Introduction
The rapid advancement of AI-driven molecular generation, particularly through diffusion models for drug discovery, has introduced significant challenges for intellectual property (IP) management. Molecular structures currently exist in a regulatory gray zone: they are generally ineligible for copyright, and the cost of patenting early-stage leads is often prohibitive. This leaves high-value AI-generated assets vulnerable to unauthorized reuse and provenance ambiguity.
Traditional digital watermarking techniques designed for images or text are unsuitable for molecular data due to three primary constraints: sparsity (few atoms to hide data), sensitivity (minor structural changes can alter chemical properties), and geometric invariance (molecules are identical regardless of 3D orientation). To address these, we introduce MolMark, the first deep-learning framework designed to embed verifiable ownership signatures directly into 3D molecular structures.
Methodology
MolMark utilizes an encoder-decoder framework that treats watermarking as a learnable, chemistry-aware transformation.
The Encoder-Decoder Architecture
- Encoder ($E_\phi$): The encoder receives a 3D molecule and a binary watermark (e.g., 16 bits). It employs an Edge Embedder to analyze interatomic correlations and a Cross Processing Module with attention mechanisms to distribute the watermark by subtly modulating atom positions. This ensures the structural changes remain within chemically acceptable bounds.
- Decoder ($D_\theta$): To achieve robustness against random rotations and translations (SE(3) invariance), the decoder calculates a Euclidean Distance Matrix (EDM) rather than processing raw coordinates. It then uses Multidimensional Scaling (MDS) to reconstruct a canonical coordinate space for accurate watermark extraction.
Training Strategy
A dynamic, balanced loss function is employed during training. Initially, the model prioritizes the decoder's ability to extract bits. As accuracy improves, the focus shifts to the encoder to minimize structural distortion, ensuring the watermarked molecule remains nearly identical to the original.
Experimental Validation
MolMark was evaluated using the QM9 (134,000 small molecules) and GEOM-DRUG (450,000 drug-like molecules) datasets. The framework was also tested for compatibility with state-of-the-art generative models, including GeoBFN and GeoLDM, to demonstrate its utility in protecting synthetic data.
Results and Discussion
Experimental results indicate that MolMark achieves high capacity and robustness while maintaining chemical integrity.
- High Extraction Accuracy: The system achieves over 95% extraction accuracy even after the molecules undergo random 3D rotations, translations, and reflections.
- Minimal Distortion: The watermark is embedded with a Root-Mean-Square Deviation (RMSD) of less than 0.03 Å. This change is virtually indistinguishable from natural thermal vibration, ensuring the molecule's biological functionality is preserved.
- Property Preservation: Watermarked molecules retained over 90% of their essential properties (stability and validity). Downstream molecular docking tests confirmed that binding affinities remained nearly identical to the original structures.
Comparisons against standard image watermarking baselines (such as HiDDeN) showed that non-specialized models effectively "broke" the molecules, reducing stability to near zero. This highlights the necessity of MolMark's chemistry-first architectural approach.
Conclusion
MolMark represents a significant step forward in securing the intellectual property of AI-generated molecular structures. By balancing high-capacity data embedding with strict chemical constraints, it provides a robust tool for provenance tracking in drug discovery. Future work will explore modulating additional properties like atom types and scaling the framework to handle larger macromolecules such as proteins.
Summary updated