Stabl: sparse and reliable biomarker discovery in predictive modeling of high-dimensional omic data

Julien Hédou, Ivana Marić, Grégoire Bellan, Jakob Einhaus, Dyani Gaudillière, François-Xavier Ladant, Franck Verdonk, Ina A. Stelzer, Dorien Feyaerts, Amy S. Tsai, Edward A. Ganio, Maximilian Sabayev, Joshua Gillard, Adam Bonham, Masaki Sato, Maïgane Diop, Martin S. Angst, David K. Stevenson, Nima Aghaeepour, Andrea Montanari, Brice GaudillièreResearch Square 2023 · DOI

In the era of precision medicine, researchers are drowning in data. High-content "omic" technologies—like transcriptomics or proteomics—allow us to measure thousands of molecular features from a single blood sample. However, we usually only have samples from a few dozen or hundred patients. This creates a "needle in a haystack" problem ($p \gg n$) where standard machine learning models often pick out "lucky" noise rather than true biological signals.

A new paper recently published in Nature Biotechnology introduces Stabl, a machine learning framework designed to identify a sparse, highly reliable set of biomarkers that actually mean something in a clinical context.

TL;DR

Stabl is a supervised machine learning framework that identifies a minimal, highly reliable set of biomarkers from high-dimensional data. By injecting artificial noise to calibrate a data-driven "reliability threshold," it achieves state-of-the-art predictive performance while drastically reducing the number of false-positive discoveries.

What Problem Does This Paper Address?

When you have 30,000 gene expressions but only 50 patients, standard Sparsity-promoting Regularization Methods (SRMs) like Lasso or Elastic Net struggle. While they are great at prediction, they are notoriously unstable. A tiny change in the training data can lead the model to select a completely different set of "biomarkers."

This lack of reliability is a massive hurdle for clinical translation. If a model identifies 500 potential biomarkers, a clinician cannot realistically develop a diagnostic test for all of them. Furthermore, if half of those 500 are just statistical noise, the biological interpretation of the disease becomes impossible.

Methods

Stabl builds on the concept of Stability Selection but adds a clever twist to make it objective. The process follows three main steps:

  1. Subsampling: The model (Lasso or EN) is trained hundreds of times on different subsets of the data to see which features consistently appear.
  2. Noise Injection: This is the core innovation. Stabl generates "artificial" features—either through random permutations or "knockoffs"—that look like real data but are guaranteed to be unrelated to the clinical outcome.
  3. The Reliability Threshold ($\theta$): Instead of the researcher picking an arbitrary cutoff (e.g., "keep features that appear in 50% of models"), Stabl uses the artificial noise to calculate a surrogate for the False Discovery Proportion ($FDP_+$). It sets the threshold $\theta$ at the point where the estimated false discoveries are minimized.

This makes the model data-driven and adaptive. If the data is very noisy, Stabl sets a high bar; if the signal is clear, it allows more features through.

Datasets

The authors validated Stabl across a diverse range of challenges:

  • Synthetic Data: Used to benchmark sparsity and reliability against a known "ground truth."
  • Preeclampsia (PE): A large-scale cell-free RNA (cfRNA) dataset (37,184 features).
  • COVID-19 Severity: A proteomic dataset (1,463 features) using the Olink platform.
  • Labor Onset: A "triple-omics" integration task combining proteomics, metabolomics, and single-cell mass cytometry.
  • Surgical Site Infection (SSI): A newly generated multi-omic dataset introduced in this paper, tracking 274 patients undergoing major surgery to predict post-operative infections from pre-operative blood.

Key Insights and Ideas

  • Noise as a Calibration Tool: Using "fake" data to set the "real" threshold is a powerful way to bypass human bias. It provides a mathematical "sanity check" for the model.
  • Sparsity is Not Just for Speed: In biology, a sparser model is a more interpretable model. By narrowing down 37,000 features to just 13, Stabl allows researchers to focus on specific biological pathways (like the CDK10 or TRIO genes in preeclampsia).
  • Multi-Omic Harmony: Different types of data (e.g., metabolites vs. proteins) have different noise levels. Stabl can calculate a unique reliability threshold for each "layer" of data before merging them, preventing one noisy dataset from drowning out a high-signal one.

Key Results

  • Dramatic Sparsity: In the Preeclampsia study, Stabl reduced the number of selected features by over 20-fold compared to Lasso, while maintaining nearly identical predictive accuracy (AUROC ~0.83).
  • Higher Reliability: In a COVID-19 validation cohort, 100% of the features selected by Stabl passed a 5% False Discovery Rate (FDR) threshold, compared to only 91% for Lasso.
  • Clinical Discovery: On the new SSI dataset, Stabl identified a pre-operative immune signature (including STAT3 signaling in neutrophils) that predicts which patients are most susceptible to infection before they even go into surgery.

Limitations and Future Work

The authors acknowledge a few areas for growth:

  • Theoretical Bounds: The reliability guarantee currently relies on the assumption that artificial noise and uninformative real features are "exchangeable," which may not always hold perfectly in complex biological systems.
  • Model Expansion: While currently optimized for Lasso and Elastic Net, the framework could theoretically be extended to other SRMs or even deep learning architectures.
  • Comparison: Future studies are needed to compare Stabl against other specialized feature-selection methods like Boruta or Double Machine Learning.

The complete Stabl package is open-source and available on GitHub.

Summary updated

Abstract

High-content omic technologies coupled with sparsity-promoting regularization methods (SRM) have transformed the biomarker discovery process. However, the translation of computational results into a clinical use-case scenario remains challenging. A rate-limiting step is the rigorous selection of reliable biomarker candidates among a host of biological features included in multivariate models. We propose Stabl, a machine learning framework that unifies the biomarker discovery process with multivariate predictive modeling of clinical outcomes by selecting a sparse and reliable set of biomarkers. Evaluation of Stabl on synthetic datasets and four independent clinical studies demonstrates improved biomarker sparsity and reliability compared to commonly used SRMs at similar predictive performance. Stabl readily extends to double- and triple-omics integration tasks and identifies a sparser and more reliable set of biomarkers than those selected by state-of-the-art early- and late-fusion SRMs, thereby facilitating the biological interpretation and clinical translation of complex multi-omic predictive models. The complete package for Stabl is available online at https://github.com/gregbellan/Stabl.
BibTeX
@inproceedings{hedou2023stabl,
  author = {Julien Hédou and Ivana Marić and Grégoire Bellan and Jakob Einhaus and Dyani Gaudillière and François-Xavier Ladant and Franck Verdonk and Ina A. Stelzer and Dorien Feyaerts and Amy S. Tsai and Edward A. Ganio and Maximilian Sabayev and Joshua Gillard and Adam Bonham and Masaki Sato and Maïgane Diop and Martin S. Angst and David K. Stevenson and Nima Aghaeepour and Andrea Montanari and Brice Gaudillière},
  title = {{Stabl: sparse and reliable biomarker discovery in predictive modeling of high-dimensional omic data}},
  booktitle = {Research Square},
  year = {2023},
  doi = {10.21203/rs.3.rs-2609859/v1},
}
Figure