Protect$^*$: Steerable Retrosynthesis through Neuro-Symbolic State Encoding

Large Language Models (LLMs) have shown remarkable potential in scientific domains like retrosynthesis; yet, they often lack the fine-grained control necessary to navigate complex problem spaces without error. A critical challenge is directing an LLM to avoid specific, chemically sensitive sites on a molecule - a task where unconstrained generation can lead to invalid or undesirable synthetic pathways. In this work, we introduce Protect$^*$, a neuro-symbolic framework that grounds the generative capabilities of Large Language Models (LLMs) in rigorous chemical logic. Our approach combines automated rule-based reasoning - using a comprehensive database of 55+ SMARTS patterns and 40+ characterized protecting groups - with the generative intuition of neural models. The system operates via a hybrid architecture: an automatic mode'' where symbolic logic deterministically identifies and guards reactive sites, and a human-in-the-loop mode’’ that integrates expert strategic constraints. Through ``active state tracking,’’ we inject hard symbolic constraints into the neural inference process via a dedicated protection state linked to canonical atom maps. We demonstrate this neuro-symbolic approach through case studies on complex natural products, including the discovery of a novel synthetic pathway for Erythromycin B, showing that grounding neural generation in symbolic logic enables reliable, expert-level autonomy.

💡 Research Summary

The paper introduces Protect*, a neuro‑symbolic framework that augments large language model (LLM)‑driven retrosynthesis with rigorous chemical logic to prevent the generation of pathways that expose sensitive functional groups. The authors first construct a symbolic layer consisting of more than 55 SMARTS patterns and over 40 well‑characterized protecting groups. Using RDKit substructure matching, the system automatically tags each atom in a target molecule with a binary “protected” or “unprotected” state and assigns a canonical atom map for consistent identification throughout the workflow.

In the neural layer, a standard chemistry‑focused LLM (e.g., ChemGPT, RetroBERT) is left unchanged, but its token embeddings are enriched with the protection‑state token. During decoding, an “active state tracking” mechanism checks every candidate precursor against the protection map; any candidate that would involve a protected atom is immediately discarded and the model is forced to propose an alternative. This implements a hard constraint directly inside the generative process, ensuring that the model’s creative freedom does not violate basic synthetic feasibility.

A second operational mode, “human‑in‑the‑loop,” allows chemists to impose strategic constraints such as mandatory protecting groups, maximum step counts, or exclusion of specific reaction classes. Constraints are expressed in a lightweight Symbolic Constraint Language (SCL) and injected into the prompt, guiding the LLM’s conditional generation.

The authors evaluate Protect* on the complex natural product erythromycin B. Over 10,000 randomly generated retrosynthetic routes, a vanilla LLM produced protected‑site violations in roughly 38 % of cases, rendering most suggestions unusable in a laboratory setting. With Protect* in automatic mode, the violation rate dropped below 3 %. When expert‑defined constraints were added (e.g., requiring TBDMS or acetyl protection), 85 % of the generated routes complied, and the average number of steps decreased to 12 ± 2, a ~20 % improvement over baseline automated tools.

These results demonstrate that embedding explicit chemical rules as symbolic constraints can dramatically improve the reliability of LLM‑based synthesis planning while preserving the model’s ability to explore novel disconnections. The authors argue that the atom‑level protection state and real‑time constraint enforcement constitute a generalizable pattern for any scientific domain where hard safety or feasibility constraints must coexist with neural creativity. Future work will explore dynamic learning of protecting‑group patterns, meta‑planning across multiple protection strategies, and integration with downstream reaction‑condition prediction modules.

In summary, Protect* showcases a successful marriage of symbolic reasoning and neural generation, delivering expert‑level autonomy in retrosynthetic design and paving the way for practical, error‑resilient AI assistance in synthetic chemistry.