KAN-PROSPECT: A Natural Product Pharmacological Effect Prediction Framework Integrating Graph Neural Networks and Kolmogorov-Arnold Networks

KAN-PROSPECT is an AI-driven framework combining Graph Neural Networks (GNN), Kolmogorov-Arnold Networks (KAN), and transfer learning, designed for large-scale prediction of the pharmacological effects and adverse reactions of natural products. Addressing the issues of scarce natural product data and insufficient generalization ability of traditional methods, this framework enhances robustness and generalization through technical integration, providing a new tool for computational drug discovery.

Natural products are important sources for new drug development, but traditional screening methods face challenges such as scarce data, high costs, and long cycles (averaging 10-15 years and costing billions of dollars). Many natural products lack experimental data, making it difficult for traditional machine learning models to generalize. Additionally, natural products have complex components, and the interaction between their adverse drug reactions (ADR) and human metabolism is hard to describe with linear models; accurate prediction under data constraints has become a key challenge.

The core innovation of KAN-PROSPECT lies in the integration of three technologies:

1. Graph Neural Networks (GNN)：Directly process molecular graph structures, capture topological relationships and chemical properties, and are more expressive than traditional fingerprint/descriptor methods.
2. Kolmogorov-Arnold Networks (KAN)：Use learnable univariate activation functions to fit complex nonlinear relationships with fewer parameters, avoiding the "flat region" problem of traditional MLPs.
3. Transfer Learning：First pre-trained on large chemical databases (e.g., ChEMBL, PubChem), then transferred to natural product tasks to mitigate the risk of overfitting on small samples.

The project is developed based on Python 3.9.20, with core dependencies including:

• RDKit (2024.03.2): Molecular structure parsing and feature extraction
• NetworkX (3.2.1): Molecular graph construction and analysis
• scikit-learn (1.5.2): Data preprocessing and evaluation
• NumPy/SciPy/Matplotlib: Scientific computing and visualization Mature open-source tools are selected to lower deployment barriers and ensure efficiency.

The application value of KAN-PROSPECT includes:

• Drug Repurposing: Predict new indications for known natural products, accelerating the repurposing of existing drugs.
• Early Toxicity Warning: Predict adverse reactions before synthesis, reducing the risk of clinical trial failure.
• Natural Product Screening: Quickly screen high-potential candidate molecules to guide experimental priorities.
• Mechanism Research Assistance: Visualize through attention mechanisms to help understand structure-activity relationships.

Limitations: Currently relies on public datasets for validation; integration with real R&D processes needs exploration; although interpretability is better than black-box models, it does not meet regulatory transparency standards. Outlook: Integrate multi-omics data for multimodal prediction; introduce physicochemical constraints to improve reliability; develop an end-to-end experimental design system connecting AI prediction and automated experimental platforms.