Sesame: Structure-Aware Molecular Generation via Spatial Density-Map Conditioning

Abstract: Generative molecular models for drug design are a promising direction with much active research. In the next phase of computational drug design, such models will need to understand small molecule structure and protein-ligand interactions, and they will need to possess the machinery to generate molecules de novo. Incorporating each feature poses a critical challenge. Equally important, yet often treated as secondary, is the ability to grow a molecule from a partial starting point -- a scaffold or fragment supplied by a chemist -- which is the central operation of lead optimization. We present Sesame (Spatial Evoformer for a Structure-Aware Molecular Engine), a diffusion-based molecular generation model that leverages a novel spatial pairformer module to condition on partial molecular structure and the surrounding protein pocket, both expressed as continuous spatial density maps. This single conditioning mechanism supports both de novo generation and fragment-conditioned lead optimization, letting a medicinal chemist prune a hit to a scaffold and have Sesame grow it in productive ways. In addition to this module, we also introduce a diffusion framework for joint denoising of atom types, bond types, and positions, along with a trajectory finetuning scheme that trains on the model's own sampling rollouts to improve generation quality. Sesame is trained on a large corpus of ligand-only and protein-ligand datasets.

• The paper introduces Sesame, a framework using spatial density-map conditioning to integrate protein pocket and fragment information for unified molecular generation.
• It employs a novel MoleculePairformer architecture with a hybrid discrete-continuous diffusion process and trajectory-based finetuning for robust structure-aware design.
• Empirical results demonstrate 94.8% scaffold retention and high molecule validity, highlighting its practical applicability in drug discovery.

Structure-Aware Molecular Generation with Spatial Density-Map Conditioning

Overview

"Sesame: Structure-Aware Molecular Generation via Spatial Density-Map Conditioning" (2606.23856) introduces a generative molecular modeling framework specifically designed to address practical drug discovery scenarios. The system leverages spatial density-map conditioning, enabling both de novo molecular generation and fragment-based lead optimization through a unified input interface. The architecture is centered around a novel Pairformer variant, integrating continuous spatial information from protein pockets and molecular fragments, augmented by a hybrid discrete-continuous diffusion process and a trajectory-based fine-tuning scheme.

Motivation and Problem Framing

Generative models for molecular design have shown increasing utility in drug discovery but often lack structural awareness necessary for real-world lead optimization. Conventional paradigms tend to treat fragment-based growth as a secondary task, constraining architectures to either purely de novo generation or rigid, atom-level constraints for fragment conditioning. Sesame directly addresses this gap by employing spatial density maps for both protein pockets and molecular fragments, enabling flexible, structure-aware conditioning without architectural changes. This approach allows medicinal chemists to incorporate domain knowledge by pruning molecules to scaffolds and seeding generative completions compatible with a target binding site.

Model Architecture and Key Innovations

Density Map Conditioning

Sesame uses continuous spatial density map representations to encode protein pocket and fragment information. Each density map encompasses multiple physical parameters (charge, hydrophobicity, hydrogen bonding, van der Waals interactions, aromaticity). The architecture employs a novel density map conditioning module: attention-based sampling of query points from molecular representations, interpolated with trilinear grid sampling from the density maps, and updated through cross-attention. This mechanism tightly couples the spatial context to the generative process, supporting both fragment-based and de novo molecular generation.

MoleculePairformer Core

The core model, MoleculePairformer, operates on single and pair representations—mirroring the AlphaFold Evoformer/Pairformer lineage—with extensive embedding layers for atom types, spatial coordinates, bond types, and diffusion timesteps. The architecture features:

• Triangle multiplication, triangle attention, and pair-biased attention modules, consistent with state-of-the-art structural prediction architectures.
• Cross-attention between learned anchor points (sampling queries) and processed density map volumes, updating molecular representations via gated residuals.

Output heads jointly predict atom types, coordinates, and bonds, accommodating variable atom counts through a pseudo-None atom type and ensuring symmetry in bond matrices.

Hybrid Diffusion Process

Sesame introduces a hybrid discrete-continuous diffusion framework, jointly modeling atom types (categorical diffusion), bond types (categorical diffusion), and atomic coordinates (Gaussian diffusion), each with schedule-agnostic corruption rates. The forward process applies independent noise to each modality, followed by permutation matching for coordinates, and marginal discrete diffusion for atom/bond types. The reverse process utilizes trajectory-level likelihood maximization for joint denoising.

Trajectory finetuning incorporates self-distillation from model rollouts: recorded denoising steps are used as fine-tuning datapoints, aligned via permutation matching to address common slot-permutation errors.

Data Pipeline and Training Regime

Sesame is trained on two primary datasets:

• ZINC22: Extensive ligand-only dataset with RDKit-generated conformers and MMFF94-derived density maps, filtered to 3–50 heavy atoms.
• SAIR: Synthetic protein-ligand complexes, processed for binding pocket extraction and density map calculation.

Training leverages heavy data augmentation, including rotational invariance (SO(3) QR sampling), fragment scaffolding (RDKit Murcko, random atom addition/removal), and density map dropout for robustness against missing or noisy conditioning information.

Loss functions comprise categorical cross-entropy (with label smoothing), position MSE, and an LDDT-inspired multi-atom distance metric, combined without weights.

Optimization uses AdamW with notable high weight decay (0.1) and linear/cosine learning rate scheduling.

Empirical Validation

Validation across six conditioning modes (fully conditioned, protein + fragment, protein-only, ligand-only, fragment-only, unconditioned) demonstrates Sesame's capacity to honor fragment-based constraints and leverage pocket information. Notably:

• Fragment retention: 94.8% of molecules generated under fragment conditioning regime retain the seeded scaffold as a substructure.
• Trajectory-level molecule validity: With simple post-processing to remove singleton atoms, molecule validity for protein + fragment and protein-only modes reach 92.4% and 88.7% respectively after finetuning.
• Drug-like properties: Generated molecular distributions closely mirror those from real SAIR ligands and largely conform to Lipinski Rule of 5, with fragment-conditioned generation yielding slightly more drug-like molecules.

Sesame further explores optimal schedules for the denoising trajectory in noise space, finding schedules that prioritize bond denoising early are favored, plausibly due to bond sparsity and structural guidance.

Practical and Theoretical Implications

Sesame's density-map conditioning allows contiguous integration of human-in-the-loop approaches with generative modeling. The ability to treat fragments as soft spatial priors makes lead optimization and hit elaboration a central workflow, not a special case. The trajectory finetuning strategy and hybrid diffusion modeling directly address slot permutation issues and schedule sensitivity—theoretically advancing the robustness and flexibility of generative chemistry models.

Practically, Sesame's framework sets a foundation for workflows where medicinal chemists seed fragments and rely on the model to elaborate them into synthetically accessible, pocket-compatible molecules. Such approaches may enable rapid iteration in lead optimization, balancing exploration and exploitation in chemical space more effectively than prior systems.

Speculation on Future Developments

Future directions will likely extend Sesame’s composability, integrating downstream synthesizability filters or reaction network decoders. Further architectural refinements may improve variable-atom count stability or address excess singleton atom errors. Adapting models to exploit richer physical property maps, possibly leveraging higher-resolution protein-ligand datasets or multi-channel field conditioning, could push structural awareness further. Integration with flow-matching approaches for improved sampling efficiency and physical realism, as well as expanded fragment-conditioned design strategies, are anticipated.

Conclusion

Sesame presents a robust structure-aware generative framework for molecular design, unifying fragment-based lead optimization and de novo synthesis through spatial density-map conditioning. The architecture's core innovations—attention-based density map sampling, hybrid discrete-continuous diffusion, and trajectory-based finetuning—enable practical, theory-aligned workflows in computational drug discovery. Empirical results confirm reliable fragment retention and molecule validity, supporting both structural insight-driven human workflows and automated exploration in chemically relevant spaces.