From Feasible to Practical: Pareto-Optimal Synthesis Planning

Friedrich Hastedt, Dongda Zhang, Antonio del Rio Chanona · 2026

Read From Feasible to Practical: Pareto-Optimal Synthesis Planning by Friedrich Hastedt, Dongda Zhang, Antonio del Rio Chanona (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-2605-07521.txt.

Introduction

Friedrich Hastedt and two co-authors tackle computer-aided synthesis planning, the task of finding routes to build a target molecule from simpler chemicals. Most tools stop once they find one workable route. Real chemists, though, must weigh cost, toxicity, sustainability, and yield against each other, so a single answer is rarely enough.

Their algorithm, MORetro*, searches for a whole Pareto front of routes, each best under some balance of the user's chosen criteria. It uses weighted scoring and guided sampling to search efficiently, and it carries a guarantee of recovering the true front under stated conditions. On standard benchmarks it surfaced strong routes that single-goal searches had missed.

Catalog record

CREATORFriedrich Hastedt, Dongda Zhang, Antonio del Rio Chanona
YEAR2026
COLLECTIONPAPERS
FORMATTXT
FILE SIZE64 KB
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2605.07521

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

From Feasible to Practical: Pareto-Optimal Synthesis Planning Friedrich Hastedt, Dongda Zhang, Antonio del Rio Chanona Abstract Current computer-aided synthesis planning (CASP) methods often treat retrosynthesis as solved once a single feasible route is identified, focusing primarily on convergence or shortest-path metrics. This view is misaligned with real-world practice, where chemists must balance competing objectives such as cost, sustainability, toxicity, and overall yield. To address this, we formulate synthesis planning as a multi-objective search problem and introduce MORetro ∗ , an algorithm that generates a Pareto front of synthesis routes to explicitly capture trade-offs among user-defined criteria. MORetro ∗ uses weighted scalarization and BO-informed sampling to efficiently navigate the combinatorial search space and prioritize promising trade-offs. Building on multi-objective A ∗ -search, we provide optimality guarantees showing that, for a fixed single-step model, MORetro ∗ recovers the true Pareto front under admissibility. Across multiple retrosynthesis benchmarks, MORetro ∗ produces diverse, high-quality Pareto fronts, uncovering solutions overlooked by single-objective approaches and better aligning CASP outputs with industrial decision-making. 1 Introduction Retrosynthesis is a chemical discipline central to accelerating drug and material discovery. Starting with a target molecule, the aim of retrosynthesis is to find synthesis routes to commercially available building blocks in a backward fashion. Once a viable synthesis route is identified, the building blocks can be purchased and the physical synthesis is carried out through wet lab experimentation. However, finding such a viable synthesis route is not trivial, as each intermediate molecule in the synthesis tree could theoretically be synthesized from hundreds of chemical transformations. This becomes particularly challenging when one tries to find a synthesis route that also optimizes key metrics, such as economic and sustainability costs. To facilitate synthesis planning, researchers proposed computational workflows that can navigate these intrinsically large search spaces efficiently ( 11 ; 38 ) . These workflows are commonly composed of two interacting tools: i) a single-step model and ii) a multi-step planning algorithm. The single-step model is usually a black-box machine learning (ML) model that takes the product molecule as input and returns a set of possible reactants ( 31 ) . The single-step model is optimized offline to faithfully reproduce feasible reaction chemistry. As the name suggests, the single-step model only considers one reaction at a time. However, synthesis routes usually consist of multiple sequential reactions. To ensure convergence to commercial building blocks across the immense reaction space, the multi-step algorithm iteratively calls the single-step model, thereby constructing a synthesis route ( 30 ) . Most multi-step algorithms to date were proposed with this exact goal in mind: returning at least one synthesis route that ends in commercial building blocks. From the perspective of the search problem, it is paramount to find at least one synthesis route per target. In practice, chemists would nonetheless investigate diverse synthesis pathways to identify the ones that best meet personal objectives, which are often a combination of financial and sustainability interests. Figure 1: Exemplary Pareto front generated by MORetro ∗ . Each red point corresponds to a Pareto-optimal synthesis route given an arbitrary number of user-specified objectives. Embedding multiple objectives directly in the search algorithm is not trivial, as it could interfere with the algorithm’s ability to find synthesis pathways. That is why the multi-step algorithm is commonly used to identify an unranked set of synthesis routes. Post-hoc analysis then optimizes with specific objectives in mind ( 8 ) , often adding relevant conditions (temperature and agents) in the process ( 34 ) . By keeping these two stages separate, it is highly probable that the algorithm misses out on promising synthesis routes. In this paper, we address this challenge by introducing MORetro ∗ , a search algorithm for multi-objective synthesis planning. Given an arbitrary number of user-defined objectives, MORetro ∗ returns the Pareto front (PF), where each point corresponds to a unique synthesis route (Figure 1 ). MORetro ∗ extends the single-objective Retro ∗ algorithm ( 4 ) to the multi-objective setting by down-sampling the high-dimensional objective space via linear (weight-based) scalarization. To effectively explore this space, we introduce several weight-sampling strategies. Moreover, we derive a lower bound on the multi-dimensional objective space that guarantees the returned front to be Pareto-optimal. This bound also enables early pruning of dominated intermediate molecules; those that cannot be part of any Pareto-optimal synthesis route if further expanded. Finally, MORetro ∗ directly integrates a reaction condition prediction module, which augments each reaction with predicted conditions such as temperature and reagents during the search. This integration allows us to define and optimize novel objectives of practical relevance to the chemical industry, including sustainability and scalability. Our contributions can be summarized as follows: • We introduce a multi-objective retrosynthesis search algorithm that returns the Pareto front for any number of user-defined objectives, using weight scalarization and sampling. It outperforms single-objective baselines in both Pareto front quality and diversity, independent of the single-step model. • We derive optimality bounds guaranteeing that, for a given single-step model, our algorithm can recover all Pareto-optimal synthesis routes. • We define industrially relevant search objectives capturing toxicity, sustainability, and scale-up potential, enabling the algorithm to prioritize practically feasible reactions. 2 Background 2.1 Related Work Computer-aided Synthesis Planning Retrosynthetic planning is commonly formulated as a tree-search problem, where branches correspond to chemical transformations predicted by a single-step model. These transformations are often represented as reaction templates, leading to template-based approaches that rely on neural policies ( 6 ; 5 ) or expert rules ( 35 ) . In contrast, template-free methods avoid predefined rules and formulate retrosynthesis as sequence-to-sequence translation ( e.g. , SMILES) ( 29 ; 14 ) or molecular graph editing ( 27 ; 32 ; 48 ) . Extending single-step prediction to multi-step synthesis requires a search algorithm to select promising intermediates during tree expansion. Monte Carlo Tree Search (MCTS) addresses this via rollout-based exploration-exploitation trade-offs ( 30 ) . Alternatives avoid rollout by using AND-OR trees with heuristic guidance, either hand-crafted or neural ( 16 ; 4 ) , with 4 proposing a neural-guided A ∗ -like algorithm. Subsequent work builds on MCTS and A ∗ to improve route-finding success rates ( 15 ; 43 ; 37 ; 47 ) . A parallel line of work applies reinforcement learning to fine-tune single-step models in the multi-step setting ( 28 ; 46 ; 20 ; 39 ) . Unlike our approach, these methods primarily aim to maximize the likelihood of discovering any valid synthesis route. Constrained Synthesis Planning Recent studies incorporate explicit constraints into retrosynthesis search, which, while not framed as multi-objective optimization, are conceptually related, as constraints impose simultaneous requirements on valid routes. Examples include starting-material-constrained retrosynthesis targeting specific building blocks ( 45 ; 1 ) , bond constraints that restrict which bonds may be broken or preserved ( 36 ; 17 ; 42 ) , and approaches that condition retrosynthesis on human prompts using large language models ( 44 ) . Multi-objective Synthesis Planning 19 first explored multi-objective synthesis planning, proposing a multi-objective Monte Carlo Tree Search (MO-MCTS) framework that maintains local Pareto fronts over molecular nodes and applying it to objectives such as stock availability, step count, synthetic complexity, and route similarity. In contrast to our work, their approach relies on stochastic

Read the complete text