Why Equivariant Diffusion Models Are Reshaping Drug Design

The drug discovery industry has long relied on high-throughput screening and virtual screening to identify promising molecules. Both approaches share a fundamental limitation: they search within existing chemical space. High-throughput screening tests physical compound libraries, typically containing 1-2 million molecules. Virtual screening expands this to billions of enumerated compounds, but the search space remains bounded by known chemical scaffolds and commercially available building blocks.

Generative molecular design promises something fundamentally different: the ability to create novel molecules from scratch, optimized for specific targets and property profiles. Over the past three years, generative models for molecules have evolved rapidly, from SMILES-based variational autoencoders and recurrent neural networks to graph-based models and, most recently, 3D-aware diffusion models.

At Gimerny AI, we have invested heavily in the last category, and the results have validated our thesis. This post explains why equivariant diffusion models represent a step change in generative drug design, and what practical lessons we have learned deploying them at scale.

The Problem with 2D Representations

Most early generative models for molecules operated on SMILES strings or 2D molecular graphs. SMILES-based models treat molecules as sequences of characters, applying the same architectures (RNNs, transformers) that work well for natural language. Graph-based models represent atoms as nodes and bonds as edges, using graph neural networks to encode and decode molecular structures.

Both approaches suffer from a critical blind spot: they ignore 3D geometry. A molecule's biological activity depends not on its 2D topology alone, but on its 3D shape, electrostatic surface, and conformational flexibility. Two molecules with identical 2D graphs can adopt radically different 3D conformations and exhibit opposite biological activities. A kinase inhibitor, for example, must present specific hydrogen bond donors and acceptors at precise spatial positions to complement the ATP binding pocket. No 2D representation captures this requirement.

Enter Equivariant Neural Networks

The key insight behind equivariant models is that molecular properties are invariant to rotation and translation (you can rotate a molecule in space without changing its energy or binding affinity), but the intermediate representations used to compute those properties should transform predictably under these operations. This is the formal definition of equivariance: if you rotate the input, the internal features rotate correspondingly.

SE(3)-equivariant neural networks enforce this symmetry by construction. They use spherical harmonics and tensor products to build features that transform correctly under 3D rotations and translations. This inductive bias means the network does not need to learn rotational invariance from data, dramatically improving sample efficiency and generalization.

Diffusion in 3D Coordinate Space

Diffusion models, originally developed for image generation (DALL-E, Stable Diffusion), learn to reverse a noise-adding process. Starting from pure noise, the model iteratively denoises to produce a sample from the data distribution. For molecules, we apply this process in 3D coordinate space: starting from a random point cloud, the model denoises atom positions and types simultaneously to produce a valid 3D molecular structure.

Our implementation at Gimerny uses a score-based diffusion framework with an SE(3)-equivariant denoising network. The forward process adds Gaussian noise to atom coordinates while simultaneously corrupting atom types through a categorical diffusion process. The reverse process, parameterized by our equivariant network, jointly predicts clean coordinates and atom types at each denoising step.

The critical advantage over autoregressive generation (adding one atom at a time) is that diffusion models generate all atoms simultaneously, naturally capturing long-range spatial correlations. A hydrogen bond donor on one side of the molecule can influence the placement of a hydrophobic group on the opposite side, even if they are separated by many bonds in the graph.

Conditioning on Target Pockets

Unconditional generation produces chemically valid but biologically arbitrary molecules. For drug design, we need molecules optimized for a specific target. We achieve this through conditional diffusion: the denoising network receives the target binding pocket as additional context.

Specifically, we represent the binding pocket as a point cloud of protein atom coordinates with associated features (atom type, partial charge, solvent accessibility). The equivariant network processes ligand and protein atoms through shared message-passing layers, allowing the generated molecule to "see" the pocket it must fill. Cross-attention between ligand and protein features ensures that generated atoms form favorable interactions with pocket residues.

In practice, we condition on multiple objectives simultaneously. Beyond pocket complementarity, we include ADMET property predictors, synthetic accessibility scores, and novelty metrics in the guidance signal. This multi-objective conditioning is implemented through classifier-free guidance, where we train the model with and without conditioning and interpolate at inference time.

Practical Results

We have generated over 2.8 million molecules using this approach across 45+ client projects. Key observations from deployment include the following.

First, synthesizability is critical. Early generative models produced beautiful molecules that could not be made. Our retrosynthetic-aware scoring, integrated into the generation loop via GimernySynth, ensures 94% of generated molecules have feasible synthesis routes of 5 steps or fewer.

Second, diversity matters as much as quality. Generating 1,000 copies of the same optimal molecule is useless. We use a diversity-promoting loss term that penalizes batch-level similarity in both chemical and shape space, ensuring each generation run produces a structurally diverse set of candidates.

Third, 3D matters. In head-to-head comparisons on 12 target-based design tasks, our 3D diffusion model outperformed the best 2D baseline (a graph transformer VAE) by 18% in docking score and 23% in binding free energy, while maintaining comparable synthetic accessibility.

What Comes Next

The field is moving toward several exciting directions. Multi-resolution generation, where coarse-grained pharmacophore patterns are generated first and atomistic details filled in subsequently, promises to bridge the gap between medicinal chemistry intuition and machine-learned design. Protein-ligand co-diffusion, where both the protein and ligand conformations are generated jointly, could unlock allosteric and cryptic binding sites that rigid-pocket methods miss.

At Gimerny, we are also exploring reinforcement learning from wet-lab feedback: using experimental binding assay results to fine-tune the diffusion model in a closed loop. Early results suggest that just 50-100 experimental data points can substantially improve hit rates for a given target, making the human-AI collaboration cycle tighter and more efficient.

The era of searching for drugs is ending. The era of designing them has begun.