Retro-Synthesis Tasks

This page describes the chemical reaction and retro-synthesis datasets used in our benchmark. These tasks investigate the influence of synthesis knowledge on molecular generation, helping models learn to generate compounds that are both optimized and synthetically accessible.

Overview

Split	Size	Description
Training	50,000 reactions	Multi-step synthesis routes with various task types
Test (ChEMBL)	1,000 molecules	Real-world synthesis prediction
Test (Enamine)	1,000 molecules	Real-world synthesis prediction

Task Distribution

The training dataset includes four main task types:

Task Type	Proportion	Description
Retro-synthesis Planning	60%	Predict complete multi-step synthesis pathways
Reactant Prediction	25%	Identify missing reactants for a reaction step
SMARTS Prediction	10%	Predict the reaction template (SMARTS notation)
Product Prediction	5%	Predict the final product of a multi-step synthesis

Synthesis Complexity

The dataset contains reactions of varying complexity:

• Single-step reactions: 20,000 (40%)
• Two-step reactions: 18,000 (36%)
• Multi-step reactions (3-5 steps): 12,000 (24%)

---

Data Generation Pipeline

We follow a methodology that employs building blocks from the Enamine catalog and 115 chemical reaction templates described in SMARTS notation to generate multi-step reactions.

References

This data generation approach is derived by:

1. Lee et al., "Rethinking Molecule Synthesizability with Chain-of-Reaction." (2025)
2. Gao et al., "Generative Artificial Intelligence for Navigating Synthesizable Chemical Space." (2024)

by using their proposed Reactant-Reaction Matrix.

Multi-Step Synthesis Generation

We generate synthetic pathways through an iterative stochastic process:

• 1. Initialization

  ---

  Select a random seed reaction and identify available reactants via the compatibility matrix. Sample up to 10 valid reactant combinations and apply the reaction using RDKit. Filter products based on physicochemical properties and atom count.

• 2. Probabilistic Product Selection

  ---

  For each valid product, compute a probability score based on a target distribution over molecular properties (QED, molecular weight, TPSA, H-bond donors/acceptors, rotatable bonds, aromatic rings). Products are selected proportionally to these scores.

• 3. Chain Extension

  ---

  With up to 5 reaction steps, iteratively select a new reaction compatible with the last product, identify available reactant partners via the matrix, apply the reaction with property-based filtering, and add the product to the synthesis chain.

• 4. Termination

  ---

  Synthesis continues until the maximum number of steps is reached or no valid reactions can be applied. This ensures all pathways are chemically feasible.

Molecular Property Filtering

Products must satisfy strict physicochemical constraints to remain in the dataset, ensuring drug-like molecules:

Property	Min	Max
QED (Drug-likeness)	0.30	1.00
Molecular Weight (Da)	0	600
TPSA (Ų)	0	160
H-Bond Acceptors	0	10
H-Bond Donors	0	10
Rotatable Bonds	1	10
Aromatic Rings	0	6
Atom Count	-	60

Target Distribution Modeling

Rather than using hard constraints alone, we compute log-probabilities for products via Beta distributions over normalized property ranges. This biases the stochastic selection toward drug-like molecules without rejecting valid synthetic products. The distribution parameters are tuned on the ZINC-250K dataset.

---

Task Types

We created ten distinct objective templates to train models on complementary synthesis reasoning tasks:

Single-Step Tasks

• Final Product Prediction

  ---

  Predict the final product of a multi-step synthesis given the last reaction's SMARTS representation and reactants.

  ---

  Training samples: ~6.5k

• Reactant Prediction

  ---

  Identify a missing reactant for a single synthesis step (always first step).

  ---

  Training samples: ~3k

• All Reactants Prediction

  ---

  Given a reaction SMARTS and target product, predict all required reactants (always first step).

  ---

  Training samples:

  • ~1k with no additional information
  • ~1.5k with a set of building blocks provided

• SMARTS Identification

  ---

  Predict the SMARTS representation for a reaction step, given the reactants and product (any step of a synthesis).

  ---

  Training samples: ~1.5k

Multi-Step / Path Tasks

• Full Synthesis Path

  ---

  Generate a complete multi-step synthesis pathway to a target molecule.

  ---

  Training samples:

  • ~6.5k with not additional information
  • ~6.5k with a set of SMARTS templates provided
  • ~6.5k with the 4, 8 or 16 most similar building blocks to the target molecule provided
  • ~3k with both SMARTS templates and most similar building blocks provided

• Full Path With Interm. Products

  ---

  Generate a complete multi-step synthesis pathway to a target molecule, given possible intermediate products to help guide the model.

  ---

  Training samples:

  • ~6.5k with not additional information
  • ~6.5k with a building blocks available (including the ones used in the synthesis)

---

Reward Functions

The reward functions for chemical reaction tasks are designed to progressively guide the model toward correct predictions:

• Reactant/Product Prediction

  ---

  \[R = \begin{cases} 1 & \text{if prediction is correct} \\ 0 & \text{otherwise} \end{cases}\]

  Evaluates correctness by verifying if using the predicted reactants/products in the reaction yields the expected product/reactants.

• SMARTS Prediction

  ---

  \[R = \frac{9 \times \mathbb{1}_{SMARTS_{pred} = SMARTS_{ref}} + \mathbb{1}_{product\_match}}{10}\]

  High reward for exact SMARTS match, small reward if applying the predicted SMARTS produces the correct product.

• Retro-Synthesis Planning

  ---

  \[R = \left(\frac{n_{valid}}{n}\right)^2 \times \text{sim}(target, \hat{y})^3\]

  Where \(n_{valid}\) is the number of valid steps, \(n\) is total steps, and \(\hat{y}\) is the last valid product. Rewards increase with valid step proportion and Tanimoto similarity to target.

Invalid Predictions

If the extracted answer is invalid (unparseable SMILES, invalid reaction), the reward is automatically set to 0.

---

Evaluation

Test Sets

Following established methodology, we evaluate on real-world synthesis prediction rather than synthetic data:

Test Set	Size	Description
ChEMBL	1,000 molecules	Drug-like molecules from the ChEMBL database
Enamine	1,000 molecules	Molecules from the Enamine catalog

For each molecule, we either:

1. Directly prompt the model to predict the synthesis route
2. Prompt the model to predict the synthesis route given a set of building blocks (4, 8, or 16 most similar to the target).

Evaluation Metrics

Model performance is evaluated based on:

• Success rate: Proportion of molecules successfully synthesized using predicted routes
• Tanimoto similarity: Similarity between target molecule and synthesized product (when synthesis fails)
• Valid step ratio: Proportion of chemically valid steps in predicted routes

---

References

1. Lee, S., et al. "Rethinking Molecule Synthesizability with Chain-of-Reaction." (2025)
2. Gao, W., et al. "Generative Artificial Intelligence for Navigating Synthesizable Chemical Space." (2024)
3. Enamine Building Blocks Catalog: https://enamine.net/building-blocks/building-blocks-catalog