Abstract
In drug discovery, lead optimization often fails to account for the legal landscape of intellectual property. We present the Molecular Optimization Model with Patentability Constraint (MOMP), a generative framework designed to optimize chemical properties while ensuring patentability. By framing optimization as a multi-domain translation task across low-property, high-property/patented, and high-property/non-patented domains, MOMP generates novel molecules that are chemically effective yet distinct from existing patented compounds. Our results demonstrate that MOMP significantly outperforms state-of-the-art models in success rate, particularly in navigating the trade-off between molecular quality and patent likeness.
Introduction
Lead optimization is a critical phase in drug discovery, where researchers refine a promising molecule's structure to enhance properties like solubility or binding affinity. However, a significant real-world barrier is the existing patent landscape. A molecule with ideal chemical properties is commercially unviable if it infringes on existing intellectual property held by competitors.
Traditional generative models focus on chemical performance but often neglect "patent space." Because patented molecules are scattered throughout the chemical space rather than clustered, standard gradient-based optimization often fails to find "safe" zones. MOMP addresses this by integrating patentability directly into the generative loop, ensuring that optimized candidates are both potent and legally novel.
Methodology: Multi-Domain Translation
The MOMP architecture utilizes a sequence-to-sequence approach based on SMILES notation. The core innovation is a multi-cycle translation framework involving three distinct domains:
- Domain A: Molecules with low property scores (initial candidates).
- Domain B: Molecules with high property scores but high similarity to existing patents (the "danger zone").
- Domain C: Molecules with high property scores and low similarity to patents (the "target zone").
The system comprises several key modules:
- Molecule-Embedding Translation Network (METN): Pre-trains encoders and decoders to map molecules into a continuous latent space.
- Embedding-Embedding Translation Network (EETN): Translators that move representations between domains (e.g., A ⇄ B and B ⇄ C).
- Extended-EETN: A direct path (A to C) that preserves structural similarity to the original input while achieving the target properties.
- Molecular Attention Mechanism: Uses molecular fingerprints from non-patented compounds to steer the generative process away from patented structural motifs during training.
Experimental Results
The model was evaluated using standard benchmarks for QED (Quantitative Estimate of Drug-likeness) and DRD2 (Dopamine Receptor D2). Patent data was sourced from SureChEMBL to calculate "Patent Likeness" (PL) scores. A key finding was the positive correlation between property scores and patent likeness, highlighting the necessity of an explicit patentability constraint.
MOMP was compared against several state-of-the-art (SOTA) models, including UGMMT, JTVAE, and REINVENT. The primary metric was "Success," defined as generating a molecule that is novel, high-performing, structurally similar to the lead, and has low patent likeness.
As shown in the table above, MOMP achieved a success rate of 0.272 for QED and 0.322 for DRD2, significantly outperforming baselines. Notably, MOMP consistently achieved lower PL scores than other models reaching similar property values, confirming its ability to find unoccupied chemical space.
Discussion and Conclusion
MOMP represents a significant advancement in aligning AI-driven drug discovery with the practicalities of the pharmaceutical industry. By treating patentability as a domain translation problem, the model successfully navigates the inherent correlation between high-quality molecules and existing patents.
While effective, the model currently relies on Tanimoto similarity, which may not capture all nuances of legal patent infringement. Future work will focus on incorporating more complex legal definitions and expanding the framework to handle multi-objective optimization for multiple chemical properties simultaneously.
Summary updated