Problem

Generative artificial intelligence in drug discovery frequently designs molecules that are highly optimized computationally but prove challenging or impossible to synthesize, creating a significant bottleneck. Current approaches, such as post-hoc filtering, often discard a large number of potentially valuable compounds. Alternatively, methods that constrain generation by using predefined building blocks or reaction templates compromise the exploratory power of generative models and limit the accessible chemical space, failing to make precise, minimal modifications needed to overcome synthetic difficulties.

Method

SynCraft redefines synthesizability optimization as a precise structural editing problem, utilizing Large Language Models (LLMs) like Gemini-2.5-Pro. Instead of generating complete molecular representations (SMILES strings), SynCraft guides the LLM through a Chain-of-Thought (CoT) prompting strategy to first reason about synthetic liabilities and then formulate a precise sequence of atom-level edits (e.g., DEL_ATOM, ADD_BOND). These instructions are then executed by a deterministic chemical toolkit, ensuring chemical validity. The framework employs a Retrieval-Augmented Generation (RAG) paradigm, where few-shot examples from a curated "Synthesis Cliff" dataset inform the LLM's reasoning. Furthermore, SynCraft integrates an interaction-aware constraint mechanism that translates 3D protein-ligand interaction data into natural language prompts, instructing the LLM to preserve critical pharmacophores during optimization.

Results

SynCraft significantly outperformed state-of-the-art projection-based baselines (ChemProjector, SynFormer, ReaSyn) in generating synthesizable analogs while maintaining high structural similarity to the original molecules. For example, on the Pocket2Mol dataset, SynCraft achieved a success rate of 42.7% at a Tanimoto similarity threshold of 0.5, surpassing the best baseline by over 12 percentage points. At a stricter threshold of 0.6, SynCraft's success rate (28.4%) nearly doubled that of the baselines. Ablation studies confirmed that the edit-based paradigm dramatically reduced structural hallucination compared to direct SMILES generation. Qualitatively, SynCraft demonstrated "surgical" editing capabilities, applying local fixes such as heteroatom substitutions that resolved synthetic bottlenecks without drastic scaffold changes, unlike projection methods which sometimes yielded strained or chemically idiosyncratic motifs. Case studies validated SynCraft's ability to replicate human medicinal chemistry intuition in optimizing PLK1 inhibitors by addressing stereochemical liabilities and successfully "rescuing" high-scoring but previously shelved RIPK1 inhibitor candidates through structure-aware modifications, preserving critical binding affinity and interactions.

Implications

SynCraft provides a crucial bridge between virtual chemical design and physical reality by approaching synthesizability optimization as a strategic reasoning task. By decoupling strategic planning from chemical execution, it effectively addresses the challenges of syntactic validity and semantic preservation inherent in traditional generative models. The framework offers surgically precise navigation of the "synthesis cliff," demonstrates the capacity to replicate expert medicinal chemistry intuition, and has the potential to rescue valuable but synthetically complex "orphan" designs. SynCraft's generation of human-readable rationales for its modifications fosters trust and collaboration between AI agents and human chemists, suggesting that such reasoning-centric LLM frameworks will become indispensable tools in the drug discovery pipeline for designing molecules that are not only potent and novel but also readily manufacturable. While highly effective for high-value hit optimization, the current inference costs make it less suitable for mass enumeration of large virtual libraries.