ANNalog: generation of MedChem-similar molecules.

Area of Science:

• Medicinal Chemistry
• Computational Chemistry
• Drug Discovery

Background:

• Generative deep learning models show promise for designing drug-like molecules.
• Medicinal chemistry requires generating analogues with structural similarity and scaffold hopping.
• Existing models often struggle with generating diverse yet relevant chemical transformations.

Purpose of the Study:

• Introduce ANNalog, a transformer-based generative model for molecule design.
• Enable generation of both structurally similar analogues and novel scaffolds.
• Address limitations in current deep learning approaches for analogue generation.

Main Methods:

• Trained a sequence-to-sequence model (ANNalog) on molecule pairs from the same bioactivity assay (ChEMBL33).
• Utilized Simplified Molecular Input Line Entry System (SMILES) encoding for molecular representation.
• Applied Levenshtein distance-guided alignment for preprocessing to enhance model performance.
• Validated scaffold-hopping capabilities using curated data and a case study on orexin-2 receptor antagonists.

Main Results:

• ANNalog successfully generates structurally similar analogues and performs scaffold hopping.
• Preprocessing with Levenshtein distance significantly improved model performance.
• The prefix control feature allowed recovery of approximately 25% of known scaffolds in a patent set.
• Generated molecules demonstrated chemical relevance and structural novelty.

Conclusions:

• ANNalog effectively generates diverse molecular analogues, including scaffold hopping, by leveraging data from shared bioactivity assays.
• The model advances generative capabilities in medicinal chemistry beyond simple similarity measures.
• ANNalog offers a powerful tool for exploring novel chemical space and accelerating drug discovery efforts.