DeepRetro: Hybrid Retrosynthetic Planner

• DeepRetro is an open-source retrosynthetic planning framework that fuses a template-based Monte Carlo Tree Search engine with a large language model reasoning module to propose viable and novel synthesis pathways.
• It employs a hybrid, iterative pipeline combining deterministic template-based searches with generative LLM suggestions, rigorously filtered for chemical validity, stability, and logical consistency.
• The framework features an interactive GUI that supports human-in-the-loop corrections and selective replanning, enhancing both the feasibility and innovation of chemical synthesis strategies.

DeepRetro is an open-source retrosynthetic planning framework integrating a template-based Monte Carlo Tree Search (MCTS) engine with a LLM reasoning module in an iterative, feedback-driven architecture. The system is designed to identify viable and potentially novel synthesis pathways for complex target molecules, moving beyond the limitations imposed by fixed reaction templates while maintaining rigorous chemical validity through a suite of specialized filters and optional human-in-the-loop guidance (Sathyanarayana et al., 7 Jul 2025).

1. System Architecture and Algorithmic Workflow

DeepRetro fuses the strengths of established template-based retrosynthesis tools with the generative and reasoning capabilities of LLMs, yielding a hybrid iterative pipeline for pathway discovery. The retrosynthetic process proceeds as follows:

• The user provides a target molecule $m$ (specified by SMILES) and a set $S$ of commercially available molecules (the "stock").
• If $m\in S$, the synthesis terminates.
• The template-based engine (Tool T), implemented with MCTS, is invoked to discover a route to $S$ using learned or predefined reaction templates. If successful, the pathway is returned.
• If the template engine fails, the LLM module is prompted for $k$ single-step disconnection proposals, each accompanied by brief chain-of-thought justifications and confidence scores.
• Each LLM-generated precursor set is subjected to:
  1. Validity filtering (chemically plausible valence, allowed elements, and parseability)
  2. Stability checks (absence of explosive or strained motifs, reasonable computed properties, heuristic structure rules)
  3. Hallucination screening (subgraph isomorphism to verify chemical logic, reagent-type consistency)

• Surviving precursor branches are recursively processed by re-invoking the entire DeepRetro loop.

• The recursion continues until all branches terminate at in-stock compounds or the search budget is exhausted.
• An interactive graphical user interface allows for pathway inspection, selective editing, and partial replanning by human experts.

2. Search Algorithms and Mathematical Foundation

The central algorithmic structure of DeepRetro is governed by a recursive function combining classic MCTS and LLM-driven retrosynthetic reasoning. The template-based engine uses a PUCT strategy for tree traversal, formulated as:

$$\pi(a\mid s) = \frac{Q(s,a) + U(s,a)}{\sum_{b}(Q(s,b)+U(s,b))}$$

where the exploration bonus is given by:

$$U(s,a) = c\,P(a\mid s)\,\frac{\sqrt{N(s)}}{1 + N(s,a)}$$

with $Q(s,a)$ denoting the mean value of roll-outs from action $a$ in state $s$, $S$0 the action visit count, and $S$1 a neural/template policy prior. Candidate actions are ranked for selection as:

$S$2

Upon expansion, simulation evaluates heuristic success (e.g., whether new precursors are in stock). Backpropagation updates action values throughout the tree.

LLM proposals are validated and filtered before recursive invocation of DeepRetro on each plausible precursor. This dual-pronged search blends deterministic template enumeration with generative natural language-based single-step prediction, supported by extensive post hoc filtering to ensure chemical soundness.

3. Molecule-Level Filtering: Validity, Stability, and Hallucination

Robust chemical screening is essential for leveraging LLM creativity while controlling for infeasible output. DeepRetro applies three layers of molecular filtering:

• Validity Checks: Evaluate atom valence adherence, restrict to allowed elements (C, H, O, N, P, S, halogens), and ensure SMILES are parseable and sanitize to valid molecular graphs.
• Stability Checks: Exclude precursors with patterns matching known explosives or highly strained rings (via SMARTS), enforce property bounds (e.g., predicted pKa, logP), and apply heuristic shape filters (e.g., maximal ring size or repetition). A stability score $S$3 is assigned, and proposals with $S$4 (such as 0.4) are discarded.
• Hallucination Checks: Require that disconnections correspond to workable graph cuts (subgraph isomorphism), and that proposed reagents match the transformation type (e.g., Suzuki coupling proposals must involve boronic acids). Reactions inconsistent between textual description and precursor structure are rejected.

This rigorous post-processing constrains the search to chemically believable and procedurally consistent retrosynthetic moves.

4. Empirical Evaluation: Benchmarks and Case Analyses

DeepRetro's performance was benchmarked on both single-step and multi-step retrosynthesis datasets, with direct comparison to ASKCOS and AiZynthFinder.

Table 1: Single-step Top-1 Accuracy on 100-compound USPTO-50k Subset

Model	LLM	Data	All-Correct (%)	Any-Correct (%)
ASKCOS	–	Reaxys	32.32	42.42
AiZynthFinder	–	Pist.	31.31	41.41
DeepRetro	Claude 3.7	Pist.	2.56	52.56
DeepRetro	DeepSeek	Pist.	1.14	47.12
AiZynthFinder	–	USPTO	29.29	39.39
DeepRetro	Claude 3.7	USPTO	1.21	43.90
DeepRetro	DeepSeek	USPTO	0.00	41.86

Table 2: Multi-step Success Rate on USPTO-190 Test Set

Model	LLM	Data	Success %
Retro*	–	USPTO	80
PDVN	–	USPTO	80
DeepRetro	Claude 3.7	Pist.	80
DeepRetro	DeepSeek	Pist.	60
DeepRetro	Claude 3.7	USPTO	50
DeepRetro	DeepSeek	USPTO	80

Search times (50–300 s per target) are comparable to AiZynthFinder.

Case studies on targets including Ohauamine C, a tetracyclic azepine derivative, and erythromycin B demonstrate DeepRetro’s ability to propose non-template-based disconnects. For Ohauamine C, the route entails LLM-guided amide bond cleavage, esterification, and peptide cyclization steps, culminating in commercially available amino acids following validation and recursive expansion. All cases required 6–12 DeepRetro calls and occasional human corrections.

5. Human-in-the-Loop Feedback and User Interface

A React-based graphical user interface supports advanced interactive capabilities:

• Pathway Visualization: Tree graph of explored syntheses with metadata overlays (confidence, scalability index).
• Experimental Controls: Target molecule and LLM selection, stability/hallucination check toggles, stock/import settings, step/yield constraints, and MCTS budget specification.
• Editable Intermediates: Direct SMILES editor to allow expert correction of LLM-predicted structures or pathway branches.
• Selective Replanning: Local replanning of any tree branch using MCTS or LLM, without restarting global search.
• Recursion with Human Edits: User modifications feed directly into subsequent DeepRetro invocations, facilitating collaborative optimization between algorithm and domain expert.

This infrastructure enables human chemists to correct errors (e.g., protecting group placement, stereochemistry) and actively steer searches toward desirable synthetic strategies.

6. Limitations, Failure Modes, and Directions for Advancement

Identified challenges and open issues encompass:

• LLM Hallucinations: Although filtered, general LLMs can produce chemically implausible suggestions, increasing computational overhead.
• Data Bias: Template and LLM pre-training data steer solutions toward known chemistries, limiting discovery of truly novel transformations.
• Resource Demands: Computational costs rise with problem complexity and LLM API use, particularly for deep multi-step searches.
• Safety Risks: Predicted routes toward controlled or illicit compounds require augmented screening and policy constraints.

Proposed enhancements include:

• Domain-adapted chemical LLMs to curtail hallucinated chemistry
• Quantum-chemistry or energy-based filters for reaction feasibility
• RL-based learning of MCTS priors from successful plans
• Closed-loop integration with laboratory automation for real-time feedback and retraining
• Advanced novelty metrics (such as graph edit-distance to existing pathways) for prioritizing innovation

7. Significance and Outlook

DeepRetro constitutes a technologically significant advance in computer-aided synthesis planning, systematically combining the robustness of template-based search with the creative, adaptive reasoning of LLMs in an iterative and validated framework. The comprehensive screening pipeline and interface for human intervention position DeepRetro as a flexible platform for both automated and collaborative exploration of complex retrosynthetic landscapes, enabling the discovery and validation of synthesis strategies not directly encoded within template or training data (Sathyanarayana et al., 7 Jul 2025).