Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

SynCraft: Guiding Large Language Models to Predict Edit Sequences for Molecular Synthesizability Optimization

Junren Li, Luhua Lai

TL;DR

SynCraft addresses the gap between computational design and practical synthesis by reframing synthesizability optimization as targeted, atom-level editing guided by chain-of-thought reasoning in LLMs. It decouples strategic planning from chemical execution, using a Retrieval-Augmented Inference pipeline and an interaction-aware prompting strategy to preserve pharmacophores while navigating the synthesis cliff with surgical edits. Across large-scale benchmarks, SynCraft outperforms projection-based baselines, replicates expert medicinal chemistry intuition in retrospective PLK1 cases, and demonstrates prospective rescue of shelved RIPK1 candidates through bioisosteric edits and scaffold hopping. The work delivers explainable, chemistry-grounded AI edits with practical implications for high-value hit optimization and suggests a scalable path toward manufacturable AI-designed molecules.

Abstract

Generative artificial intelligence has revolutionized the exploration of chemical space, yet a critical bottleneck remains that a substantial fraction of generated molecules is synthetically inaccessible. Current solutions, such as post-hoc filtering or projection-based methods, often compromise structural novelty or disrupt key pharmacophores by forcing molecules into pre-defined synthetic templates. Herein, we introduce SynCraft, a reasoning-based framework that reframes synthesizability optimization not as a sequence translation task, but as a precise structural editing problem. Leveraging the emergent reasoning capabilities of Large Language Models, SynCraft navigates the "synthesis cliff" where minimal structural modifications yield significant gains in synthetic feasibility. By predicting executable sequences of atom-level edits rather than generating SMILES strings directly, SynCraft circumvents the syntactic fragility of LLMs while harnessing their chemical intuition. Extensive benchmarks demonstrate that SynCraft outperforms state-of-the-art baselines in generating synthesizable analogs with high structural fidelity. Furthermore, through interaction-aware prompting, SynCraft successfully replicates expert medicinal chemistry intuition in editing PLK1 inhibitors and rescuing high-scoring but previously discarded RIPK1 candidates in previous molecular generation literatures.

SynCraft: Guiding Large Language Models to Predict Edit Sequences for Molecular Synthesizability Optimization

TL;DR

SynCraft addresses the gap between computational design and practical synthesis by reframing synthesizability optimization as targeted, atom-level editing guided by chain-of-thought reasoning in LLMs. It decouples strategic planning from chemical execution, using a Retrieval-Augmented Inference pipeline and an interaction-aware prompting strategy to preserve pharmacophores while navigating the synthesis cliff with surgical edits. Across large-scale benchmarks, SynCraft outperforms projection-based baselines, replicates expert medicinal chemistry intuition in retrospective PLK1 cases, and demonstrates prospective rescue of shelved RIPK1 candidates through bioisosteric edits and scaffold hopping. The work delivers explainable, chemistry-grounded AI edits with practical implications for high-value hit optimization and suggests a scalable path toward manufacturable AI-designed molecules.

Abstract

Generative artificial intelligence has revolutionized the exploration of chemical space, yet a critical bottleneck remains that a substantial fraction of generated molecules is synthetically inaccessible. Current solutions, such as post-hoc filtering or projection-based methods, often compromise structural novelty or disrupt key pharmacophores by forcing molecules into pre-defined synthetic templates. Herein, we introduce SynCraft, a reasoning-based framework that reframes synthesizability optimization not as a sequence translation task, but as a precise structural editing problem. Leveraging the emergent reasoning capabilities of Large Language Models, SynCraft navigates the "synthesis cliff" where minimal structural modifications yield significant gains in synthetic feasibility. By predicting executable sequences of atom-level edits rather than generating SMILES strings directly, SynCraft circumvents the syntactic fragility of LLMs while harnessing their chemical intuition. Extensive benchmarks demonstrate that SynCraft outperforms state-of-the-art baselines in generating synthesizable analogs with high structural fidelity. Furthermore, through interaction-aware prompting, SynCraft successfully replicates expert medicinal chemistry intuition in editing PLK1 inhibitors and rescuing high-scoring but previously discarded RIPK1 candidates in previous molecular generation literatures.
Paper Structure (20 sections, 1 equation, 4 figures, 1 table)

This paper contains 20 sections, 1 equation, 4 figures, 1 table.

Table of Contents

  1. Introduction
  2. Methods
  3. Dataset Construction and Definition of the Synthesis Cliff
  4. The SynCraft Framework: Generative Editing via In-Context Reasoning
  5. Discrete Editing Action Space
  6. Constructing the Reference Dataset
  7. Retrieval-Augmented Inference
  8. Interaction-Aware Prompting Strategy
  9. Results and discussion
  10. Quantitative Evaluation of Synthesizability Optimization
  11. SynCraft Navigates the Synthesis Cliff via Bioisosteric Editing while Preserving Pharmacophores
  12. Navigate the Synthesis Cliff via Bioisosteric Editing
  13. Preserving Pharmacophoric Integrity
  14. Retrospective Validation: Replicating Expert Intuition in Hit Optimization
  15. Prospective Rescue of Shelved Candidates: Interaction-Aware Optimization of High-Scoring Designs
  16. ...and 5 more sections

Figures (4)

  • Figure 1: The SynCraft framework. (a) Concept of the synthesis cliff. Generative models often produce molecules in the "unsynthesizable highlands." SynCraft navigates to the accessible "lowlands" not by regenerating the whole molecule, but by predicting a precise sequence of discrete edits (e.g., DEL_ATOM, ADD_BOND), transforming the target into a synthesizable analog. (b) The workflow of SynCraft. The process begins by retrieving similar "unsynthesizable-to-synthesizable" pairs from a pre-constructed Synthesis Cliff Dataset. These pairs serve as few-shot demonstrations for the Large Language Model (Gemini-2.5-Pro). The model operates via a Chain-of-Thought (CoT) mechanism: first reasoning about synthetic liabilities and (optional) biological constraints derived from PLIP interaction analysis, and then generating an executable edit sequence. This sequence is deterministically applied to the input molecule to yield the final optimized candidate.
  • Figure 2: Qualitative comparison of structural modifications by SynCraft versus projection-based baselines. (a) Navigating the synthesis cliff via bioisosteric editing. The input molecule features a complex fused carbocycle. SynCraft (bottom path) strategically replaces a methylene group with an oxygen atom, creating a synthetically accessible ether linkage while preserving the scaffold shape. In contrast, projection-based baselines converge on a strained benzocyclobutene-like motif, likely an artifact of projecting into a sparse region of the synthesizable chemical space. (b, c) Preserving pharmacophoric integrity against oversimplification. (b) For a nitrogen-containing ring, the baseline (ReaSyn) discards the nitrogen atom, losing a potential interaction site. SynCraft aromatizes the ring to a pyridine derivative, stabilizing the structure while preserving the heteroatom. (c) For a fused bicyclic core, the baseline breaks the scaffold into simpler fragments (scaffold hopping), potentially disrupting shape complementarity.
  • Figure 3: Retrospective validation of SynCraft’s reasoning capability against human expert intuition in TransPharmerxie2025transpharmer. (a) The initial generated molecule (lig-886) contains a 2,5-dimethylpiperazine ring, which introduces two chiral centers and poses a risk of forming complex stereoisomeric mixtures. (b) SynCraft correctly identifies this synthetic liability in its reasoning trace, explicitly stating the need to remove stereochemical ambiguity. (c) The edit sequence generated by SynCraft removes the two methyl groups on the carbon skeleton. (d) The actual chemically synthesized lead compound from the original study (IIP0944). SynCraft’s decision to simplify the core to an achiral piperazine converges precisely with the strategy adopted by human experts to ensure developability.
  • Figure 4: Prospective rescue of shelved high-scoring drug candidates targeting RIPK1. (a) Case I: SynCraft introduces an ether linker to replace a challenging C-C bond between two electron-deficient aromatic rings. This modification facilitates synthesis via C-O bond formation (proposed retrosynthetic pathway shown below) while maintaining the binding mode (Vina score: -10.7 to -10.2 kcal/mol. (b) Case II: SynCraft replaces a complex, non-planar fused polycyclic system containing a quaternary carbon with a planar, synthetically modular 2,4-disubstituted pyrimidine scaffold (Vina score: -10.2 to -9.9 kcal/mol). Green sticks: Original molecule; Magenta sticks: SynCraft-optimized molecule; Blue lines: Hydrogen bonds; Grey dashed lines: Hydrophobic interactions.