Accurate estimation of conformational strain in small-molecule ligands is critical for successful structure-based drug design. While torsional strain in acyclic bonds is readily assessed, evaluating strain in saturated ring systems currently necessitates conformational sampling and compute-intensive quantum-mechanical calculations. To address this limitation, RingAnalyzer is introduced, i.e., a machine learning framework predicting conformational strain directly from single three-dimensional molecular geometries. The method introduces a ring-centric 3D equivariant neural network trained on a rigorously curated dataset of 4,226 drug-like molecules and 27,581 conformations. Calculated at the high-fidelity $\omega$B97XD/Def2TZVP level of theory, this dataset captures diverse stereoelectronic effects. Following extensive data augmentation strategies yielding 76,636 conformations, the computational model accurately predicts strain energies with a mean absolute error of 0.54 kcal/mol, demonstrating robustness to structural artifacts and incomplete structures. Prospective application of RingAnalyzer across diverse structural repositories, including small-molecule conformer generators, 3D generative models, co-folding architectures, docking datasets, the Protein Data Bank, and the Cambridge Structural Database, highlights a widespread occurrence of high-energy ring artifacts. As a computationally efficient and quantum-accurate geometric filter, RingAnalyzer enables the rapid identification and elimination of strained ring systems to streamline structure-based molecular design.

Snakebite envenoming is almost exclusively treated with cold-chain-dependent polyclonal antibody antivenom with limited cross-species coverage and frequent adverse reactions. Snake toxin families share highly conserved sequence identi ties with significantly different binding profiles, rendering sequence- and analog based drug design unfeasible. Here, we investigate whether binding-pocket geometry, independent of sequence and ligand similarity, constitutes a gen uine, exploitable design axis for the rapid discovery of novel polyspecific low molecular-weight antitoxins. Using BEAR, a pharmacologically anchored repos itory comprising 12,358 binding pockets across 1,197 drug targets, we systemat ically compared pocket geometry against full-length sequence, domain-trimmed sequence, and ligand chemical space across the druggable proteome. These three modalities were largely non-overlapping. Within tight structural neigh borhoods, pocket geometry reliably recovered pharmacologically related protein pairs and uniquely captured such relationships across evolutionarily distant fam ilies where sequence similarity failed. Leveraging this principle, we deployed a pocket-conditioned generative model to design candidate antitoxins directly from venom toxin structures, entirely bypassing the need for existing chemical matter or homologous targets. These findings demonstrate that binding-pocket geometry exposes functional relationships invisible to both evolutionary lineage and ligand profiles, providing a structural foundation for de novo ligand design against orphan targets in neglected global health applications.

Traditional screening-based drug discovery is inherently limited by the astronomical scale of the underlying chemical space. Generative modelling offers a compelling alternative to the classical search paradigm and enables rational, bottom-up design of novel and target-specific small molecules. However, its impact has been hampered by challenges in synthetic accessibility of the designed compounds and lack of large-scale experimental validation. Here, we introduce LDDM (Large Drug Discovery Model), a generative framework that supports a wide range of drug discovery tasks, including constrained and unconstrained docking, fragment linking and growing, and de novo design. We further introduce a programmable design algorithm that enables accurate and synthesizable design of compounds satisfying various fine-grained objectives. We experimentally validated optimized and de novo designed peptide and small-molecule binders for seven therapeutically relevant proteins: CDO1, cathepsin S, Pin1, PGK1, KRAS, BRD4, and SARS-CoV-2 Nsp3. The best designs were structurally characterized through competition experiments, NMR spectroscopy, and X-ray crystallography, demonstrating high 3D modelling accuracy. Overall, LDDM provides a scalable and flexible platform for the rapid and tailored design of small molecules and non-natural peptides for diagnostic and therapeutic applications.

Medicinal chemistry decisions rarely rely on a single prediction. Instead, in a MedChem project one must balance potency, ADME, physicochemical properties, novelty, and project-specific constraints. To be impactful, AI/ML methods must therefore be integrated into decision-making workflows rather than used as standalone prediction tools.

Through examples from small-molecule drug discovery projects, I will show how generative chemistry, machine learning models for PhysChem and ADME properties, and multi-parameter optimization can be combined with structural and medicinal chemistry expertise to design, prioritize, and refine compounds. Particular emphasis will be placed on prediction reliability through uncertainty estimation and iterative model improvement. By integrating AI/ML approaches with medicinal chemistry expertise and structure- and physics-based methods, these workflows enable more informed compound prioritization, ultimately driving the identification of molecules with improved in vitro quality and more balanced target property profiles throughout the optimization process.