Traditional lead discovery has relied on quantitative structure–activity relationships (QSAR) and physics-based screening, but exhaustively searching chemical space is infeasible. Modern workflows therefore employ deep learning to learn predictive structure–property mappings. Graph neural networks (GNNs) and transformer models have become widely adopted for molecular property prediction and design, as they natively operate on graph-structured or sequential chemical representations. Variational autoencoders, generative adversarial networks and related architectures embed molecules in continuous latent spaces, enabling inverse QSAR: one can sample or optimize structures to match target bioactivity and physicochemical criteria. These generative models can propose entirely new scaffolds with desired attributes, effectively ‘designing’ candidate leads beyond known libraries. Despite these advances, significant challenges remain. Data sparsity and bias limit model robustness, and many molecular properties (e.g. ADMET endpoints) are measured on limited datasets. Interpretability is limited – deep models often act as black boxes, motivating development of explainable AI techniques. Ensuring scalability to ultra-large libraries and embedding chemical constraints (synthetic feasibility, drug-likeness) is nontrivial. Moreover, lead optimization is inherently multi-objective: models must balance potency, selectivity, and pharmacokinetics, requiring complex trade-offs during design. Looking ahead, emerging strategies promise to address these gaps. Self-supervised pretraining on massive unlabelled chemical corpora is improving feature learning, while explainable AI methods aim to highlight key substructures driving predictions. Early quantum-enhanced machine learning frameworks show promise for accelerating optimization and generation of candidates. Multimodal models that integrate chemical structure with biological assays and omics data may yield richer lead profiles. Federated