Atom-anchored LLMs speak Chemistry: A Retrosynthesis Demonstration

Alan Kai Hassen; Andrius Bernatavicius; Antonius P. A. Janssen; Djork-Arn\'e Clevert; Gerard J. P. van Westen; Mike Preuss

arxiv: 2510.16590 · v2 · pith:UWP2HW5Hnew · submitted 2025-10-18 · 💻 cs.LG · cs.AI· q-bio.BM

Atom-anchored LLMs speak Chemistry: A Retrosynthesis Demonstration

Alan Kai Hassen , Andrius Bernatavicius , Antonius P. A. Janssen , Mike Preuss , Gerard J. P. van Westen , Djork-Arn\'e Clevert This is my paper

Pith reviewed 2026-05-22 11:56 UTC · model grok-4.3

classification 💻 cs.LG cs.AIq-bio.BM

keywords large language modelsretrosynthesismolecular reasoningzero-shot promptingfew-shot promptingchemistrydrug discoverychain of thought

0 comments

The pith

General-purpose LLMs perform single-step retrosynthesis by anchoring reasoning to unique atomic identifiers without task-specific training

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper introduces a prompting framework that lets off-the-shelf large language models reason about molecules by tagging each atom with a unique identifier. This grounding allows the model to first identify relevant molecular fragments and their chemical transformation classes in a zero-shot step. An optional few-shot step then uses those position-aware labels to predict the actual reactants. On academic benchmarks and expert-reviewed drug molecules the method reaches at least 90 percent success at locating plausible reaction sites, 40 percent at naming reaction classes, and 74 percent at recovering the correct reactants. A reader would care because the approach removes the usual requirement for large labeled chemistry datasets or model fine-tuning.

Core claim

By assigning unique atomic identifiers to molecules, general-purpose LLMs can perform chain-of-thought reasoning that first locates chemically relevant fragments and their associated transformation classes and then, when supplied with class examples, predicts the reactant molecules that would produce the target in a single retrosynthetic step, achieving success rates of at least 90 percent for reaction-site identification, 40 percent for named reaction classes, and 74 percent for final reactant prediction across benchmarks and real drug-discovery compounds without any task-specific training.

What carries the argument

Atom-anchored chain-of-thought reasoning, in which unique atomic identifiers are attached to the molecular graph so the LLM can reference specific atoms when identifying fragments and transformation classes.

If this is right

LLMs can reach competitive performance on retrosynthesis benchmarks using only prompting rather than fine-tuning.
The same anchoring technique supplies a reusable blueprint for other molecular-reasoning tasks that involve identifying sites and transformations.
Drug-discovery workflows can incorporate general-purpose LLMs to propose reactant sets for newly designed target molecules.
Performance on named reaction classes improves when the second step supplies concrete class examples.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

The method could be chained across multiple steps to tackle full retrosynthetic routes by reusing the same atomic identifiers at each stage.
Because no domain-specific weights are updated, the approach lowers the barrier for labs that lack large GPU clusters to experiment with molecular reasoning.
Similar identifier-anchoring might be tested on other structured scientific objects such as protein binding sites or crystal lattices.

Load-bearing premise

That supplying unique atomic identifiers is enough for a general-purpose LLM to perform chemically valid molecular reasoning and transformation prediction without any chemistry-specific training or fine-tuning.

What would settle it

Running the same prompts on a fresh set of expert-validated molecules but with the atomic identifiers removed or randomized and observing whether reactant-prediction accuracy drops below 50 percent.

Figures

Figures reproduced from arXiv: 2510.16590 by Alan Kai Hassen, Andrius Bernatavicius, Antonius P. A. Janssen, Djork-Arn\'e Clevert, Gerard J. P. van Westen, Mike Preuss.

**Figure 2.** Figure 2: A) Position model performance on USPTO-LLM. The plot compares various foundation models on the task of reaction position prediction, measured by four evaluation metrics: achieving a partial positional match, maximizing the Jaccard metric, identifying the exact position, and predicting the correct reaction (conditional on a partial match). B) Confusion matrix of predicted versus ground-truth reaction classe… view at source ↗

**Figure 3.** Figure 3: Transition Model Performance on USPTO-LLM. The plot evaluates various LLMs on their ability to predict [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗

**Figure 4.** Figure 4: Zero-shot position model prediction for compound LEI-515 Jiang et al. [2023] using the PaRoutes reaction [PITH_FULL_IMAGE:figures/full_fig_p007_4.png] view at source ↗

**Figure 5.** Figure 5: Distribution of reaction names in the USPTO-50k test set. From this dataset, we created a balanced subsample [PITH_FULL_IMAGE:figures/full_fig_p014_5.png] view at source ↗

**Figure 6.** Figure 6: Confusion matrix of predicted versus ground-truth reaction names for the Gemini 2.5 Pro model. The [PITH_FULL_IMAGE:figures/full_fig_p016_6.png] view at source ↗

**Figure 7.** Figure 7: Performance difference between known and unknown reaction names. For unknown reactions, no equivalent name reaction examples within the USPTO50k training dataset are provided, illustrating the importance of the reaction name as a chemical anchor for retrieving reaction examples and chemical reasoning. Gemini 2.5 Pro (short no examples) Gemini 2.5 Pro (detailed no examples) Gemini 2.5 Pro (short with exampl… view at source ↗

**Figure 8.** Figure 8: An ablation study on the impact of prompt instruction detail and the inclusion of in-context examples on the performance of the Gemini 2.5 Pro transition model. We evaluate four settings: 1) a simple prompt without examples (see Appendix 3); 2) a detailed prompt without examples (see Appendix 2); 3) a simple prompt with examples; and 4) a detailed prompt with examples. 17 [PITH_FULL_IMAGE:figures/full_fig… view at source ↗

**Figure 9.** Figure 9: Confusion Matrix Highlighting the performance of different Transition Models on respective reaction name [PITH_FULL_IMAGE:figures/full_fig_p018_9.png] view at source ↗

**Figure 10.** Figure 10: Five real-world drug disovery molecules used in our case study: DH376 Deng et al. [2017], LEI-102 Li et al. [PITH_FULL_IMAGE:figures/full_fig_p019_10.png] view at source ↗

**Figure 11.** Figure 11: Correct reactant prediction for LEI-515 Jiang et al. [2023] by the Transition model (position priority 3, [PITH_FULL_IMAGE:figures/full_fig_p022_11.png] view at source ↗

**Figure 12.** Figure 12: Correct reactant template prediction for LEI-515 Jiang et al. [2023] by the Transition model (position priority [PITH_FULL_IMAGE:figures/full_fig_p023_12.png] view at source ↗

**Figure 13.** Figure 13: Correctly flagged chemically invalid prediction for LEI-515 Jiang et al. [2023] by the Transition Model [PITH_FULL_IMAGE:figures/full_fig_p024_13.png] view at source ↗

**Figure 14.** Figure 14: Syntactically invalid SMILES prediction for LEI-515 Jiang et al. [2023] by the Transition model (position [PITH_FULL_IMAGE:figures/full_fig_p025_14.png] view at source ↗

read the original abstract

Applications of machine learning in chemistry are often limited by the scarcity and expense of labeled data, restricting traditional supervised methods. In this work, we introduce a framework for molecular reasoning using general-purpose Large Language Models (LLMs) that operates without requiring task-specific model training. Our method anchors chain-of-thought reasoning to the molecular structure by using unique atomic identifiers. First, the LLM performs a zero-shot task to identify relevant fragments and their associated chemical labels or transformation classes. In an optional second step, this position-aware information is used in a few-shot task with provided class examples to predict the chemical transformation. We apply our framework to single-step retrosynthesis, a task where LLMs have previously underperformed. Across academic benchmarks and expert-validated drug discovery molecules, our work enables LLMs to achieve high success rates in identifying chemically plausible reaction sites ($\geq90\%$), named reaction classes ($\geq40\%$), and final reactants ($\geq74\%$). Ultimately, our work establishes a general blueprint for applying LLMs to challenges where molecular reasoning and molecular transformations are key, positioning atom-anchored LLMs as a powerful solution for data-scarce chemistry domains.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

Atom anchoring gives general LLMs a workable zero-shot path into retrosynthesis, but the missing anchored-vs-plain comparison leaves the mechanism unproven.

read the letter

The paper's main contribution is a prompting setup that adds unique atomic IDs to SMILES strings so an off-the-shelf LLM can first spot relevant fragments and labels in zero-shot mode, then optionally use few-shot examples to name the reaction class and output reactants. This is applied to single-step retrosynthesis, where prior LLM attempts have been weak, and it reports usable numbers across standard benchmarks plus some expert-validated drug molecules: at least 90% on site identification, 40% on class naming, and 74% on final reactant recovery. The approach avoids any task-specific training or fine-tuning, which is the practical angle for data-scarce chemistry work. That part is straightforward and worth noting. The two-step structure is also a clean way to separate structure awareness from reaction knowledge. The soft spot is the lack of a direct control. The manuscript does not show what happens when the same molecules and tasks are run with ordinary SMILES prompts instead of the anchored versions, so it is still possible the LLM is succeeding mainly on patterns it already saw in pre-training rather than because the identifiers are doing the anchoring work. Without that comparison the central claim stays harder to evaluate. The abstract also gives success rates without error bars, dataset sizes, or a clear description of how chemical plausibility was checked by experts, which makes the empirical side thin until the full methods are examined. This is aimed at groups already running LLM experiments on molecular tasks who need something that works with limited labels. A reader looking for prompting ideas in low-data chemistry will find a concrete recipe here. The work has a clear method and some external benchmarks, so it is worth sending to peer review even though the ablations and fuller validation details will need to be added.

Referee Report

2 major / 2 minor

Summary. The paper presents a framework for single-step retrosynthesis using general-purpose LLMs without task-specific fine-tuning. It anchors chain-of-thought reasoning to molecular structures via unique atomic identifiers, first identifying fragments and reaction classes in zero-shot mode, then optionally using few-shot examples to predict transformations. The central empirical claim is that this enables LLMs to achieve ≥90% success on chemically plausible reaction sites, ≥40% on named reaction classes, and ≥74% on final reactants across academic benchmarks and expert-validated drug-discovery molecules.

Significance. If the reported success rates are reproducible and the atomic-anchoring mechanism is shown to be causal, the work would supply a practical, training-free blueprint for LLM-based molecular reasoning in data-scarce chemistry domains. The absence of any free parameters or invented entities in the reported approach is a methodological strength that distinguishes it from supervised retrosynthesis models.

major comments (2)

[Methods / Results] The manuscript provides no ablation that isolates the effect of the unique atomic identifiers. A direct comparison of anchored versus standard SMILES prompts on identical molecules and tasks is required to establish that the identifiers, rather than pre-trained reaction knowledge or prompt phrasing, drive the reported accuracies (≥90% sites, ≥40% classes, ≥74% reactants). Without this control the central claim that atom-anchoring enables chemically valid zero-shot/few-shot retrosynthesis remains unproven.
[Abstract / Results] The abstract and results sections state success rates but supply no information on experimental protocol, baseline comparisons, error bars, dataset sizes, or how chemically plausible reactants were validated by experts. These omissions prevent assessment of whether the empirical claims are statistically supported or reproducible.

minor comments (2)

[Methods] Clarify the exact format of the atomic identifiers and how they are injected into the LLM prompt (e.g., SMILES augmentation syntax).
[Methods] Specify the number of few-shot examples used in the optional second step and whether performance varies with example count.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive feedback. The comments help clarify the need for stronger evidence on the role of atomic identifiers and for greater transparency in reporting experimental details. We respond to each major comment below and will revise the manuscript to address them where possible.

read point-by-point responses

Referee: [Methods / Results] The manuscript provides no ablation that isolates the effect of the unique atomic identifiers. A direct comparison of anchored versus standard SMILES prompts on identical molecules and tasks is required to establish that the identifiers, rather than pre-trained reaction knowledge or prompt phrasing, drive the reported accuracies (≥90% sites, ≥40% classes, ≥74% reactants). Without this control the central claim that atom-anchoring enables chemically valid zero-shot/few-shot retrosynthesis remains unproven.

Authors: We agree that an explicit ablation comparing anchored prompts against standard SMILES prompts on the same molecules and tasks would strengthen the causal argument. The current manuscript demonstrates the overall framework performance but does not include this head-to-head control. In the revised version we will add the requested ablation across the benchmark sets and expert-validated molecules, reporting success rates for both zero-shot fragment identification and few-shot reactant prediction. This addition will directly test whether the identifiers, rather than general pre-trained knowledge, account for the observed accuracies. revision: yes
Referee: [Abstract / Results] The abstract and results sections state success rates but supply no information on experimental protocol, baseline comparisons, error bars, dataset sizes, or how chemically plausible reactants were validated by experts. These omissions prevent assessment of whether the empirical claims are statistically supported or reproducible.

Authors: The full manuscript contains a Methods section that specifies the datasets (academic benchmarks and expert-curated drug-like molecules), the number of molecules evaluated, the expert validation protocol for chemical plausibility, and the prompting procedure. Error bars and statistical details appear in the Results figures and tables. To address the referee’s concern about accessibility, we will revise the abstract to include a concise statement of dataset scale and validation approach, and we will add explicit cross-references in the Results section to the Methods details on reproducibility. We will also ensure baseline comparisons (where performed) and error reporting are highlighted more clearly. revision: partial

Circularity Check

0 steps flagged

No significant circularity: empirical demonstration on external benchmarks

full rationale

The paper introduces an atom-anchored prompting framework for zero-shot and few-shot retrosynthesis with general-purpose LLMs and reports empirical success rates (≥90% reaction sites, ≥40% named classes, ≥74% reactants) on academic benchmarks plus expert-validated drug discovery molecules. These metrics are measured against independent external test sets rather than quantities defined by the authors' own fitted parameters or self-referential equations. No derivation chain, mathematical ansatz, or uniqueness theorem is presented that reduces the central claims to inputs by construction. The work is self-contained as an empirical demonstration; any self-citations that may exist are not load-bearing for the reported accuracies.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

The central claim rests on the premise that atomic identifiers enable reliable chemical reasoning in general LLMs; no free parameters, invented entities, or additional axioms are stated in the abstract.

axioms (1)

domain assumption General-purpose LLMs can perform chemically valid molecular reasoning when chain-of-thought is anchored to unique atomic identifiers
This premise is invoked to justify the zero-shot and few-shot stages without task-specific training.

pith-pipeline@v0.9.0 · 5767 in / 1240 out tokens · 39165 ms · 2026-05-22T11:56:41.674170+00:00 · methodology

discussion (0)

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

IndisputableMonolith/Cost/FunctionalEquation.lean washburn_uniqueness_aczel unclear

?

unclear
Relation between the paper passage and the cited Recognition theorem.

Our method anchors chain-of-thought reasoning to the molecular structure by using unique atomic identifiers... zero-shot task to identify relevant fragments and their associated chemical labels or transformation classes... few-shot task with provided class examples to predict the chemical transformation.

What do these tags mean?

matches: The paper's claim is directly supported by a theorem in the formal canon.
supports: The theorem supports part of the paper's argument, but the paper may add assumptions or extra steps.
extends: The paper goes beyond the formal theorem; the theorem is a base layer rather than the whole result.
uses: The paper appears to rely on the theorem as machinery.
contradicts: The paper's claim conflicts with a theorem or certificate in the canon.
unclear: Pith found a possible connection, but the passage is too broad, indirect, or ambiguous to say the theorem truly supports the claim.

Reference graph

Works this paper leans on

19 extracted references · 19 canonical work pages · 1 internal anchor

[1]

DeepSeek-R1: Incentivizing Reasoning Capability in LLMs via Reinforcement Learning

ISSN 1758-2946. doi:10.1186/s13321-025-00986-6. Seongmin Kim, Yousung Jung, and Joshua Schrier. Large Language Models for Inorganic Synthesis Predictions, June 2024. Joseph M. Cavanagh, Kunyang Sun, Andrew Gritsevskiy, Dorian Bagni, Thomas D. Bannister, and Teresa Head-Gordon. SmileyLlama: Modifying Large Language Models for Directed Chemical Space Explor...

work page internal anchor Pith review Pith/arXiv arXiv doi:10.1186/s13321-025-00986-6 2024
[2]

Thinking

ISSN 0022-2623, 1520-4804. doi:10.1021/acs.jmedchem.6b01482. Xiaoting Li, Hao Chang, Jara Bouma, Laura V . De Paus, Partha Mukhopadhyay, Janos Paloczi, Mohammed Mustafa, Cas Van Der Horst, Sanjay Sunil Kumar, Lijie Wu, Yanan Yu, Richard J. B. H. N. Van Den Berg, Antonius P. A. Janssen, Aron Lichtman, Zhi-Jie Liu, Pal Pacher, Mario Van Der Stelt, Laura H. ...

work page doi:10.1021/acs.jmedchem.6b01482 2023
[3]

disconnection

**Step 1: Identify All Candidate Transformations** 15Process steps A - L sequentially. For each step, you must perform a complete and independent analysis to identify all transformations that fit its description. A finding in one step does not exclude findings in others. 16* **Input:** The ‘product_smiles ‘. 17* **Process:** 18* A) **Symmetry Analysis:** ...

work page
[4]

disconnection

**Step 2: Assign Candidate Reactions** 35* **Input:** The list of transformation strings from Step 1. 36* **Process:** For each transformation, determine all appropriate forward reaction names. A single transformation may have multiple corresponding reactions. 37* **Output (Internal):** A list of objects, where each object contains a transformation and a ...

work page
[5]

42* **Process:** Expand the input into a flat list by creating a **new, separate entry for each reaction** associated with a transformation

**Step 3: Expand and Evaluate Pairs** 41* **Input:** The list of objects from Step 2. 42* **Process:** Expand the input into a flat list by creating a **new, separate entry for each reaction** associated with a transformation. Then, for each of these new entries, apply the Retrosynthetic Analysis Framework to assign a ‘Retrosynthesis Importance‘ value and...

work page
[6]

isInOntology

**Step 4: Final Formatting and Priority Assignment** 46* **Input:** The flat list of objects from Step 3. 47* **Process:** For each object, format it according to the ‘Constraints & Formatting Rules‘. Then, calculate a ‘ Priority‘ number for each entry by ranking them based on two criteria: 1. ‘"isInOntology"‘ (‘true‘ before ‘false ‘), and 2. ‘"Retrosynth...

work page
[7]

23* **Process:** If a ‘forward_reaction_name‘ is provided, use it as the sole reaction

**Step 1: Determine Reaction(s) to Model** 22* **Input:** The ‘forward_reaction_name‘ (optional) and ‘reaction_center_atoms‘ from the user. 23* **Process:** If a ‘forward_reaction_name‘ is provided, use it as the sole reaction. If not, analyze the ‘ reaction_center_atoms‘ to generate a list of potential ‘forward_reaction_name‘s. 24* **Output (Internal):**...

work page
[8]

28* **Process:** For each ‘forward_reaction_name‘, use your expert chemical knowledge and the provided examples to determine the **precise and complete reaction center**

**Step 2: Refine Reaction Center** 27* **Input:** The list of ‘forward_reaction_name‘s (Step 1), the users ‘reaction_center_atoms‘, and any ‘ retrosynthesis_reaction_examples ‘. 28* **Process:** For each ‘forward_reaction_name‘, use your expert chemical knowledge and the provided examples to determine the **precise and complete reaction center**. The user...

work page
[9]

template

**Step 3: Extract Atom-Level Reaction Template** 32* **Input:** The list of ‘forward_reaction_name‘s from Step 1, the **precise reaction center** from step 2, and the user-provided ‘retrosynthesis_reaction_examples ‘. 33* **Process:** For each ‘forward_reaction_name‘, analyze its corresponding valid example(s). Your primary goal is to extract the **struct...

work page
[10]

**Step 4: Generate Precursor Molecule(s)** 95* **Input:** The ‘product_smiles‘ and ‘precise_reaction_center_atoms ‘. 96* **Process:** Based on the number of fragments implied by the transformation type (e.g., two for an intermolecular disconnection, one for an FGI, three for a 3-component MCR), generate the corresponding core precursor molecule(s ). This ...

work page
[11]

101* **Process:** For each reactions template, apply the extracted retrosynthetic template to the precursor(s)

**Step 5: Apply Reaction Template to Generate Reactant Permutations** 100* **Input:** The precursor(s) (Step 4) and the reaction templates (Step 3). 101* **Process:** For each reactions template, apply the extracted retrosynthetic template to the precursor(s). The ‘ precise_reaction_center_atoms‘ provided by the user defines the **locality** of the transf...

work page
[12]

**Fragment-Role Permutations:** For a disconnection into multiple fragments with distinct reactive groups, you must generate reactant sets for **all** possible assignments of those groups to the fragments

work page
[13]

organohalide,

**Intra-Group Class Permutations:** If a generated reactive group belongs to a general chemical class (e.g., an "organohalide," "leaving␣group," or "protecting␣group"), you are required to generate an exhaustive list of separate options for **all chemically distinct members of that class known to be compatible with the reaction.** 104The model is **explic...

work page
[14]

109* **Process:** For each generated option, perform a rigorous chemical validation

**Step 6: Validate and Justify Each Option** 108* **Input:** The list of potential reactant options from Step 5. 109* **Process:** For each generated option, perform a rigorous chemical validation. 110* A) **Stability:** Are the proposed reactants chemically stable? 111* B) **Chemoselectivity:** Would the reaction be selective? Are there other functional ...

work page
[15]

**Group Options:** Begin by grouping the list of validated options by their ‘forward_reaction_name ‘

work page
[16]

It signals that a member of this chemical class (e.g

**Extract Wildcard Reaction Class** Looking at the validated options and their reaction names, you must deduct a general reaction class template if possible using the ‘<CLASS:..>‘ tag. It signals that a member of this chemical class (e.g. ‘<CLASS:AmineProtectingGroup >‘) should be used instead of an explicit molecular structure

work page
[17]

This object serves as the general, machine-readable representation for the entire transformation class and should be placed at the **beginning** of the ‘reactant_permutations‘ list

**Generate General Template Entry (if applicable):** For each extracted general reaction class template, you ** should** create one additional, special permutation object derived from the two provided general reaction classes . This object serves as the general, machine-readable representation for the entire transformation class and should be placed at th...

work page
[18]

This list will now contain the general template entry at the top (if applicable), followed by all the validated, specific examples from Step 6

**Assemble Final List:** For each unique reaction, create a single object containing the ‘forward_reaction_name‘ and its final ‘reactant_permutations‘ list. This list will now contain the general template entry at the top (if applicable), followed by all the validated, specific examples from Step 6

work page
[19]

product":

**Finalize and Clean:** Assemble these grouped objects into the final ‘reaction_analysis‘ list according to the ‘ Output Schema‘. Keep the original atom mapping of the product where possible and do not introduce new atom maps on the reactant side, but use unmapped atoms. 131* **Output:** The final, single JSON object. 132 133**Output Schema - Strict JSON ...

work page

[1] [1]

DeepSeek-R1: Incentivizing Reasoning Capability in LLMs via Reinforcement Learning

ISSN 1758-2946. doi:10.1186/s13321-025-00986-6. Seongmin Kim, Yousung Jung, and Joshua Schrier. Large Language Models for Inorganic Synthesis Predictions, June 2024. Joseph M. Cavanagh, Kunyang Sun, Andrew Gritsevskiy, Dorian Bagni, Thomas D. Bannister, and Teresa Head-Gordon. SmileyLlama: Modifying Large Language Models for Directed Chemical Space Explor...

work page internal anchor Pith review Pith/arXiv arXiv doi:10.1186/s13321-025-00986-6 2024

[2] [2]

Thinking

ISSN 0022-2623, 1520-4804. doi:10.1021/acs.jmedchem.6b01482. Xiaoting Li, Hao Chang, Jara Bouma, Laura V . De Paus, Partha Mukhopadhyay, Janos Paloczi, Mohammed Mustafa, Cas Van Der Horst, Sanjay Sunil Kumar, Lijie Wu, Yanan Yu, Richard J. B. H. N. Van Den Berg, Antonius P. A. Janssen, Aron Lichtman, Zhi-Jie Liu, Pal Pacher, Mario Van Der Stelt, Laura H. ...

work page doi:10.1021/acs.jmedchem.6b01482 2023

[3] [3]

disconnection

**Step 1: Identify All Candidate Transformations** 15Process steps A - L sequentially. For each step, you must perform a complete and independent analysis to identify all transformations that fit its description. A finding in one step does not exclude findings in others. 16* **Input:** The ‘product_smiles ‘. 17* **Process:** 18* A) **Symmetry Analysis:** ...

work page

[4] [4]

disconnection

**Step 2: Assign Candidate Reactions** 35* **Input:** The list of transformation strings from Step 1. 36* **Process:** For each transformation, determine all appropriate forward reaction names. A single transformation may have multiple corresponding reactions. 37* **Output (Internal):** A list of objects, where each object contains a transformation and a ...

work page

[5] [5]

42* **Process:** Expand the input into a flat list by creating a **new, separate entry for each reaction** associated with a transformation

**Step 3: Expand and Evaluate Pairs** 41* **Input:** The list of objects from Step 2. 42* **Process:** Expand the input into a flat list by creating a **new, separate entry for each reaction** associated with a transformation. Then, for each of these new entries, apply the Retrosynthetic Analysis Framework to assign a ‘Retrosynthesis Importance‘ value and...

work page

[6] [6]

isInOntology

**Step 4: Final Formatting and Priority Assignment** 46* **Input:** The flat list of objects from Step 3. 47* **Process:** For each object, format it according to the ‘Constraints & Formatting Rules‘. Then, calculate a ‘ Priority‘ number for each entry by ranking them based on two criteria: 1. ‘"isInOntology"‘ (‘true‘ before ‘false ‘), and 2. ‘"Retrosynth...

work page

[7] [7]

23* **Process:** If a ‘forward_reaction_name‘ is provided, use it as the sole reaction

**Step 1: Determine Reaction(s) to Model** 22* **Input:** The ‘forward_reaction_name‘ (optional) and ‘reaction_center_atoms‘ from the user. 23* **Process:** If a ‘forward_reaction_name‘ is provided, use it as the sole reaction. If not, analyze the ‘ reaction_center_atoms‘ to generate a list of potential ‘forward_reaction_name‘s. 24* **Output (Internal):**...

work page

[8] [8]

28* **Process:** For each ‘forward_reaction_name‘, use your expert chemical knowledge and the provided examples to determine the **precise and complete reaction center**

**Step 2: Refine Reaction Center** 27* **Input:** The list of ‘forward_reaction_name‘s (Step 1), the users ‘reaction_center_atoms‘, and any ‘ retrosynthesis_reaction_examples ‘. 28* **Process:** For each ‘forward_reaction_name‘, use your expert chemical knowledge and the provided examples to determine the **precise and complete reaction center**. The user...

work page

[9] [9]

template

**Step 3: Extract Atom-Level Reaction Template** 32* **Input:** The list of ‘forward_reaction_name‘s from Step 1, the **precise reaction center** from step 2, and the user-provided ‘retrosynthesis_reaction_examples ‘. 33* **Process:** For each ‘forward_reaction_name‘, analyze its corresponding valid example(s). Your primary goal is to extract the **struct...

work page

[10] [10]

**Step 4: Generate Precursor Molecule(s)** 95* **Input:** The ‘product_smiles‘ and ‘precise_reaction_center_atoms ‘. 96* **Process:** Based on the number of fragments implied by the transformation type (e.g., two for an intermolecular disconnection, one for an FGI, three for a 3-component MCR), generate the corresponding core precursor molecule(s ). This ...

work page

[11] [11]

101* **Process:** For each reactions template, apply the extracted retrosynthetic template to the precursor(s)

**Step 5: Apply Reaction Template to Generate Reactant Permutations** 100* **Input:** The precursor(s) (Step 4) and the reaction templates (Step 3). 101* **Process:** For each reactions template, apply the extracted retrosynthetic template to the precursor(s). The ‘ precise_reaction_center_atoms‘ provided by the user defines the **locality** of the transf...

work page

[12] [12]

**Fragment-Role Permutations:** For a disconnection into multiple fragments with distinct reactive groups, you must generate reactant sets for **all** possible assignments of those groups to the fragments

work page

[13] [13]

organohalide,

**Intra-Group Class Permutations:** If a generated reactive group belongs to a general chemical class (e.g., an "organohalide," "leaving␣group," or "protecting␣group"), you are required to generate an exhaustive list of separate options for **all chemically distinct members of that class known to be compatible with the reaction.** 104The model is **explic...

work page

[14] [14]

109* **Process:** For each generated option, perform a rigorous chemical validation

**Step 6: Validate and Justify Each Option** 108* **Input:** The list of potential reactant options from Step 5. 109* **Process:** For each generated option, perform a rigorous chemical validation. 110* A) **Stability:** Are the proposed reactants chemically stable? 111* B) **Chemoselectivity:** Would the reaction be selective? Are there other functional ...

work page

[15] [15]

**Group Options:** Begin by grouping the list of validated options by their ‘forward_reaction_name ‘

work page

[16] [16]

It signals that a member of this chemical class (e.g

**Extract Wildcard Reaction Class** Looking at the validated options and their reaction names, you must deduct a general reaction class template if possible using the ‘<CLASS:..>‘ tag. It signals that a member of this chemical class (e.g. ‘<CLASS:AmineProtectingGroup >‘) should be used instead of an explicit molecular structure

work page

[17] [17]

This object serves as the general, machine-readable representation for the entire transformation class and should be placed at the **beginning** of the ‘reactant_permutations‘ list

**Generate General Template Entry (if applicable):** For each extracted general reaction class template, you ** should** create one additional, special permutation object derived from the two provided general reaction classes . This object serves as the general, machine-readable representation for the entire transformation class and should be placed at th...

work page

[18] [18]

This list will now contain the general template entry at the top (if applicable), followed by all the validated, specific examples from Step 6

**Assemble Final List:** For each unique reaction, create a single object containing the ‘forward_reaction_name‘ and its final ‘reactant_permutations‘ list. This list will now contain the general template entry at the top (if applicable), followed by all the validated, specific examples from Step 6

work page

[19] [19]

product":

**Finalize and Clean:** Assemble these grouped objects into the final ‘reaction_analysis‘ list according to the ‘ Output Schema‘. Keep the original atom mapping of the product where possible and do not introduce new atom maps on the reactant side, but use unmapped atoms. 131* **Output:** The final, single JSON object. 132 133**Output Schema - Strict JSON ...

work page