Biologically-driven generative chemistry: using biological data to guide de novo drug design

Review article, 2026

Drug discovery is a dauntingly difficult process due to the enormous space of possible molecules and the complexity of biological systems. Traditional in silico drug design methods have focused primarily on optimising chemical properties, often overlooking or only implicitly handling key biological effects that determine whether or not a compound is a viable drug candidate. In recent years, deep generative models have emerged as a powerful class of machine learning methods capable of generating novel molecular structures tailored to specific design objectives. Increasingly, these models are being conditioned not only on chemical features but also on diverse categories of biological data, from bioassays and protein structures to gene expression signatures and cell morphology profiles. This biologically-guided paradigm holds the potential to improve both the relevance and clinical translatability of generated compounds. In this review, we provide a comprehensive and accessible overview of biologically-conditioned generative models for de novo drug design. We describe core generative modalities while critically examining current challenges in benchmarking and validation. Our goal is to orient researchers with a machine learning background to the data, design choices, and evaluation gaps that define biologically-conditioned molecular generation.