Pesti-Gen: Unleashing a Generative Molecule Approach for Toxicity Aware Pesticide Design

Taehan Kim, Wonduk SeoarXiv 2025 · DOI · arXiv

TL;DR

This paper introduces Pesti-Gen, a novel generative AI model that designs new pesticide molecules with optimized toxicity profiles. By combining a general molecular structure representation with specific livestock and aquatic toxicity data, Pesti-Gen aims to accelerate the development of safer and more sustainable pest management solutions.

What Problem Does This Paper Address?

Global agriculture faces a critical challenge: increasing crop resilience and pesticide efficacy are hampered by climate change, leading to an over-reliance on synthetic pesticides. While essential for pest management, these chemicals pose significant health and environmental risks. Existing machine learning approaches in this domain have primarily focused on classifying or predicting properties of existing pesticides, leaving a crucial gap in the generation of entirely new, safer molecular structures. This paper tackles this gap by proposing a generative approach to design pesticides that are not only effective but also environmentally conscious.

Methods

Pesti-Gen employs a two-stage generative approach based on Variational Autoencoders (VAEs).

  1. Latent Space Training: The model is first pre-trained on a large dataset of general molecular structures (from the ZINC dataset) to learn a comprehensive representation of chemical features in a latent space. This stage focuses on accurately encoding and decoding molecular structures (SMILES strings) and creating a smooth, meaningful latent embedding.
  2. Fine-Tuning on Target Properties: In the second stage, the pre-trained encoder is frozen, and the decoder is fine-tuned using specialized toxicity metrics. This includes incorporating both livestock toxicity (LD50) and aquatic ecotoxicity (LC50) data. A modified loss function is used, which includes reconstruction loss, KL divergence, and weighted terms for livestock and aquatic toxicity, to guide the generation towards molecules with optimized safety profiles. Latent noise injection is also used to improve robustness and generalization.

Datasets

  • Training (Latent Space): ZINC dataset (50,000 diverse molecular structures).
  • Training/Evaluation (Toxicity Fine-tuning): A custom dataset compiled from the Gyeonggi Province Pesticide Open Dataset. This dataset includes SMILES representations and toxicity grades (LD50 for livestock, LC50 for aquatic) for 56,360 data points and 520 unique pesticides. This dataset was curated to address the lack of publicly available benchmarks for generative pesticide design, specifically incorporating both livestock and aquatic toxicity metrics.

Key Insights and Ideas

  • Generative Design for Pesticides: The paper demonstrates that generative AI, successfully applied in drug discovery, can be effectively adapted for designing safer pesticides, moving beyond mere prediction.
  • Multi-Objective Optimization for Safety: By simultaneously optimizing for both livestock and aquatic toxicity, Pesti-Gen offers a more holistic approach to designing environmentally friendly pesticides.
  • Bridging Drug Discovery and Agriculture: The work highlights the potential of leveraging advanced computational techniques from one domain (drug discovery) to solve critical problems in another (agricultural sustainability).
  • Importance of Custom Datasets: The creation of a specialized dataset with detailed toxicity metrics was crucial for training and validating the model, underscoring the need for domain-specific benchmarks.

Key Results

  • Structural Validity: Pesti-Gen achieved an approximate 68% structural validity rate in generating new molecular structures, indicating a good balance between novelty and chemical feasibility.
  • Property Distribution: Generated molecules exhibited a slightly broader LogP range, suggesting the model can explore a wider chemical space.
  • Synthetic Accessibility: The model generated molecules with a slightly lower mean Synthetic Accessibility Score (SAS), indicating improved synthetic feasibility on average.
  • Qualitative Improvement: Structural analysis showed that reconstructed molecules retained critical functional groups while incorporating subtle modifications to improve biodegradability and reduce environmental persistence.

Limitations and Future Work

  • Dataset Limitations: Reliance on publicly accessible, license-free datasets limits the model's ability to learn nuanced pesticide-specific knowledge. Predictions still require wet-lab validation.
  • Single-Objective Optimization: While the multi-objective approach is versatile, more focused single-objective optimization could further refine specific toxicity profiles.
  • Cross-Resistance Tracking: Incorporating cross-resistance tracking could aid agronomists in developing effective rotation strategies.
  • Future Directions: The authors suggest that incorporating more specialized datasets or objectives, and leveraging the detailed geographic coordinates in their custom dataset for environment-specific analysis, could lead to further improvements in pesticide design and sustainable agricultural practices.

Summary updated

BibTeX
@misc{kim2025pesti,
  author = {Taehan Kim and Wonduk Seo},
  title = {{Pesti-Gen: Unleashing a Generative Molecule Approach for Toxicity Aware Pesticide Design}},
  year = {2025},
  eprint = {2501.14469},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2501.14469},
}
Figure