Augmenting Genetic Algorithms with Deep Neural Networks for Exploring the Chemical Space

AkshatKumar Nigam; Pascal Friederich; Mario Krenn; Alán Aspuru-Guzik

Source-linked AI summary

Augmenting Genetic Algorithms with Deep Neural Networks for Exploring the Chemical Space

AkshatKumar Nigam, Pascal Friederich, Mario Krenn, Alán Aspuru-Guzik

arXiv:1909.11655v4Sep 25, 2019cs.NEcs.LGphysics.chem-phphysics.comp-ph

TL;DR

Designing organic molecules with desired properties requires efficient exploration of chemical space beyond time-consuming and expensive experimentation. The paper augments a genetic algorithm with a neural-network discriminator to promote diversity and steer search, reporting stronger optimization performance than other generative models and interpretable design rules. Its demonstrations also expose limitations of objective functions and show that reference-distribution mimicry can constrain exploration.

Problem
Designing tailor-made organic molecules requires efficient chemical-space exploration because purely experimental approaches are time consuming and expensive.
Method
The paper augments a genetic algorithm with a neural-network-based adaptive penalty that promotes molecular diversity and steers the search.
Results
The algorithm outperforms other generative models in molecular-design optimization tasks, and analysis of generated molecules yields design rules for high-performing molecules.
Takeaways & Limitations
The approach supports interpretable molecular optimization and is presented as independent of domain knowledge, enabling application beyond chemistry.
Takeaways & Limitations
Highest-performing molecules can exploit deficiencies in the penalized logP objective by generating chemically irrelevant motifs such as sulfur chains.

Abstract

from arXiv · show

Challenges in natural sciences can often be phrased as optimization problems. Machine learning techniques have recently been applied to solve such problems. One example in chemistry is the design of tailor-made organic materials and molecules, which requires efficient methods to explore the chemical space. We present a genetic algorithm (GA) that is enhanced with a neural network (DNN) based discriminator model to improve the diversity of generated molecules and at the same time steer the GA. We show that our algorithm outperforms other generative models in optimization tasks. We furthermore present a way to increase interpretability of genetic algorithms, which helped us to derive design principles.

1 INTRODUCTION

The paper presents a genetic algorithm for molecular design that uses a neural-network adaptive penalty to promote exploration and diversity. Starting from simple methane molecules, the approach outperforms other generative models and does not require domain knowledge.

The proposed molecular-design genetic algorithm adds a neural-network-based adaptive penalty to its search.The paper presents this as the core enhancement to the genetic algorithm.
The adaptive penalty promotes exploratory behavior and increases the diversity of generated molecules.
SELFIES robustness removes the need for expert-based mutation or crossover rules.
Starting from only simple methane molecules, the algorithm outperforms other generative models in molecular-design optimization tasks.
The approach requires no domain knowledge and is therefore applicable beyond chemistry-specific optimization questions.

2 RELATED WORKS

The related work frames molecular inverse design as optimization addressed by VAEs, GANs, reinforcement learning, and genetic algorithms. Existing generative models often mimic reference-data distributions, while genetic algorithms can become trapped in local optima.

Inverse design develops structures with desired properties and has been formulated as an optimization problem in chemistry.
VAEs and GANs generate molecular strings or graphs by encoding discrete representations into a continuous latent space.
Molecular validity remains challenging for latent-space models, motivating follow-up methods operating on strings, graphs, and reinforcement-learning policies.
Generative models trained to mimic reference-data distributions have limited exploration ability.
Genetic algorithms are likely to become trapped in regions of local optima, motivating strategies such as multiple restarts.

3 GA-D ARCHITECTURE

The architecture combines a molecular genetic algorithm with a neural-network discriminator that adaptively penalizes persistent or similar high-fitness molecules, promoting exploration and diversity. Mutations and discriminator-guided selection are used to reduce stagnation and replace long-surviving molecules.

Overview: The generator evaluates each molecule using molecular properties and a discriminator score combined into a single fitness function.The discriminator is trained on GA-generated molecules and a reference dataset at the end of each generation.
Overview: Random mutations of high-fitness molecules replace inferior members, while the best-performing molecules continue into the next generation.Replacement probability is based on a smooth logistic function of fitness rank within the generation.
Mutation & Cross-over Rules: A low mutation rate slows chemical-space exploration and can cause fitness stagnation.SELFIES mutations preserve molecular validity, with insertions and replacements forming the basic mutation operations.
Role of the Discriminator: The discriminator reduces the fitness of persistent high-fitness molecules, increasing their probability of being removed from later generations.Figure 2 illustrates this process for o-terphenyl, which is completely replaced by generation 16.
Role of the Discriminator: The adaptive neural-network penalty assigns similar scores to similarly represented molecules and weakens scores for long-surviving molecules.This reduces their fitness and lowers the chance of prolonged stagnation while forcing exploration of different chemical-space regions.

4 EXPERIMENTS

The experiments optimize the penalized logP objective while comparing the model with literature methods under different exploration scopes. The comparison emphasizes that reference-distribution models and freely exploring genetic algorithms pursue different objectives.

Objective: The optimized objective is penalized logP, combining logP with synthetic-accessibility and ring-size penalties.The ring penalty applies linearly to rings larger than six, while the quantities are normalized using 250,000 ZINC molecules.
Objective: logP measures the water-octanol partition coefficient, while synthetic accessibility discourages chemically infeasible molecules.The objective also penalizes rings larger than six.
Comparison: Table 1 compares maximum penalized logP scores across literature models and GA-based approaches with different exploration scopes.Upper-table models optimize within reference-data distributions, whereas lower-table GA approaches can freely explore chemical space.
Comparison: Direct comparisons require accounting for the differing objectives and scopes of the compared models.The GA parameter β controls the balance between exploration and exploitation.

4.1 UNCONSTRAINED OPTIMIZATION AND COMPARISON WITH OTHER GENERATIVE MODELS

The GA achieves substantially higher molecular-design objective scores than prior methods, while time-dependent adaptive penalties help it recover from stagnation and explore new solutions. However, the highest-scoring structures can exploit deficiencies in the penalized logP objective and be chemically irrelevant.

13.31 ± 0.63 (β = 10) and 12.61 ± 0.81 (β = 0) exceed the highest literature value of 7.87±0.07.These maximum J(m) scores were averaged over 10 independent runs.
The time-dependent adaptive penalty significantly outperforms previous methods in maximizing J(m).During saturation, discriminator weighting is increased from 0 to 1000 until stagnation is overcome.
After steep decreases in maximum J(m), the adaptive-penalty runs recover and can reach values higher than earlier plateaus.Figure 3 compares 10 runs without a discriminator against 5 runs using the time-dependent penalty; solid lines show run averages, shaded regions show run boundaries, and dashed lines show average maxima.
The GA can exploit deficiencies in penalized logP by generating chemically irrelevant motifs such as sulfur chains.This limits the direct application relevance of some highest-performing molecules despite demonstrating efficient chemical-space exploration.

4.3 ANALYSIS OF MOLECULE CLASSES EXPLORED BY THE GA

The GA sequentially explores increasingly diverse molecular classes and identifies structural motifs associated with high objective scores. Clustering and trajectory analysis make these patterns interpretable as design rules.

The GA begins with relatively small molecules, then explores larger classes containing aromatic rings, conjugated carbon chains, and long linear sulfur chains.In the analyzed trajectory, the initial small-molecule phase lasts approximately 40 generations before broader exploration begins.
Class 1 contains the highest J(m) scores found in this work, exceeding 20.The classes were generated by clustering fingerprint vectors of the 50 best-performing molecules from each generation.
Clustering the generated molecules supports design rules for high penalized logP scores based on aromatic rings, conjugated carbon chains, and linear sulfur chains.Representative high-performing molecules were selected by their short distance to cluster centers.
Constructed maximally long molecules score 31.79 for sulfur chains, 8.58 for conjugated carbon chains, and 10.02 for benzene-ring chains with sulfur bridges.The linear sulfur-chain score exceeds the best scores found by the GA and other generative models.
PCA snapshots show sequential exploration of different chemical-space regions and discovery of multiple high-performing molecular families.The trajectories were colored by chemical family and annotated with J(m) scores.

4.4 CONSTRAINED OPTIMIZATION

The GA was evaluated on generating molecules with target properties and improving specific molecules under similarity constraints. It achieved the desired properties in 90.0% of random target-property cases, while the supplied passages do not report the constrained-improvement values.

Generating molecules with specific properties: 90.0% of 250 random target-property cases produced molecules with squared property error below 1.0.Targets varied logP from -5 to 10, SA from 1 to 5, and ring penalty from 0 to 3; runs were limited to 100 generations and 81-character canonical SMILES.
Generating molecules with specific properties: The target-property experiment minimized summed squared differences between actual and desired logP, SA, and ring-penalty values.Each run used a randomly selected set of target properties and counted successful proposals below the specified squared-difference threshold.
Improving penalized logP scores of specific molecules: The constrained-improvement experiment optimized penalized logP for 800 low-scoring ZINC molecules while requiring similarity to each starting molecule.Experiments ran for 20 generations with population size 500 and no discriminator.
Improving penalized logP scores of specific molecules: Success in constrained improvement required increasing penalized logP without violating the similarity constraint.SimilarityPenalty(m) was 0 when sim(m,m′) > δ and −10^6 otherwise.

4.5 SIMULTANEOUS LOGP AND QED OPTIMIZATION

The GA was adapted to optimize logP and drug-likeness simultaneously. It densely sampled the high-value edge of both datasets’ property distributions and generated examples with jointly high scores.

Simultaneous logP and QED optimization: The GA efficiently and densely sampled the edge of the logP and QED distributions in the ZINC and GuacaMol datasets.Because solubility metric logP and drug-likeness cannot be maximized simultaneously, the experiment targeted the boundary of both distributions.
Simultaneous logP and QED optimization: The modified objective function included the drug-likeness score QED alongside logP.The experiment also presented molecules that simultaneously maximize logP and QED.

4.6 MODIFICATION OF THE HYPERPARAMETER β

The hyperparameter β controls the balance between target-property optimization and resemblance to the reference dataset. Its effect is not a smooth interpolation: around β ≈20, the generated distributions transition abruptly toward the reference distribution.

Hyperparameter role: β weights the discriminator score against the molecular target-property score in the fitness function.Large β values favor molecules resembling the reference dataset, whereas small values permit exploration outside its distribution.
Hyperparameter effects: Low β values produce higher average J(m), while very high β values constrain the algorithm toward the reference distribution with mean J(m)=0.The comparison used 11 β values and analyzed both property and discriminator scores.
Hyperparameter effects: At β ≈20, J(m) shifts rapidly from approximately 12-13 toward a distribution centered around 0.The discriminator-score distributions show the same abrupt transition; β ≈18 yields a somewhat wider intermediate distribution that may reflect incomplete convergence.
Hyperparameter effects: β = 0 generates large molecules unlike the reference data, whereas β = 50 generates molecules with more reference-like structures and properties.For β = 50, the J(m) distribution is comparable to the dataset distribution, with mean 0 and unit standard deviation.

5 CONCLUSIONS AND FUTURE WORK

The paper presents a hybrid GA and ML-based generative model for molecular design and uses generated molecules to derive interpretable design rules. It also outlines extensions for expensive property evaluations and broader scientific applications.

Conclusions: The hybrid GA and ML-based generative model outperformed literature approaches in generating molecules with desired properties.The conclusion positions the model as an approach for molecular design.
Conclusions: Analyzing GA-generated molecules enabled interpretation of the model and the derivation of rules for designing high-performing molecules.The authors report that expert-designed molecules inspired by these rules outperformed molecules created by generative models.
Future work: The authors propose an on-the-fly trained ML property evaluator for computationally more expensive property evaluations.They describe this as a future extension for inverse-design problems in chemistry and materials science.
Future work: Because the approach is independent of domain knowledge, the authors plan to generalize GA-D beyond chemistry to broader generative modelling.The stated scope extension targets design questions in other scientific disciplines.

6 SUPPLEMENTARY INFORMATION

The supplementary figures compare property distributions for generated molecules against ZINC and GuacaMol data sets, including random SELFIES and two GA objectives. They also show molecular modifications associated with improved penalized logP scores under similarity constraints.

Molecular optimization examples: Figure S1 presents molecular modifications that produce the largest penalized logP improvements under similarity constraints sim(m, m′) > 0.4 and 0.6.The figure shows examples of molecules optimized in Section 4.4.
Property-distribution comparisons: Figures S2–S4 compare logP, SA, QED, and molecular-weight distributions across generated molecules and the ZINC and GuacaMol data sets.The generated sets include random SELFIES, GA molecules optimized for penalized logP, and GA molecules optimized with logP and QED.
Property-distribution comparisons: Average logP scores of random SELFIES are low, but their distribution extends into a high-value tail.The supplementary comparison passage explicitly notes this distributional pattern.
Property-distribution comparisons: Randomly generated SELFIES are compared with ZINC and GuacaMol molecules across logP, SA, QED, and molecular weight.These comparisons are presented in Figure S2.
Property-distribution comparisons: GA-generated SELFIES with a penalized logP objective are compared with ZINC and GuacaMol molecules across the same four property distributions.These comparisons are presented in Figure S3.
Property-distribution comparisons: GA-generated SELFIES using logP and QED as objectives are compared with ZINC and GuacaMol molecules across logP, SA, QED, and molecular weight.These comparisons are presented in Figure S4.