violate privacy rules. Standard "Federated Learning" (where models are sent to the data instead of the other way around) usually fails here because the "model" itself is too heavy to send back and forth. It's like trying to mail a library of books just to update a single page.

2. The Solution: The "Smart Highlighter" (Adaptive Sampling)

FL-Sailer's first trick is Adaptive Leverage Score Sampling.

Imagine you have a 1,000-page document, but only 200 pages actually contain the important story; the rest are just blank pages or repetitive footnotes. Instead of mailing the whole 1,000 pages to your friends, you use a "Smart Highlighter" to pick out only the 200 most important pages.

How it works: The algorithm mathematically identifies which genomic regions (the "pages") are biologically interesting and discards the rest.
The Result: It shrinks the data size by 80%. Instead of sending a heavy truckload of data, they send a small, lightweight package. Crucially, the paper claims this doesn't just save space; it actually improves the puzzle by removing the "noise" (blank pages) that confuses the solver.

3. The Second Trick: The "Universal Translator" (Invariant VAE)

Even if you shrink the data, different hospitals might have used slightly different microscopes or protocols. This creates "batch effects"—like if one group of friends drew the puzzle pieces in blue ink and another in red ink. If you just mix them, the picture looks messy.

FL-Sailer uses a special architecture called an Invariant VAE (Variational Autoencoder). Think of this as a Universal Translator.

It learns to separate the "story" (the actual biological signal) from the "accent" (the technical noise caused by different labs).
It strips away the "accent" so that a cell from Hospital A looks exactly the same as a similar cell from Hospital B, even if they were measured differently. This allows the global model to see the true biological patterns without being confused by the lab equipment differences.

4. The Assembly: Building the Picture Together

Now, the process works like this:

Local Highlighting: Each hospital uses the "Smart Highlighter" to pick the top 20% of their most important data.
Local Translation: They train a small model locally to learn the "story" while ignoring their specific "accent."
Sending Updates: They send only the learned rules (the model updates) back to a central server, not the data itself. Because the data was shrunk, these updates are tiny.
Global Assembly: The server combines these rules to build a better, global understanding of the puzzle.

What Did They Prove?

The paper doesn't just say "it works"; they provide a mathematical proof (a "convergence guarantee") showing that this method will eventually find the correct answer, even with such aggressive data shrinking.

In their tests, they compared FL-Sailer to two other approaches:

Centralized Method: Trying to put all data in one place (impossible due to size/privacy).
Standard Federated Learning: Trying to share data without shrinking it (failed because it was too heavy and noisy).

The Result: FL-Sailer didn't just work; it outperformed the centralized method in many cases. By removing the "noise" (the 80% of data that wasn't needed), the model actually saw the biological patterns more clearly than if it had tried to process the messy, full dataset.

Summary

FL-Sailer is a privacy-preserving tool that allows hospitals to collaborate on massive genetic puzzles without ever sharing the raw pieces. It does this by:

Throwing away the junk (shrinking data by 80% to make it fast and private).
Ignoring the accents (removing technical noise so different labs can compare apples to apples).
Proving mathematically that this shortcut leads to the right answer.

The paper concludes that this approach turns a "computationally impossible" problem into a practical, superior way to study human biology across institutions.

Technical Summary: FL-Sailer

Problem Statement

Single-cell ATAC-seq (scATAC-seq) enables high-resolution mapping of chromatin accessibility but faces significant barriers to multi-institutional collaboration. While Federated Learning (FL) offers a privacy-preserving alternative to data centralization, standard FL algorithms fail in the scATAC-seq domain due to three compounding challenges:

Ultra-high Dimensionality: Genomic datasets contain $10^5$ to $10^7$ features (peaks/regions), making per-round communication costs prohibitive (gigabytes per round).
Extreme Sparsity: Data matrices are approximately 95% sparse, obscuring biological signals and causing high variance in local stochastic gradients.
Severe Cross-Institutional Heterogeneity: Technical confounders (e.g., sequencing depth, batch effects) vary significantly across institutions, often correlating with biological signals, leading to model divergence or the learning of spurious batch-specific features.

Existing FL methods for single-cell data (e.g., scRNA-seq) are insufficient because they do not address the order-of-magnitude increase in feature dimensionality and sparsity specific to scATAC-seq.

Methodology: FL-Sailer

The authors propose FL-Sailer, the first FL framework specifically engineered for scATAC-seq. It integrates two core innovations to address dimensionality and heterogeneity simultaneously:

1. Adaptive Leverage Score Sampling

To overcome communication bottlenecks, FL-Sailer employs a feature selection strategy based on column leverage scores.

Mechanism: Clients compute approximate leverage scores locally using randomized QR decomposition (avoiding expensive SVD). The server aggregates these scores to construct a global sampling distribution that prioritizes biologically informative genomic regions.
Effect: This reduces the feature dimension from $d$ to $s$ (where $s \ll d$ ), achieving an 80% reduction in dimensionality.
Theoretical Guarantee: The sampling acts as a subspace embedding, preserving the optimization geometry and gradient norms of the original high-dimensional problem with bounded error. Crucially, it selects features without replacement to ensure diverse coverage of regulatory elements.

2. Invariant VAE Architecture

To handle heterogeneity and batch effects, the framework adapts the SAILER architecture into a federated setting.

Architecture: A chromosome-block Variational Autoencoder (VAE) that learns a latent representation $z$ of the chromatin accessibility profile $x$ .
Invariance Objective: The model explicitly minimizes the mutual information $I(z; c)$ between the latent representation $z$ and technical confounders $c$ (e.g., sequencing depth, institution ID) via a variational upper bound. This disentangles biological signals from technical noise.
Federated Training: Clients train the VAE on their sampled data ( $n \times s$ matrices) and aggregate updates to learn a unified biological representation robust to local technical variations.

Algorithm Overview

The training process (Algorithm 1) consists of two phases:

Federated Dynamic Feature Selection: Clients compute local leverage scores; the server aggregates them to sample a global set of features $S$ , which is broadcast to all clients.
Subspace Federated Training: Clients perform local training on the reduced feature space using the invariant VAE objective. Updates are aggregated via FedAvg.

Key Contributions

Adaptive Leverage Score Sampling: A novel sampling module that preserves biological interpretability by directly selecting genomic regions. It reduces communication costs by orders of magnitude (up to 80%) while improving clustering performance by focusing computation on high-signal regulatory regions.
Invariant Representation Learning: A VAE architecture adapted for FL that explicitly disentangles biological signals from technical confounders through mutual information minimization, addressing severe batch effects common in multi-institutional data.
Convergence Theory: The paper establishes the first convergence guarantees for FL with aggressive feature sampling. It proves that FL-Sailer converges to an approximate solution of the original high-dimensional problem with a bounded error term $O(\epsilon \cdot L(U^*))$ , where $\epsilon \to 0$ as dimensionality increases.
Systematic Benchmarking: Extensive evaluation on synthetic and real-world datasets (Brain PFC, PsychENCODE, PBMC) demonstrating that FL-Sailer enables previously infeasible collaborations and often outperforms centralized methods.

Experimental Results

The authors evaluated FL-Sailer on three real-world scATAC-seq datasets (Brain PFC: 127K features, PsychENCODE: 423K features, PBMC: 108K features) and simulated data.

Performance vs. Centralized: FL-Sailer outperformed centralized SAILER (the state-of-the-art baseline) on real-world data. For example, on the PsychENCODE dataset, FL-Sailer achieved an Adjusted Rand Index (ARI) of 0.781, compared to 0.626 for centralized SAILER with full data access.
Robustness to Heterogeneity: In scenarios with confounded heterogeneity (where technical noise correlates with cell type), centralized methods failed (ARI = 0.221), whereas FL-Sailer achieved an ARI of 0.871 by disentangling signals.
Communication Efficiency: At an optimal sampling rate ( $\rho = 0.2$ ), FL-Sailer reduced total communication by approximately 80% (e.g., from 20.40 GB to 4.12 GB for PsychENCODE over 100 rounds) while simultaneously improving model accuracy.
Comparison to Standard FL: FL-Sailer significantly outperformed generic FL baselines (FedAvg, FedProx, SCAFFOLD, etc.), which often suffered from performance collapse (e.g., FedAvg ARI = 0.418 on PsychENCODE).
Rare Cell Detection: The framework successfully preserved rare cell populations (e.g., Common Myeloid Progenitors at ~2.1% prevalence) under extreme class imbalance, achieving perfect clustering (ARI = 1.000) in synthetic tests.

Significance and Claims

The paper claims that FL-Sailer establishes a superior paradigm for collaborative epigenomic research when tailored to domain-specific constraints.

Feasibility: It transforms a computationally impossible problem (federated training on millions of features) into a practical solution.
Paradigm Shift: The work demonstrates that federated learning, when combined with domain-specific techniques like adaptive sampling and invariant learning, can not only match but surpass centralized methods. The authors attribute this to the sampling strategy acting as an "implicit regularizer" that suppresses technical noise and aligns gradients across heterogeneous clients.
Privacy and Scalability: It enables privacy-preserving multi-institutional studies without centralizing sensitive raw profiles, addressing critical regulatory and storage barriers in genomics.

The authors conclude that their work fills a theoretical void by providing convergence guarantees for FL under aggressive feature selection and empirically validates that decentralized, privacy-preserving frameworks can yield higher quality biological insights than centralized approaches in the presence of technical heterogeneity.