🧬 biology

HiRes: Inspectable Precedent Memory for Reaction Condition Recommendation

HiRes is a retrieval-augmented reaction condition recommendation system that combines hierarchical graph representations with k-NN precedent retrieval to achieve state-of-the-art prediction accuracy while providing chemists with inspectable, justifying precedents for practical synthesis planning.

Original authors: Shreyas Vinaya Sathyanarayana, Raja Sekhar Pappala, Deepak Warrier

Published 2026-05-21

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Shreyas Vinaya Sathyanarayana, Raja Sekhar Pappala, Deepak Warrier

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ⚕️ This is an AI-generated explanation of a preprint that has not been peer-reviewed. It is not medical advice. Do not make health decisions based on this content. Read full disclaimer

Imagine you are a chef trying to recreate a complex dish you saw in a magazine. You know the final result (the "product"), and you've figured out the main ingredients you need to start with (the "reactants"). But you're stuck on the most crucial part: how to actually cook it. You need to know exactly which stove temperature (catalyst), which pot of oil or water (solvent), and which specific spice blend (reagent) will make the recipe work.

This is the problem chemists face every day, and this paper introduces a new AI tool called HiRes to solve it.

Here is a simple breakdown of what HiRes does, using everyday analogies:

1. The Problem: The "Recipe" Gap

In chemistry, figuring out what ingredients to mix is only half the battle. The other half is knowing the conditions (the heat, the liquid, the helpers) to make the reaction happen.

Old AI tools were like a chef who could guess the ingredients but couldn't explain why they picked them. They just gave you a label like "Use Salt" without showing you a similar dish that worked before.
HiRes is like a chef who not only guesses the ingredients but also pulls out a physical cookbook, points to three similar recipes that worked in the past, and says, "I'm suggesting this because these three dishes turned out great using the same method."

2. How HiRes Works: The "Smart Library"

HiRes is built on a concept called Hierarchical Reaction Representations. Think of it as a library that organizes chemical reactions in three layers, just like a human chef thinks:

Layer 1 (The Ingredients): It looks at the individual molecules (the atoms and bonds) like a chef looking at raw vegetables.
Layer 2 (The Cut): It looks at how the ingredients change when mixed (the "disconnection"). It's like noticing that a carrot was chopped into a specific shape to fit a stew.
Layer 3 (The Full Dish): It combines everything into a complete picture of the reaction.

The Magic Trick:
Usually, AI models are either predictors (guessing the answer) or search engines (looking up old answers). HiRes does both at the same time.

It learns a "language" of chemistry.
When you ask it for a recipe, it uses its brain to predict the best catalyst, solvent, and reagent.
Simultaneously, it uses that same "brain" to search its memory bank for the top 10 most similar past reactions.
It then combines its own guess with the evidence from those past reactions to give you the final answer.

3. The Results: Better Guesses with Proof

The paper tested HiRes on a massive database of over 680,000 chemical reactions (the USPTO-Condition dataset). Here is what they found:

It's a Top Performer: HiRes became the best at predicting the right "solvent" and "reagent" compared to other famous AI models. It tied for the best at predicting "catalysts."
The "Hybrid" Power: The researchers found that the AI's "guessing brain" (the learned model) and its "search engine" (the memory of past reactions) are better together than apart.
- Analogy: Imagine a student taking a test. If they just rely on memorization, they might get stuck. If they just look up answers, they might miss the nuance. HiRes is the student who studies hard and has a cheat sheet of similar problems to double-check their work. This combination made the predictions significantly more accurate.
Transparency: Unlike other models that just give you a number, HiRes gives you the precedents. It shows you the specific past experiments that support its recommendation. This is crucial for chemists who need to trust the AI before they start mixing chemicals in a lab.

4. Why This Matters

The paper argues that for AI to be truly useful in a chemistry lab, it can't just be a "black box" that spits out answers. It needs to be inspectable.

HiRes bridges the gap between accuracy (getting the right answer) and interpretability (showing you the proof). It provides a single system that acts as both a high-speed calculator and a searchable library of chemical history, helping chemists move from a theoretical idea to a real, working experiment with confidence.

In short: HiRes is an AI assistant that doesn't just tell you what to use to make a chemical reaction; it shows you the proof from history that it will work, making the process safer and more reliable.

Technical Summary: HiRes – Inspectable Precedent Memory for Reaction Condition Recommendation

Problem Statement

Reaction condition recommendation is a critical downstream task in AI-driven synthesis planning, occurring after retrosynthetic disconnection selection. While chemists require accurate predictions for catalysts, solvents, and reagents, they also require the concrete chemical precedents that justify those recommendations to ensure experimental feasibility.

Existing approaches typically optimize only one half of this workflow:

Parametric classifiers map reaction graphs or strings directly to condition labels but lack inherent interpretability or precedent retrieval.
Non-parametric memories (e.g., Morgan fingerprints, DRFP) offer efficient nearest-neighbor lookups but may lack the nuanced understanding of complex reaction transformations.
Semi-parametric systems in other domains suggest that combining learned predictors with external datastores improves performance, but this integration has not been fully realized in a unified chemical representation that serves both as a classifier feature and an inspectable precedent index.

The core challenge addressed by this paper is: Can a single learned reaction representation support both direct condition prediction and a precedent index queryable at decision time?

Methodology: HiRes (Hierarchical Reaction Representations)

HiRes is a retrieval-augmented condition recommendation system built on a hierarchical representation of chemical reactivity, progressing from molecular structure to bond disconnections and finally to full reaction patterns.

1. Hierarchical Architecture

The model follows a three-level progression:

Level 1: Molecule Representations: Reactants and products are encoded using a shared four-layer Graph Attention Network (GATv2) with virtual nodes. This enforces a common molecular basis before reasoning about reaction direction.
Level 2: Disconnection Representations: Atom-level embeddings interact via bidirectional cross-attention. If atom maps are available, a soft attention bias is applied; otherwise, the model defaults to standard cross-attention. The model generates two zero-parameter decompositions:
- $z_\Delta = p - r$ : Emphasizes the chemical transformation.
- $z_\Sigma = p + r$ : Preserves the reaction context.
Level 3: Reaction and Condition Representations: The full reaction vector $z_{rxn}$ $z_{r x n}$ is a gated fusion of six explicit streams:
1. Concatenated reactant-product context.
2. Difference ( $z_\Delta$ ).
3. Sum ( $z_\Sigma$ ).
4. Engineered descriptors (including Differential Reaction Fingerprints, DRFP).
5. DFT-derived features.
6. Reaction-center difference.
  This fused vector (256-dimensional) is concatenated with $z_\Delta$ and fed into residual MLP heads for predicting catalyst, solvent, and reagent roles. Training utilizes ambiguity-aware multi-hot supervision with focal-weighted binary cross-entropy to handle multiple valid condition sets.

2. Retrieval-Augmented Prediction

HiRes integrates a non-parametric retrieval layer to complement the parametric heads:

Indexing: Training reactions are indexed by their learned embeddings ( $z_{rxn}$ ) using FAISS.
Hybrid Inference: For a query, the system retrieves $k=10$ nearest neighbors. The final prediction is a weighted combination of the learned head probabilities ( $p_{head}$ ) and the k-NN neighbor voting probabilities ( $p_{knn}$ ):
$p_{hyb}(y | R) = \alpha p_{head}(y | R) + (1 - \alpha) p_{knn}(y | R)$
where $\alpha = 0.5$ in the fixed configuration.

Key Contributions

Unified Representation: A single hierarchical representation that acts simultaneously as a strong classifier feature and an inspectable precedent memory, pairing every recommendation with concrete neighbors.
State-of-the-Art Performance: Achieved top-tier results on the primary-slot USPTO-Condition benchmark, outperforming established baselines like REACON on solvent and reagent selection while matching the best reported catalyst scores.
Statistically Validated Complementarity: Demonstrated via paired bootstrap analysis that fusing learned prediction heads with k-NN retrieval yields statistically significant gains for solvent and reagent selection over purely parametric approaches.
Integrated Evaluation: Showed that a unified framework can streamline the computational pipeline for chemist-facing assistants by exposing inspectable neighbors alongside predictions.

Experimental Results

The system was evaluated on the USPTO-Condition dataset (680,741 reactions) using the standard train/validation/test splits and the REACON Table S9 evaluation protocols.

Primary Slot Accuracy (Acc@1):
- Catalyst: 0.929 (Ties best reported baseline, e.g., REACON D-MPNN).
- Solvent: 0.534 (Outperforms REACON by +3.0 points).
- Reagent: 0.530 (Outperforms REACON by +3.0 points).
Hybrid vs. Parametric: The hybrid model (HiRes-Hybrid) improved Solvent Acc@1 by +1.1 points and Reagent Acc@1 by +1.0 points over the purely parametric End2End heads. These gains were confirmed to be statistically significant (95% CI) via paired bootstrap over 10,000 resamples.
Baseline Comparison: HiRes-Hybrid outperformed exact-search DRFP baselines on solvent and reagent prediction, demonstrating that learned representations surpass static fingerprints in these complex slots.
Catalyst Imbalance Note: The authors explicitly note that 87% of catalyst test rows map to "absent/None." While the overall Acc@1 is high (0.929) due to this imbalance, the "present-only" Acc@1 is 0.604, indicating robust active predictive power.

Significance and Claims

The paper claims that HiRes bridges the gap between predictive accuracy and chemical interpretability. By using the same learned reaction space for both classification and retrieval, HiRes provides:

Competitive Recommendations: High-accuracy predictions for catalysts, solvents, and reagents.
Actionable Precedents: Immediate access to the specific chemical precedents that justify the recommendation, which is essential for practical synthesis planning and expert intervention.

The authors position HiRes not as an autonomous execution engine, but as a decision support system for chemists. It is designed to reduce manual search effort and increase transparency in AI-assisted synthesis workflows, with the explicit caveat that recommendations must be verified by qualified chemists and constrained by safety procedures. The work establishes that integrating retrieval with learned condition heads provides a complementary signal that is particularly valuable for solvent and reagent selection, where precedent data is dense and highly informative.