Single-Cell Concept Bottleneck Generative Models for Interpretable and Controllable Cellular Editing

Alma Andersson, Aya Abdelsalam Ismail, Edward De Brouwer, Doron Haviv, Tommaso Biancalani, Kyunghyun Cho, Gabriele Scalia, Aicha BenTaieb, Hector Corrada Bravo

Abstract:

Understanding how any cellular phenotype changes in response to genetic and/or chemical interventions is fundamental to the understanding of cell biology and the design of effective and safe therapeutics. Single-cell RNA sequencing enables characterization at cellular resolution, yet the combinatorial space of possible interventions renders exhaustive experimental mapping infeasible. In-silico counterfactual generation is an attractive alternative but current scRNA-seq generative methods fall short: disentanglement models rarely support interventions, and most intervention-based approaches perform conditional generation that synthesizes new cells rather than editing existing ones.

We introduce Single-Cell Concept Bottleneck Generative Models (scCBGMs), unifying counterfactual reasoning and generative modeling, to overcome these limitations. \method~adapts concept bottleneck architectures for single-cell data through decoder skip connections and a cross-covariance penalty that promotes disentanglement without dimensional constraints. Counterfactual edits are done via a principled abduction–action–prediction procedure that supports fine-grained per-cell perturbations and zero-shot generalization to novel concept combinations. Moreover, conditioning modern generative procedures, e.g., flow matching, on scCBGM embeddings combines state-of-the-art single cell data generation quality with precise controllability.

Across multiple real datasets, including a rigorous evaluation based on a newly developed synthetic benchmark with ground-truth counterfactuals, scCBGM demonstrates superior accuracy in zero-shot generalization and cell-level counterfactual prediction compared to state-of-the-art methods, while providing interpretable control over biological concepts. We also show how counterfactual generation scCBM can be used in-silico to understand mechanism of action of cellular response to chemical perturbation by editing concepts capturing pathway activity scores that mediate response in real scRNA-seq data. scCBGM establishes a principled framework for high-fidelity in silico cellular experimentation and hypothesis testing in single-cell biology.

Submission Number: 44