Molecular generation is fundamentally a geometric problem. Molecules are 3D objects with spatial symmetries: rotating or translating a molecule in space does not change its properties. Any neural network that generates molecules must respect these symmetries to produce physically valid structures. This requirement led us to build NutraDiscover's core generative engine on equivariant graph neural networks (EGNNs).
Traditional approaches to molecular generation operate on SMILES strings or molecular fingerprints, which are 1D representations that discard 3D structural information. While effective for simple property prediction, these representations struggle with tasks that depend on molecular shape: binding pose prediction, conformational analysis, and scaffold hopping across structurally diverse chemotypes.
EGNNs operate directly on molecular graphs embedded in 3D space. Each atom is represented as a node with position coordinates and feature vectors; bonds are edges with geometric attributes. The key property is equivariance: if the input molecule is rotated, the network's internal representations rotate accordingly, and the output predictions transform consistently. This is achieved through carefully designed message-passing layers that separate scalar (invariant) and vector (equivariant) features.
Our implementation extends the E(n) equivariant framework with several innovations. First, we use a hierarchical graph structure that represents molecules at both atomic and fragment levels, enabling the model to capture long-range interactions without excessive computational cost. Second, we introduce a novel attention mechanism that weights inter-atomic interactions based on both distance and chemical context. Third, we incorporate a differentiable force field regularizer that penalizes generated structures with physically unrealistic bond lengths, angles, or steric clashes.
Training the model required substantial computational resources: 256 NVIDIA A100 GPUs for 12 days on a dataset of 180 million 3D conformers generated from ChEMBL compounds using RDKit. We used a denoising diffusion objective, training the model to reconstruct molecular structures from progressively corrupted versions. At inference time, molecules are generated by iteratively denoising random noise into valid 3D structures.
Benchmarking against alternatives reveals clear advantages. On the GuacaMol benchmark suite, our EGNN-based generator outperforms SMILES-based VAEs on novelty (87% vs. 72%) and uniqueness (99.1% vs. 94.3%) while maintaining comparable validity (96.8% vs. 97.2%). On 3D-aware metrics (shape similarity, pharmacophore matching), the advantage is more pronounced: our method achieves 0.78 shape Tanimoto vs. 0.51 for SMILES-based approaches on a prospective drug design task.
The computational cost is the primary tradeoff. Generating a single molecule with our EGNN takes approximately 2.3 seconds on an H100, compared to milliseconds for SMILES-based methods. However, the quality improvement means fewer generated candidates need experimental testing, resulting in net time and cost savings. For a typical hit-finding campaign, we estimate that the higher quality of EGNN-generated candidates reduces required wet-lab validation by 5-8x.
Looking ahead, we are exploring two extensions: conditional generation guided by protein pocket geometry (enabling true structure-based drug design) and multi-objective optimization that simultaneously optimizes potency, selectivity, and ADMET properties. Both are active areas of research, and we plan to release updated benchmarks in Q4 2026.