Perturbation Guided Drug Molecule Design via Latent Rectified Flow

ICLR 2026 Conference Submission20223 Authors

Keywords: Multi-modal generation, Perturbation biology, Molecular generation

Abstract: Phenotypic drug discovery generates rich multi-modal biological data, yet translating complex cellular responses into molecular design remains a computational bottleneck. Existing generative methods operate on single modalities (transcriptomic or morphological alone) and condition on post-treatment measurements without leveraging paired control-treatment dynamics. We present **Pert2Mol**, the first framework for multi-modal phenotype-to-structure generation that integrates transcriptomic and morphological features from paired control-treatment experiments. Pert2Mol employs separate ResNet and cross-attention encoders for microscopy images and gene expression profiles, with bidirectional cross-attention between control and treatment states to capture perturbation dynamics rather than simple differential measurements. These multi-modal embeddings condition a rectified flow transformer that learns velocity fields along straight-line trajectories from noise to molecular structures, enabling deterministic generation with superior efficiency over diffusion models. We introduce Student-Teacher Self-Representation (SERE) learning where an exponential moving average teacher supervises student representations across network depths, stabilizing training in high-dimensional multi-modal spaces. Unlike previous approaches that require preprocessed differential expression vectors, Pert2Mol learns perturbation effects directly from raw paired experimental data. Experiments on large-scale datasets demonstrate the first successful multi-modal framework for phenotype-driven molecular generation.

Primary Area: applications to physical sciences (physics, chemistry, biology, etc.)

Submission Number: 20223