Gene Regulatory Networks

Problem

Single-cell RNA sequencing measures gene expression in individual cells, and the data are used to infer which genes regulate which. The data are sparse and noisy, and researchers process them in different ways before inference. If the inferred network depends on that processing, the conclusions drawn from it do too.

Approach and findings

We first measured that dependence. On five human tissue datasets, the networks inferred after six different zero-imputation methods overlapped very little, with Jaccard similarity below 0.1 for most pairs. On simulated data with a known network, adding technical noise brought inference down to chance, and no imputation method restored it: the best AUC-ROC across all methods and datasets was 0.530.

Because inference from scratch fails in this way, BEACON instead completes a network that is partly known from experiments. It learns separate embeddings for a gene as a regulator and as a target, so that the direction of regulation can be represented, and on simulated data it reached an AUROC between 94.85 and 97.24. A replication-testing framework, presented at RECOMB-Genetics 2026, asks whether the edge rankings that inference methods produce can be reproduced. Separately, we built a compendium of 81,488 transcription factor cascades from the STRING database and used network centrality to select candidate therapeutic targets.

Papers

A Replication-Testing Framework for Single-Cell Gene Regulatory Edge Rankings
Yunwei Zhao, Ankit Bhardwaj, Lakshminarayanan Subramanian
RECOMB Satellite Conference on Genetics (RECOMB-Genetics), 2026. Oral presentation.

BEACON: Bayesian Completion of Gene Regulatory Networks from Single-Cell Expression and Partial Priors
Yunwei Zhao, Ankit Bhardwaj, Preetha Balasubramanian, Lakshminarayanan Subramanian
Under review. Full version of the NeurIPS 2025 workshop paper listed below.

BEACON: Bayesian Contrastive Learning for Single-Cell Gene Regulatory Inference
Yunwei Zhao, Ankit Bhardwaj, Lakshminarayanan Subramanian
NeurIPS Workshop on AI Virtual Cells and Instruments, 2025. [Paper]

Limitations of scRNA-seq Zero-Imputation Methods for Network Inference
Ankit Bhardwaj, Joshua Weiner, Preetha Balasubramanian, Lakshminarayanan Subramanian
ICML Workshop on Machine Learning for Life and Material Science, 2024. [Paper] [Code]

Generation of a Compendium of Transcription Factor Cascades and Identification of Potential Therapeutic Targets Using Graph Machine Learning
Sonish Sivarajkumar, Romy Roy, Pratyush Tandale, Ankit Bhardwaj, Kipp W. Johnson, Anoop Titus, Benjamin S. Glicksberg, Shameer Khader, Kamlesh K. Yadav, Lakshminarayanan Subramanian
Genes, 2025. [Paper]

Collaborators

• Yunwei Zhao

• Joshua Weiner

• Preetha Balasubramanian

• Lakshminarayanan Subramanian (NYU, advisor)