1753-Biological Embedding Alignment for Perturbation Effects

PertOmni is a novel multimodal representation-learning framework designed to analyze cellular perturbation effects across both genetic and small-molecule experiments. By employing a CLIP-style architecture with masked contrastive learning, the system aligns transcriptomic profiles and microscopic cell images with large language model-derived text embeddings of biological descriptions. This approach effectively isolates within-cell-type perturbation dynamics while minimizing confounding heterogeneity across different cell lines. The framework demonstrates robust predictive capabilities in bi-directional cross-modality retrieval, drug-gene interaction inference, and retrieval-based perturbation prediction. Ultimately, this method provides a unified computational strategy to link interventions directly to their functional targets and improve downstream drug discovery applications.

References:

• Long W, Liu T, Szałata A, et al. Learning Perturbation Effects Through Contrastive Alignment of Multimodal Biological Embeddings[J]. bioRxiv, 2026: 2026.06. 23.734145.