Molecule Generation

AI generates and ranks candidate molecules for early drug discovery and lead identification.

Business impact

• Theoretical molecules explored — PolarisQB explores approximately 10^30 theoretical molecules.
• Desirable molecules generated — Shionogi generated 10 times more desirable molecules than classical machine-learning algorithms in the second phase of its project.
• Time to generate molecules — QuADD reduces drug design time from years to hours; one comparison in the article says QuADD generated 3,000 molecules in roughly 30 minutes.
• Binding affinity — QuADD-generated structures showed stronger predicted binding affinity than the comparison system.

Data requirements

• Molecular and protein-target data (text) — The models are trained or evaluated against protein targets and known ligands.
• Chemical structure data (text) — Candidate molecules are generated and compared by structure and physicochemical properties.
• Physicochemical property data (tabular) — The study evaluates molecular weight, LogP, KL divergence, and docking scores.

AI methods and techniques

• Quantum annealing — Used in QuADD and Shionogi’s project to search and optimize candidate molecules under chemical constraints.
• Generative modeling — Used by DiGress, MolRNN, and GraphINVENT to create novel molecules for protein targets.
• Reinforcement learning — Used in the academic study as one training approach for the generative models.
• Docking-based evaluation — Used to score generated molecules against protein targets and compare candidate quality.

AI models and model families

D-Wave Advantage, D-Wave Advantage2, DiGress, MolRNN, GraphINVENT

Industries

Real-world evidence

3 documented case studies on record.

Companies using this: Polaris Quantum Biotech (PolarisQB), Shionogi, University College London and the University of Urbino Carlo Bo.