Abstract
Chemical space exploration is a major task of the hit-finding process during the pursuit of novel chemical entities. Compared with other screening technologies, computational de novo design has become a popular approach to overcome the limitation of current chemical libraries. Here, we reported a de novo design platform named systemic evolutionary chemical space explorer (SECSE). The platform was conceptually inspired by fragment-based drug design, that miniaturized a "lego-building" process within the pocket of a certain target. The key to virtual hits generation was then turned into a computational search problem. To enhance search and optimization, human intelligence and deep learning were integrated. Application of SECSE against phosphoglycerate dehydrogenase (PHGDH), proved its potential in finding novel and diverse small molecules that are attractive starting points for further validation. This platform is open-sourced and the code is available at http://github.com/KeenThera/SECSE.
Introduction
The search for new drugs is often a search for a needle in a haystack—except the haystack is unimaginably large. While modern "ultra-large" chemical libraries contain up to $10^{26}$ compounds, the total estimated "druglike" chemical space is a staggering $10^{63}$. Traditional screening methods like High-Throughput Screening (HTS) or standard Virtual Screening (VS) are limited because they can only test what already exists in a predefined library. To find truly novel "hits" for challenging biological targets, scientists need de novo design—building molecules atom-by-atom or fragment-by-fragment. However, de novo design often struggles with "combinatorial explosion" and high computational costs.
The SECSE Platform
SECSE addresses these challenges through a three-module evolutionary workflow designed to explore the chemical universe efficiently.
Molecular Generator
This module uses a library of over 3,000 curated medicinal chemistry rules—derived from literature and expert medicinal chemists—to evolve molecules. These rules cover growing, mutation, bio-isosteric replacement, and chemical reactions, ensuring the generated molecules maintain drug-like properties.
Fitness Evaluator
New molecules are "docked" into a target protein's 3D structure to predict binding affinity. To handle the massive volume of generated structures, the team integrated Chemprop, a graph-based deep learning model. This AI learns to predict docking scores, allowing the system to evaluate millions of molecules without the heavy computational cost of full physics-based simulations.
Genetic Selector
Using a "survival of the fittest" approach, the system selects the best-performing molecules to serve as "seeds" for the next generation. It prioritizes molecules that maintain a consistent binding mode (low RMSD) while improving their docking scores or ligand efficiency, a concept termed "hereditary binding."
Key Insights and Results
- AI as a Force Multiplier: The integration of deep learning (DL) is the "aha moment." Instead of docking every single generated molecule, the DL model learns the "logic" of the protein pocket. By the fifth generation, the DL model achieved an $R^2$ of 0.85, allowing it to accurately pre-screen 66 million molecules.
- Hereditary Binding: SECSE enforces a "hereditary" constraint. It looks for molecules that "remember" how their parents bound to the protein, ensuring evolution stays on a productive path rather than jumping randomly across the chemical space.
- Escaping "Flatland": Traditional HTS libraries are often "flat" (2D-like). SECSE’s rule-based generation produces more spherical, 3D molecules, which are often better at fitting into complex protein pockets.
- PHGDH Case Study: SECSE successfully "rediscovered" 10 known ligands for PHGDH from a simple benzene starting point. In novel discovery, average docking scores surpassed the reference compound within just three generations. The entire five-generation process took only 40.5 hours on an 80-core machine.
Conclusion
SECSE demonstrates that combining medicinal chemistry rules with genetic algorithms and deep learning can significantly accelerate the discovery of novel chemical entities. By exploring vast chemical spaces in a fraction of the time required by traditional methods, it offers a promising path forward for challenging drug targets.
Summary updated