De novo design
Generating entirely new molecular structures from scratch to meet a set of target properties — rather than selecting from existing compound libraries.
De novo molecular design means generating novel chemical structures that weren’t previously known, targeting a specified combination of properties: binding affinity for a particular protein, drug-likeness, synthetic accessibility, solubility, or metabolic stability. The contrast is with virtual screening, which scores existing compounds from a library.
AI has substantially changed what de novo design looks like in practice. Earlier approaches used rule-based fragment assembly or evolutionary algorithms; modern tools use generative models — variational autoencoders, diffusion models, reinforcement learning, or language models trained on SMILES strings — that can explore chemical space more efficiently and produce structurally diverse candidates.
Common approaches:
- Reinforcement learning: generate molecules and score them against a reward function (e.g., predicted binding affinity minus synthetic difficulty); iteratively shift generation toward high-scoring regions
- Diffusion models: learn the distribution of known drug-like molecules, then sample new structures by reversing a noise process
- LLM-based generation: treat SMILES as a language; fine-tune on property-annotated datasets to steer generation toward desired profiles
Key tools for de novo design: REINVENT 4 (AstraZeneca, open-source), RFdiffusion (protein-focused), Molecule.one (incorporates synthesizability constraints)
Critical caveat: generative models can produce structures that look plausible computationally but are difficult or impossible to synthesize. Filtering candidates through synthesizability scores (SA score, SYBA) or retrosynthesis tools before ordering them for experimental validation is standard practice.
Related terms: Retrosynthesis, Virtual Screening, SMILES Notation
Related guide: Chemistry