Hierarchical generative modeling for the design of multi-component systems
Rhyan Barrett, Robin Curth, Julia Westermayr · 2026
Read Hierarchical generative modeling for the design of multi-component systems by Rhyan Barrett, Robin Curth, Julia Westermayr (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2604-12607.txt.
Introduction
Rhyan Barrett and two colleagues design multi-part chemical systems. Catalysts and enzymes work through interplay of parts, not lone molecules. Brute-force search over mixes and layouts is infeasible. Their plan pairs a genetic search with a generative molecule model. Shape and mix refine in a closed loop. A test case builds catalytic surroundings for a Claisen rearrangement. Checks show a 30 percent cut in activation barrier. The plan extends to catalysts, enzyme sites, and new materials.
Catalog record
| CREATOR | Rhyan Barrett, Robin Curth, Julia Westermayr |
|---|---|
| YEAR | 2026 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 58840 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2604.12607
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
Hierarchical generative modeling for the design of multi-component systems Rhyan Barrett, Robin Curth, Julia Westermayr Abstract The functionality of catalysts, enzymes, and supramolecular assemblies emerges not from individual molecules alone, but from the subtle interplay between multiple components arranged in complex systems. Designing such systems is a grand challenge, the combinatorial explosion of possible chemical compositions and spatial arrangements makes brute-force exploration infeasible, while many current generative approaches remain limited to isolated molecules. In this work, we introduce a hierarchical generative optimization framework that overcomes this barrier by coupling a genetic algorithm for configurational search with a generative model for molecular design. This closed-loop approach enables simultaneous refinement of geometry and composition, efficiently steering discovery toward systems with targeted functionality. As a proof of concept, we design catalytic environments for the Claisen rearrangement of p-tolyl ether by optimizing surrounding components around a fixed reference transition-state geometry. Despite this constraint during the search phase, post-hoc validation via Climbing-Image Nudged Elastic Band calculations confirm a 30% reduction in activation barrier. Beyond this example, our framework provides a general strategy for data-driven discovery of functional multi-component systems, opening the door to automated design of catalysts, enzyme active sites, and advanced materials. I Introduction A central challenge in molecular discovery is the efficient exploration of chemical space to identify molecules and molecular systems with desired properties 46 ; 49 ; 55 . In cheminformatics, computational methods have become indispensable for this task by enabling virtual screening, property prediction, and data driven prioritization of candidate structures before experimental testing 42 ; 19 ; 20 . However, the number of synthetically and functionally accessible compounds remains enormous, and structure property relationships are often highly non-linear, making exhaustive evaluation of predefined libraries increasingly impractical as molecular complexity grows 61 ; 54 ; 49 . As a result, many current workflows still rely on screening based approaches that rank existing candidates rather than proposing fundamentally new ones 19 ; 20 . This limitation has motivated growing interest in inverse design strategies, particularly generative models that learn molecular representations from data and propose previously unseen structures tailored to target properties 18 ; 36 ; 15 . Generative machine learning models offer a powerful route to inverse design by learning statistical patterns in molecular data and proposing novel structures beyond those observed during training 50 ; 51 . They can be trained on simple low-dimensional descriptors such as SMILES strings 26 ; 6 or on three-dimensional molecular structures. Prominent examples include autoregressive models such as G-SchNet 16 ; 17 and diffusion-based approaches 9 ; 27 , which can generate molecular structures directly. When combined with supervised models that capture relationships between chemical structure and target properties, generative models can be steered toward desired regions of chemical space or conditioned to propose molecules with improved properties 17 ; 58 ; 52 ; 63 . Existing case studies, e.g., on the design of molecules using generative models for drug discovery and property optimization 48 ; 5 ; 52 , demonstrate promising progress, but they remain largely restricted to isolated molecular systems. However, molecules rarely exist in isolation; instead, they function within complex chemical and biological environments. Examples include protein ligand binding sites in enzymes, receptor ligand interactions central to drug discovery, and solvation or molecular complex formation in condensed phases 56 ; 10 . In each of these cases, the functionality of the system emerges not only from the properties of the individual components but also from their intricate interactions 43 . Such systems are inherently multi-component, and because the design space of each constituent is already vast, the combined search space grows explosively when considered as a whole 29 ; 35 . Attempting to generate entire multi-component systems directly within a single generative framework 62 ; 57 proves extremely challenging, even when using conditional generative approaches. When designing catalysts or reactive sites, not only the chemical composition needs to be optimized, but also the geometric features like relative orientation, distances, or symmetries of the individual components, to improve a reaction 28 . This combined compositional and configurational space is high-dimensional and heavily constrained by chemical validity and stability requirements. Consequently, rather than generating complete local environments in a single step, recent strategies decompose the problem into smaller, more tractable sub-tasks. Classical hierarchical or multiscale workflows can reduce the problem by first selecting a fixed set of candidate components, then optimizing their spatial arrangement using heuristic search strategies such as Monte Carlo sampling 47 , evolutionary algorithms 7 , or fragment-based assembly 40 . While effective in narrowing down large candidate spaces, these methods remain constrained by the diversity of the initial library and cannot autonomously propose new chemical components. As a result, the search is limited to recombining what is already known, leaving much of chemical space unexplored. In this work, we combine the best of both worlds in a single framework. In particular, we combine the possibility to go beyond known chemical structures with generative models 14 and the multi-component design using hierarchical models and propose a combined hierarchical generative optimization framework that breaks the problem of designing multi-component systems into manageable sub-problems, as illustrated in Figure 1 . In the first stage, a global optimization layer explores the spatial arrangement of components by treating positions, distances, and orientations as tunable parameters. A genetic algorithm searches this configurational space to identify arrangements that minimize a chosen target property. Subsequently, a generative modeling layer focuses on the chemical composition of the subunits themselves. By learning from the best-performing environments found in previous iterations, the generative model proposes new molecules with similar or improved stabilizing features, enriching the pool of candidates for subsequent optimization. Together, these two stages form a closed-loop pipeline that iteratively refines both the chemical composition of the subunits and the global structure of the whole system. A key application of our framework is the automated design of local catalytic environments, where subtle changes in electrostatics, dispersion, or hydrogen bonding can dramatically alter reaction rates. As a proof of concept, we apply the method to the Claisen rearrangement reaction of p-tolyl ether 24 and reduce the activation barrier by over 30%, revealing clear interaction motifs that stabilize the transition state, providing further chemical insights into design principles. II Results In this work, we apply the hierarchical multi-component optimization framework to the design of local chemical environments for reaction optimization by using the p-tolyl ether Claisen rearrangement reaction as an example. The proposed workflow is illustrated in Figure 1 and will be explained in the following. II.1 Multi-Component System Optimization Workflow The developed workflow (Figure 1 ) for designing a multi-component system can be understood as an assembly of several molecular or structural units placed in proximity to one another, as shown in the middle of the left side of Figure 1 . Each configuration is defined by two key elements: (i) a set of candidate building blocks that can occupy the system, and (ii) a global geometry that specifies how these blocks are arranged. The geometry is described by parameters such as distances, orientations, and positions, which together determine how the units are spatially organized, see also Figure 5 a. These two elements give rise to two coupled optimization