Source-linked AI summary
MolecularRNN: Generating realistic molecular graphs with optimized properties
Mariya Popova, Mykhailo Shvets, Junier Oliva, Olexandr Isayev
TL;DR
MolecularRNN addresses de novo molecular design by generating molecules with desired properties, a challenging need in drug discovery. It combines graph recurrent generation with likelihood pretraining, policy-gradient optimization, valency-based rejection sampling, and structural penalties. The authors report property distribution shifts, performance beyond state-of-the-art methods, and 100% validity during inference.
Problem
De novo design seeks novel molecules with predefined properties for drug discovery and development.
Method
MolecularRNN recurrently generates typed molecular graphs, then uses policy gradients with a critic to optimize target properties.
Results
MolecularRNN shifts generated distributions toward desired lipophilicity, drug-likeness, and melting-point ranges while outperforming baselines on both reported property tasks.
Takeaways & Limitations
Valency-based rejection sampling yields 100% valid inference molecules, while structural penalties extract training signal from invalid intermediates.
Takeaways & Limitations
The BFS ordering assumes the first node is carbon, reflecting a scope boundary for the modeled organic molecules.
Abstract
from arXiv · showhide
Designing new molecules with a set of predefined properties is a core problem in modern drug discovery and development. There is a growing need for de-novo design methods that would address this problem. We present MolecularRNN, the graph recurrent generative model for molecular structures. Our model generates diverse realistic molecular graphs after likelihood pretraining on a big database of molecules. We perform an analysis of our pretrained models on large-scale generated datasets of 1 million samples. Further, the model is tuned with policy gradient algorithm, provided a critic that estimates the reward for the property of interest. We show a significant distribution shift to the desired range for lipophilicity, drug-likeness, and melting point outperforming state-of-the-art works. With the use of rejection sampling based on valency constraints, our model yields 100% validity. Moreover, we show that invalid molecules provide a rich signal to the model through the use of structure penalty in our reinforcement learning pipeline.
1 Introduction
Molecular design seeks novel molecules with desired properties to address the costly, lengthy, and inefficient drug-development process. MolecularRNN combines graph generation, validity constraints, reinforcement learning, and large-scale analysis for this goal.
- Less than one out of every 10,000 drug candidates becomes an approved marketed drug, while development takes approximately 10-15 years and costs $1-3 billion.
- De novo molecular design aims to create novel molecules with desired properties through generation, scoring, and optimization.
- Molecular graphs directly map atoms to nodes and bonds to edges, making them a natural molecular representation.
- MolecularRNN uses a recurrent generative model for molecular graphs, valency-based rejection sampling, structural penalties, and reinforcement-learning property optimization.
- The model targets drug-likeness, lipophilicity, and melting temperature while supporting large-scale experimental analysis.
2 Related work
Prior de novo molecule-generation methods use string or graph representations, each with distinct validity, optimization, scalability, or evaluation limitations. These limitations motivate MolecularRNN’s graph-based approach and distribution-focused evaluation.
- SMILES-based recurrent models can generate chemically invalid samples because they must learn complex grammatical rules.
- SMILES representations are difficult to extend naturally to scaffold optimization from a given molecular core.
- Graph-based methods support valency constraints and offer representations that are more interpretable and intuitive to chemists.
- An earlier sequential graph-generation method considered molecular graphs with at most 20 heavy atoms, limiting its practical size range.
- GCPN and related work often report only top-three molecules, which may not represent property distributions from many generated samples.
- Property-optimization approaches include fine-tuning, transfer learning, reinforcement learning, and adversarial training, with logP, molecular weight, and melting temperature used as proxies.
3 Methods
MolecularRNN sequentially generates typed molecular graphs with recurrent node and edge models, while enforcing chemical validity and optimizing rewards through policy gradients. Its training also uses invalid intermediate structures as localized feedback.
- 3 Methods: MolecularRNN extends GraphRNN to generate graphs with node and edge types, incorporating rejection sampling, reinforcement learning, and structural penalties.
- 3.1 Background: GraphRNN model: Breadth-first node ordering reduces graph-representation complexity and the number of edge predictions.
- 3.2 MolecularRNN: Molecular graphs represent atoms as nodes and categorical bond types as edges, including no, single, double, and triple bonds.
- 3.2 MolecularRNN: Each generated node receives an atom class, while a node with no edges to previous nodes is treated as terminal and its atom-class prediction is ignored.
- 3.2 MolecularRNN: The first node is always a carbon atom in the BFS ordering, because every organic molecule contains at least one carbon atom.
- 3.3 Valency-based rejection sampling: Valency-based rejection sampling enforces per-atom chemical constraints during inference, and incomplete valencies are complemented with hydrogens.
- 3.2 MolecularRNN: NodeRNN predicts successive atom types, while EdgeRNN predicts bond types linking each new atom to preceding atoms.
- 3.4 Property optimization: Policy gradients treat MolecularRNN as a policy that selects atom labels and connections, maximizing expected rewards over valid final molecular graphs.
4 Experiments
Experiments evaluate MolecularRNN’s generation quality at large scale and against graph-generation baselines, then test reinforcement-learning optimization of molecular properties. The model produces realistic, diverse molecules and shifts generated distributions toward desired property values, while valency-based rejection sampling ensures validity.
- 4.1 Unsupervised likelihood training: MolecularRNN generates 1 million novel, diverse, realistic molecular graphs after unsupervised likelihood training on three molecular datasets.The datasets are ChEMBL, ZINC 250k, and MOSES.
- 4.1 Unsupervised likelihood training: MolecularRNN is comparable to GCPN and JT-VAE on validity, uniqueness, and novelty, while exceeding JT-VAE in internal diversity.GCPN tends to produce overly complex molecules with high synthetic accessibility scores.
- 4.2 Property optimization with reinforcement learning: MolecularRNN policy-gradient optimization shifts generated molecules toward higher penalized logP and QED values and outperforms all baselines in both tasks.The optimization starts from a likelihood-pretrained model and uses rewards based on the target properties.
- 4.2 Property optimization with reinforcement learning: MolecularRNN shifts the full maximized-QED distribution farther toward maximum values than GCPN.The paper evaluates the optimized distribution rather than only comparing the top three molecules.
- 4.2 Property optimization with reinforcement learning: Melting-temperature optimization shifts the predicted-property distribution and rediscoveries link fused aromatic rings and polar functional groups with increased Tmelt.The melting-point predictor is trained on 37,940 training objects and 9,458 test objects, reaching 39.5°C RMS error.
5 Summary
MolecularRNN generates realistic molecular graphs through unsupervised pretraining and achieves valid, property-optimized molecules with reinforcement learning.
- MolecularRNN generates realistic molecular graphs using unsupervised pretraining over diverse molecular distributions.
- 100% valid molecules are generated during inference while invalid molecules still provide negative feedback during training.
- Policy-gradient optimization enables MolecularRNN to generate molecules with desired properties.
- Optimized MolecularRNN outperforms other state-of-the-art methods on benchmark tasks.
- A predictive model acts as a critic for optimizing melting temperature, which cannot be calculated directly from a molecular graph.