Abstract
This work introduces a method to tune a sequence-based generative model for molecular de novo design that through augmented episodic likelihood can learn to generate structures with certain specified desirable properties. We demonstrate how this model can execute a range of tasks such as generating analogues to a query structure and generating compounds predicted to be active against a biological target. As a proof of principle, the model is first trained to generate molecules that do not contain sulphur. As a second example, the model is trained to generate analogues to the drug Celecoxib, a technique that could be used for scaffold hopping or library expansion starting from a single molecule. Finally, when tuning the model towards generating compounds predicted to be active against the dopamine receptor type 2, the model generates structures of which more than 95% are predicted to be active, including experimentally confirmed actives that have not been included in either the generative model nor the activity prediction model.
Introduction
Drug discovery is a daunting search through a chemical space estimated at $10^{60}$ to $10^{100}$ potentially synthesizable molecules. Traditional methods for navigating this "haystack" often rely on computationally expensive docking simulations or rule-based transformations limited by human bias. This paper addresses the challenge of inverse QSAR (Quantitative Structure-Activity Relationship): starting with a desired biological activity and working backward to design the molecule that causes it.
Methodology
The proposed method utilizes a sequence-based approach centered on SMILES (Simplified Molecular Input Line Entry System) strings. SMILES represents chemical structures as text, which allows the use of powerful natural language processing architectures like Recurrent Neural Networks (RNNs).
Molecular Representation
To process molecules, the model converts SMILES strings into a one-hot encoded matrix. This representation allows a 3-layer Gated Recurrent Unit (GRU) based RNN to learn the "grammar" of chemistry—how to construct valid, drug-like molecules.
The Reinforcement Learning Framework
The core innovation is a two-step training process using Reinforcement Learning (RL):
- The Prior: An RNN is first pre-trained on 1.5 million molecules from the ChEMBL database. This "Prior" model learns general chemical syntax and the distribution of known drug-like space.
- The Agent: To generate molecules with specific properties, the RNN is treated as an RL agent. The authors introduce Augmented Episodic Likelihood to fine-tune the Agent.
Standard RL often suffers from "mode collapse," where a model finds a simple "cheat code" to maximize rewards. To prevent this, the cost function anchors the Agent to the Prior. The Agent is rewarded for satisfying a user-defined scoring function (e.g., activity against a specific receptor) but is penalized if it drifts too far from the realistic chemical space defined by the Prior.
Experiments and Results
The model was evaluated on three distinct tasks:
- Sulfur-Free Design: As a proof of concept, the model was tasked with avoiding sulfur. It successfully shifted its output from 34% sulfur-containing molecules to nearly 0% while maintaining high "drug-likeness."
- Celecoxib Analogues: The model generated close analogues to the drug Celecoxib. Remarkably, even when all molecules similar to Celecoxib were removed from the Prior's training data, the Agent was able to "rediscover" this chemical space, demonstrating its ability for scaffold hopping.
- DRD2 Target Activity: When optimized for the Dopamine Receptor Type 2 (DRD2), over 95% of the generated molecules were predicted to be active by an external classifier, compared to only 3% from the starting Prior model. Crucially, the model generated molecules experimentally known to be active that were not present in any of the training sets, proving it learns underlying bioactivity features rather than just memorizing data.
Discussion and Future Work
While highly effective, the current approach primarily focuses on single-objective optimization. Real-world drug discovery requires balancing multiple parameters, including toxicity, solubility, and synthetic accessibility. Future work will likely involve multi-parametric scoring functions to better simulate the complex trade-offs required in clinical development. Additionally, explicitly accounting for the ease of synthesis would further bridge the gap between computational design and laboratory realization.
Summary updated