Source-linked AI summary
Mol-CycleGAN - a generative model for molecular optimization
Łukasz Maziarka, Agnieszka Pocha, Jan Kaczmarczyk, Krzysztof Rataj, Michał Warchoł
TL;DR
Drug development requires optimizing molecules across multiple complex properties while navigating high-dimensional chemical space. Mol-CycleGAN uses a CycleGAN-based molecular model to generate structurally similar compounds with optimized properties. In constrained penalized-logP optimization of drug-like molecules, it significantly outperforms previous results while retaining controllable similarity to starting molecules.
Problem
Optimizing drug-like molecules across multiple properties in high-dimensional chemical space remains challenging.
Method
Mol-CycleGAN extends CycleGAN to molecular embeddings and learns transformations between compounds with undesired and desired molecular properties.
Results
In constrained optimization of drug-like molecules, Mol-CycleGAN significantly outperforms previous results while generated molecules remain close to starting molecules with similarity controlled by a hyperparameter.
Takeaways & Limitations
The model supports constrained molecular optimization by combining desired-property improvement with structural similarity to the starting compound.
Takeaways & Limitations
After many iterations of unconstrained optimization, the molecule distribution can move away from the starting distribution.
Abstract
from arXiv · showhide
Designing a molecule with desired properties is one of the biggest challenges in drug development, as it requires optimization of chemical compound structures with respect to many complex properties. To augment the compound design process we introduce Mol-CycleGAN - a CycleGAN-based model that generates optimized compounds with high structural similarity to the original ones. Namely, given a molecule our model generates a structurally similar one with an optimized value of the considered property. We evaluate the performance of the model on selected optimization objectives related to structural properties (presence of halogen groups, number of aromatic rings) and to a physicochemical property (penalized logP). In the task of optimization of penalized logP of drug-like molecules our model significantly outperforms previous results.
Introduction
Mol-CycleGAN addresses multiparameter molecular optimization by generating structurally similar compounds with desired properties. It uses graph-based CycleGAN modeling to preserve validity and similarity while optimizing structural or physicochemical objectives.
- Related approaches: Existing molecule-generation approaches include virtual screening, de novo design, VAEs, GANs, and reinforcement-learning models.These approaches use representations including SMILES, molecular graphs, adjacency matrices, and graph-convolutional policy networks.
- Motivation: Multiparameter optimization of drug-like molecules remains a major challenge in high-dimensional chemical space.The introduction frames hit-to-lead and lead optimization as requiring simultaneous optimization of activity and physicochemical or pharmacokinetic properties.
- Limitation: Generated compounds from prior approaches can be difficult or impossible to synthesize, limiting their practical utility.The paper identifies synthetic accessibility as a major obstacle in practical molecule-generation approaches.
- Contribution: Mol-CycleGAN generates a structurally similar molecule with desired characteristics from a starting molecule.Structural similarity is intended to facilitate synthesis and reduce the likelihood of disrupting other optimized properties during multiparameter optimization.
- Contribution: Graph-based molecular representation enables the algorithm to always return valid compounds.This contrasts with earlier SMILES-based models that could generate invalid strings not corresponding to molecules.
- Evaluation: The model is evaluated on halogen-group presence, aromatic-ring count, and penalized logP optimization.For drug-like molecules, the penalized-logP task is reported to significantly outperform previous results.
Methods
Mol-CycleGAN extends CycleGAN to JT-VAE molecular latent embeddings, learning mappings between molecule sets with selected property values while preserving structural similarity. Its losses combine adversarial, cycle-consistency, and identity objectives for property optimization.
- Model representation: Mol-CycleGAN extends CycleGAN to molecular embeddings in the JT-VAE latent space.Each molecule is represented as the mean of its variational encoding distribution.
- Model architecture: The model learns mappings between sets X and Y containing molecules with and without a selected property, using generators G and F plus discriminators DX and DY.The generators are trained to produce samples close to the target-set distributions.
- Optimization workflow: Given an input molecule from X, G generates a Y-like latent embedding, which is decoded into a structurally similar optimized molecule.This design targets lead optimization by retaining similarity to the starting compound.
- Training objective: Cycle consistency regularizes the mappings, supports learning from both transformation directions, and may help in low-data settings.The identity loss further prevents generated molecules from lying far from the starting molecule in latent space.
- Evaluation: Experiments evaluate structural modifications and molecule optimization on ZINC-250K, including aromatic-ring, halogen-group, and penalized-logP objectives.Penalized logP is defined as logP minus synthetic accessibility, and the model is assessed for both property improvement and similarity.
Results
Mol-CycleGAN performs structural and constrained penalized-logP optimization while preserving molecular similarity and drug-likeness. Aromatic-ring transformations are harder than halogen transformations, while unconstrained optimization eventually sacrifices drug-likeness and similarity.
- Structural transformations: Structural transformations modify halogen presence or aromatic-ring counts, with aromatic-ring changes reported as more difficult.The aromatic-ring task involves transitions between molecules with different ring counts, whereas halogen transformations are simpler.
- Structural transformations: Higher success rates occur for halogen-moiety transformations, partly because 64.9% of ZINC-250K molecules lack halogen moieties.The reported imbalance may explain the higher success rate for removing halogen moieties.
- Similarity and drug-likeness: Molecular similarity and drug-likeness are achieved across all experiments.Tanimoto-similarity analyses indicate the procedure is not random sampling or latent-space memorization.
- Constrained molecule optimization: Constrained optimization maximizes penalized logP subject to a Tanimoto-similarity threshold, selecting the highest-scoring distinct decoded molecule that satisfies the constraint.The procedure samples K = 80 equally spaced points along the latent-space path and reports the best valid molecule.
- Constrained molecule optimization: Mol-CycleGAN significantly outperforms previous results in mean penalized-logP improvement while achieving comparable mean similarity for δ > 0.Success rates are comparable for δ = 0 and 0.2, whereas the model has lower success rates under the more stringent constraints δ = 0.4 and 0.6.
Conclusions
Mol-CycleGAN introduces a CycleGAN-based approach that learns property-directed transformations in the JT-VAE latent space. It generates molecules close to their starting structures, achieves the desired properties, and outperforms previous results in constrained optimization, while activity cliffs remain a future challenge.
- Conclusions: Mol-CycleGAN is presented as a new CycleGAN-based model for molecule generation and optimization.The model operates in the latent space learned by JT-VAE.
- Conclusions: The model learns transformation rules from compounds with desired and undesired property values.Its use is demonstrated on structural and physicochemical properties.
- Conclusions: Generated molecules remain close to starting molecules, with similarity controllable through a hyperparameter.The approach is designed for property optimization while retaining structural similarity.
- Conclusions: In constrained optimization of drug-like molecules, Mol-CycleGAN significantly outperforms previous results.The paper identifies constrained molecule optimization as a principal demonstrated outcome.
- Conclusions: Future work includes multi-parameter optimization and cases where small structural changes produce drastic property changes, known as activity cliffs.Activity cliffs are described as hard to model.
Supporting Information Available
The paper makes its implementation materials available online.
- Supporting Information Available: All code used to produce the reported results is available online at the project repository.The repository is https://github.com/ardigen/mol-cycle-gan.