DD3S: A Privacy-Preserving Artificial Intelligence Platform for End-to-End Early-Stage Drug Discovery

The drug discovery and development process remains one of the most resource-intensive endeavours in modern science, typically requiring 10–15 years and investments exceeding $2–3 billion per approved therapeutic agent^[1]. High attrition rates — often exceeding 90% from lead identification to market approval — stem from challenges in target validation, compound optimization, toxicity prediction, and pharmacokinetic profiling. Traditional high-throughput screening and empirical methods are increasingly inadequate to explore the vast chemical space (estimated at 10⁶⁰ possible small molecules), driving the adoption of computational approaches to accelerate early-stage discovery.

In recent years, artificial intelligence and machine learning have transformed cheminformatics and virtual screening paradigms. Graph Neural Networks, generative models (such as variational autoencoders, generative adversarial networks, and diffusion models), and large language models enable rapid property prediction, de novo molecular design, and Absorption, Distribution, Metabolism, Excretion and Toxicity assessment^[3,7,15]. Platforms integrating these techniques have demonstrated success in hit identification and lead optimization. However, several persistent limitations hinder widespread adoption in academic and industrial settings.

Fragmentation of tools remains a key challenge, as most existing solutions separate molecular sketching, three-dimensional visualization, property prediction, literature search, and explanatory analysis across multiple disconnected applications or cloud-based services. Data privacy and security concerns also persist, since reliance on external application programming interfaces or cloud-hosted models exposes proprietary molecular structures and research data to potential breaches — a critical issue in pharmaceutical research and development where intellectual property is paramount. Furthermore, limited interactivity and usability characterize current platforms, as few offer real-time HTML5 drawing synchronized with dual three-dimensional viewers, integrated local large language model chat for toxicity and risk explanation, or educational interfaces tailored for students and researchers. Finally, ethical and citation shortcomings are prevalent, as online literature retrieval often relies on increasingly restrictive commercial application programming interfaces or lacks proper attribution, risking academic integrity.

To address these gaps, we present DD3S (AI-Powered Drug Discovery System), a fully local, privacy-preserving, web-based platform that unifies interactive molecular design, high-fidelity visualization, accurate property prediction, and AI-assisted analysis within a single browser environment. DD3S leverages RDKit for robust cheminformatics and SMILES-to-graph conversion; an augmented Graph Neural Network (Add-GNN) fusing structural embeddings with physicochemical descriptors for toxicity, ESOL solubility, and bioactivity prediction; dual three-dimensional rendering synchronizing an HTML5 two-dimensional canvas with 3Dmol.js and Mol*; locally hosted large language models (Gemma-3:4b and Gemma-2:2b); a preloaded database of 14,197 molecules; and a Semantic Scholar-backed search engine for ethical citations and AI-generated summaries. By running all inference locally, DD3S guarantees 100% data privacy while delivering sub-2-second latencies for most operations.

MATERIALS AND METHODS

Molecular representation: Molecules are represented via Simplified Molecular Input Line Entry System (SMILES) notation. RDKit (version 2023.09.1; RDKit Open-Source Cheminformatics, rdkit.org) was employed to convert these strings into graph structures where atoms function as nodes and bonds as edges.

Property prediction (Add-GNN): A Graph Neural Network was utilized to predict toxicity, solubility (ESOL), and bioactivity. The model achieves superior performance by fusing graph embeddings with physicochemical descriptors including LogP, topological polar surface area, and molecular weight. Model performance was assessed using mean absolute error (MAE) as the primary metric; results are reported across the full validation dataset (n = 14,197 molecules).

High-fidelity visualization with Mol*: Mol* (version 3.x; PDBe, EMBL-EBI) was integrated for web-based three-dimensional visualization, enabling rendering of complex protein–ligand interactions retrieved from the RCSB Protein Data Bank, with support for high-performance WebGL graphics even for large macromolecules. For responsive embedded views, 3Dmol.js (version 2.x; University of Pittsburgh) provides lightweight three-dimensional rendering.

Interactive web interface: The platform comprises four core pages built with Bootstrap 5, Django 4.2 (Django Software Foundation), and JavaScript libraries. The Home/Dashboard page showcases platform features and an AI Analysis Tool using Gemma-2:2b for dynamic insights across the 14,197 preloaded molecules. The Search page queries the open Semantic Scholar Graph API alongside PubMed and Europe PMC, ensuring open-access literature retrieval with full ethical citations and AI-generated summaries via Gemma-3:4b. The Design page features an interactive HTML5 canvas for two-dimensional molecular sketching, with real-time SMILES generation driving a lightweight three-dimensional WebGL render via 3Dmol.js and a secondary full-screen Mol* mode. The Analyzer page computes molecular descriptors, two-dimensional and three-dimensional views, solubility, toxicity filters, and Lipinski’s Rule of Five compliance upon entry of a SMILES string. All pages maintain sub-2-second response times with local inference only.

Statistical analysis: System performance metrics including SMILES validation accuracy and prediction latency were measured across the full dataset of 14,197 preloaded molecules and on standard reference compounds including Aspirin and Caffeine. Validation accuracy is reported as the percentage of correctly processed SMILES strings. Latency values represent mean response time across ten repeated measurements under standard consumer hardware conditions.

RESULTS AND DISCUSSION

The system was validated on standard molecules, including Aspirin and Caffeine, as well as the full dataset of 14,197 molecules. The compound details view is shown in Figure 1, which demonstrates the integration of physicochemical properties, solubility prediction, Lipinski’s Rule of Five validation, and toxicity analysis within a unified interface. The interactive molecular design interface is illustrated in Figure 2, highlighting the real-time synchronization between the two-dimensional sketching canvas and the three-dimensional visualization.

The advanced visualization capabilities are further supported by the Mol* viewer (Figure 3), which provides high-fidelity structural rendering, and the 3Dmol.js viewer (Figure 4), which enables lightweight, real-time WebGL-based interaction. The dual-viewer approach ensures smooth performance and flexibility, even on standard consumer hardware.

Figure 1. Compound details view showing comprehensive profile displaying three-dimensional and two-dimensional molecular structures, basic physicochemical information, predicted solubility, Lipinski’s Rule of Five compliance, and integrated AI assistant.

Figure 2. Interactive molecular design interface featuring a real-time two-dimensional sketching canvas with synchronized three-dimensional viewer and integrated search bar.

Figure 3. Mol* viewer showing interactive molecular visualization displaying the current structure in high-quality three-dimensional rendering with support for orbiting, zooming, and residue selection.

Figure 4. 3Dmol.js showing real-time interactive three-dimensional display of molecules with smooth WebGL rendering, multiple visualization styles, and seamless synchronization with the design canvas.