🤖 AI

From Feasible to Practical: Pareto-Optimal Synthesis Planning

This paper introduces MORetro*, a multi-objective search algorithm that generates Pareto-optimal synthesis routes by balancing competing criteria like cost and sustainability, thereby aligning computer-aided synthesis planning with real-world industrial decision-making.

Original authors: Friedrich Hastedt, Dongda Zhang, Antonio del Rio Chanona

Published 2026-05-11

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Friedrich Hastedt, Dongda Zhang, Antonio del Rio Chanona

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are a master chef trying to create a new, complex dish. You know the final flavor you want (the target molecule), but you need to figure out how to build it from basic ingredients you can buy at the grocery store (commercial building blocks).

In the world of chemistry, this is called retrosynthesis. For a long time, computer programs designed to help chefs (chemists) with this task had a simple goal: "Find any recipe that works." If the computer found a path from the grocery store to the final dish, it stopped and said, "Here is a solution!"

But in the real world, a "working" recipe isn't enough. A chef also cares about:

Cost: Is it too expensive?
Safety: Are the ingredients toxic or dangerous?
Sustainability: Does it waste energy or create pollution?
Ease: Is it easy to separate the good parts from the waste?

The paper you provided introduces a new tool called MORetro*. Think of it as a "Super Chef's Assistant" that doesn't just find one recipe, but finds the perfect balance of all these competing goals.

The Problem: The "One-Size-Fits-All" Trap

Previous computer programs were like a GPS that only cared about the shortest distance. It would tell you the fastest route to your destination, even if that route went through a dangerous construction zone or a toll road you couldn't afford.

Chemists, however, need to make trade-offs. Maybe they want a slightly longer route if it avoids a toxic chemical. Maybe they want a cheaper route even if it takes more steps. The old computers couldn't handle this "balancing act" well; they usually just gave you one option and asked you to figure out the rest later.

The Solution: The "Pareto Front" (The Menu of Perfect Choices)

The authors created an algorithm called MORetro*. Instead of giving you one single "best" route, it gives you a menu of perfect choices, known as a Pareto Front.

The Analogy: The Car Dealership
Imagine you are buying a car. You care about three things: Price, Speed, and Gas Mileage.

A cheap car might be slow and drink a lot of gas.
A fast car might be expensive and drink a lot of gas.
A gas-efficient car might be slow and expensive.

There is no single "best" car. Instead, there is a list of cars where you can't improve one feature without making another worse.

Car A: Cheap, Slow, Good Gas.
Car B: Expensive, Fast, Good Gas.
Car C: Medium Price, Medium Speed, Bad Gas.

MORetro* generates this exact list for chemical reactions. It shows the chemist: "Here is the cheapest route. Here is the safest route. Here is the most sustainable route. And here are all the smart compromises in between."

How It Works (The Magic Trick)

The paper explains that the computer uses a clever trick called Weighted Sampling and Bayesian Optimization.

The Weighted Dice: Imagine the computer rolls a set of dice to decide how much it cares about cost vs. safety vs. speed. It rolls the dice thousands of times, trying different combinations (e.g., "Today I care 80% about safety, 20% about cost").
The Smart Explorer: Instead of just guessing randomly, the computer uses a "smart explorer" (Bayesian Optimization). If it finds a combination of weights that leads to a great new recipe, it remembers that area and explores it more deeply. If a combination leads to a dead end, it stops wasting time there.
The Pruning: The computer is also very good at ignoring bad paths early. If it sees a path that is clearly worse than a path it already found (e.g., it's more expensive and more toxic), it cuts that branch off the tree immediately. This saves a massive amount of computing power.

What They Found

The researchers tested this new tool against old methods using real chemical data.

Better Variety: The old tools gave chemists a few similar recipes. MORetro* gave them a wide variety of distinct recipes, offering real choices.
Better Quality: The recipes found by MORetro* were generally "better" across the board (cheaper, safer, or more sustainable) than the single recipes found by the old tools.
Still Reliable: Even though it was looking for many different things at once, it didn't lose its ability to find a solution. It still found a working recipe almost as often as the old tools did.

The Bottom Line

This paper doesn't claim to have invented a new drug or a new chemical. Instead, it invented a better way to plan how to make them.

It moves computer-aided chemistry from asking, "Can we make this?" to asking, "What is the best way to make this, considering money, safety, and the environment?" It gives chemists a dashboard of options so they can make the smartest decision for their specific needs, rather than being forced to take the first path the computer finds.

Technical Summary: From Feasible to Practical: Pareto-Optimal Synthesis Planning

Problem Statement

Current Computer-Aided Synthesis Planning (CASP) methods predominantly treat retrosynthesis as a solved problem once a single feasible route is identified, optimizing primarily for convergence or shortest-path metrics. This single-objective perspective misaligns with industrial reality, where chemists must balance competing, often conflicting, objectives such as cost, sustainability, toxicity, and overall yield. Existing approaches typically separate route generation from optimization, using multi-step algorithms to find unranked routes and applying post-hoc analysis to optimize specific criteria. This separation risks missing promising synthesis pathways that do not excel in a single metric but offer superior trade-offs. Furthermore, while constrained synthesis planning exists (e.g., fixing starting materials), true multi-objective optimization that explicitly captures trade-offs via a Pareto front has been underexplored, particularly with theoretical guarantees of optimality in the complex AND-OR graph structure of retrosynthesis.

Methodology: MORetro*

The authors introduce MORetro, a multi-objective search algorithm designed to generate a Pareto front (PF) of synthesis routes for any number of user-defined objectives. The method extends the single-objective Retro* algorithm (Chen et al., 2020) and RetroGraph (Xie et al., 2022) to a multi-objective setting using the following core components:

1. Multi-Objective A* Search on AND-OR Graphs

The search space is modeled as a directed acyclic graph (DAG) where molecules are OR nodes (requiring only one reaction to be solved) and reactions are AND nodes (requiring all reactants to be solved).

Scalarization: To navigate the high-dimensional objective space, MORetro* employs weighted linear scalarization. It generates a set of weight vectors $W$ sampled from the probability simplex. Each weight vector $w_j$ converts the vector-valued cost $C(\Gamma)$ into a scalar value $V_{t,j} = w_j^\top V_t$ .
Parallel Search: The algorithm maintains multiple parallel searches, one for each sampled weight vector. In each iteration, frontier molecules are selected based on the scalarized value function for each weight, expanded using single-step predictors, and costs are propagated upward and downward through the graph.

2. Bayesian Optimization (BO) for Weight Sampling

To efficiently explore the trade-off space, the paper proposes a BO-informed weight sampling strategy.

Mechanism: After an initial warm-up phase (using grid or Sobol sampling), a surrogate model is fitted to the observed hypervolume (HV) improvements of previously evaluated weights.
Acquisition: New weight vectors are sampled to maximize the expected information gain (IG) using the GIBBON acquisition function, which promotes diversity in the batch of weights while focusing on promising regions of the trade-off space.

3. Optimality Guarantees and Pruning

A critical theoretical contribution is the derivation of a vector-valued pruning number ($pn$) and a bound ( $V_{bound}$ ) for the AND-OR graph.

Theorem: The authors prove that if the search terminates only when the "bound dominance condition" holds for all frontier molecules (i.e., a discovered route dominates the lower bound of any unexpanded node), the algorithm recovers the true Pareto front.
Pruning: This bound enables the safe pruning of dominated intermediate molecules that cannot be part of any Pareto-optimal route, significantly reducing the search space.

4. Practical Objectives

The framework integrates a reaction condition prediction module (QUARC) to augment reactions with predicted temperatures and reagents. This enables the definition of industrially relevant objectives:

Sustainability: A combination of reaction temperature penalties and atom economy.
Toxicity: Aggregates hazards of catalysts/solvents and penalizes toxic moieties in reactants.
Scale-up Potential: Estimates separation difficulty based on logP differences and incorporates building block market prices.

Key Contributions

Algorithmic Framework: Introduction of MORetro*, a multi-objective search algorithm that returns the Pareto front for arbitrary user-defined objectives using weight scalarization and BO-informed sampling. It outperforms single-objective baselines in both PF quality and diversity.
Theoretical Guarantees: Derivation of optimality bounds proving that, for a fixed single-step model, the algorithm can recover all Pareto-optimal synthesis routes. This includes a novel pruning mechanism for AND-OR graphs that generalizes admissibility concepts from standard OR graphs.
Industrial Relevance: Definition of search objectives capturing toxicity, sustainability, and scale-up potential, allowing the algorithm to prioritize practically feasible reactions over purely theoretical shortest paths.

Experimental Results

The method was evaluated on the USPTO-190, Pistachio-reachable, and ChEMBL datasets using three different single-step models (NeuralSym, PDVN, and Graph2Edits).

Pareto Front Quality: MORetro* (with BO sampling) consistently outperformed single-objective Retro** and fixed-weight scalarization baselines. It achieved higher Hypervolume (HV) improvements (approx. 5–8%) and lower $R^2$ indicators, indicating better convergence and coverage of the trade-off space.
Success Rate: MORetro* maintained a success rate comparable to the single-objective Retro** (failing on only 2–3 additional molecules out of 190), whereas fixed-weight baselines failed significantly more often (up to 25% fewer targets solved). This highlights the importance of adaptively balancing the guidance objective.
Diversity: The algorithm produced significantly more diverse synthesis routes (measured by pairwise dissimilarity) compared to baselines, demonstrating that multi-objective optimization naturally explores a broader chemical space.
Pruning Efficiency: Using the $V_{bound}$ for pruning reduced the search space by an average of 54%, with up to 90% reduction for specific molecules, while improving PF quality and reducing runtime.
Sampling Strategy: BO-informed sampling consistently outperformed Sobol and grid sampling in terms of HV, success rate, and diversity, particularly for complex models like PDVN.

Significance and Claims

The paper claims that MORetro* bridges the gap between theoretical feasibility and industrial practicality in synthesis planning. By moving beyond the "single feasible route" paradigm, it provides a flexible foundation for decision-making that explicitly captures trade-offs between economic, environmental, and safety metrics.

The authors emphasize that their approach:

Aligns with Industrial Needs: By directly optimizing for sustainability and scale-up, the outputs are more actionable for chemists than simple shortest-path solutions.
Provides Theoretical Rigor: Unlike stochastic approaches (e.g., MO-MCTS), MORetro* offers guarantees of Pareto-optimality under specific conditions, ensuring that no superior trade-off is missed due to random sampling.
Is Model-Agnostic: The method is independent of the underlying single-step predictor, making it applicable to various template-based and template-free models.

The paper concludes that while current heuristics are not provably admissible (limiting pruning efficiency in practice), the framework successfully recovers high-quality, diverse Pareto fronts, offering a significant step forward in making CASP outputs more relevant to real-world chemical synthesis.