Chem2Gen-Bench: Benchmarking Chemical-to-Genetic Translation in Perturbation Response Space

Researchers introduce Chem2Gen-Bench, a comprehensive benchmark dataset containing over 1.3 million chemical and genetic perturbation profiles designed to evaluate how accurately computational models can translate chemical perturbations into genetic responses. The study reveals that while translation between these perturbation types is measurable, it remains heterogeneous across different cellular contexts, and current foundation-model embeddings don't consistently outperform simpler baseline approaches.

The emergence of virtual-cell and perturbation models represents a significant shift in biomedical research, enabling researchers to predict cellular responses without extensive wet-lab experimentation. However, chemical and genetic perturbations affect cells through fundamentally different mechanisms, making direct translation between them non-trivial. Chem2Gen-Bench addresses a critical gap in the field by providing the first large-scale, systematically organized benchmark for evaluating this translation task across matched cell-target contexts.

The benchmark's scale—spanning 260,084 chemical and 1,099,045 genetic perturbation profiles—enables rigorous evaluation of alignment metrics, retrieval success, and feature space quality. The researchers' findings that foundation-model embeddings fail to consistently improve over gene-delta baselines carries important implications for AI development in biology. This suggests that sophisticated representation learning approaches may not automatically translate to practical improvements without careful domain-specific validation.

For the biomedical AI and computational biology sectors, this work establishes critical evaluation standards that will influence how future perturbation models are developed and validated. The heterogeneous translation fidelity across contexts highlights that context-specific optimization remains necessary rather than pursuing one-size-fits-all solutions. The finding that background adjustment increases pairwise similarity associations but decreases retrieval success under certain conditions underscores the complexity of preprocessing decisions in this domain.

Moving forward, researchers must focus on understanding which cellular contexts enable reliable chemical-to-genetic translation and why. The benchmark itself becomes a valuable tool for validating new model architectures and representation learning approaches, potentially driving innovation in virtual-cell prediction systems that could accelerate drug discovery timelines.

• →Chem2Gen-Bench provides the first large-scale benchmark for evaluating chemical-to-genetic perturbation translation across 1.3+ million profiles.
• →Translation fidelity between chemical and genetic perturbations is measurable but inconsistent across different cellular contexts.
• →Current foundation-model embeddings fail to consistently outperform simpler gene-delta baseline approaches in target-matched evaluations.
• →Background adjustment methodology shows trade-offs: increasing similarity associations while decreasing retrieval success in certain settings.
• →The benchmark establishes an auditable framework for validating when chemical and genetic perturbations align and when representation improvements are truly supported by evidence.