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

Protect* is a neuro-symbolic framework that enhances the reliability of large language models in retrosynthesis by integrating automated rule-based logic and expert constraints to steer generative processes away from chemically sensitive sites, enabling the discovery of valid and novel synthetic pathways for complex molecules like Erythromycin B.

The Gist

Imagine you are hiring a brilliant, incredibly creative chef (the Large Language Model or LLM) to cook a very complex, multi-course meal (a chemical synthesis). This chef has read every cookbook in the world and can invent amazing new recipes on the fly.

However, there's a problem: the chef is so eager to be creative that they sometimes ignore the specific instructions you gave them. You might say, "Don't touch the garlic; it's too spicy for this dish," but the chef, in their excitement, might accidentally chop the garlic anyway, ruining the flavor. In chemistry, this is like a computer trying to build a molecule but accidentally reacting with the wrong part, creating a mess instead of the medicine you wanted.

This paper introduces Protect*, a new "smart kitchen manager" designed to help the chef cook perfectly without making mistakes.

The Problem: The "Over-Enthusiastic Chef"

In the world of making medicines (retrosynthesis), computers use AI to figure out how to build complex molecules step-by-step. The current AI is great at brainstorming ideas, but it struggles with fine-grained control. It often forgets that certain parts of a molecule are "fragile" or "sensitive" and need to be covered up (protected) before other parts can be worked on. Without this protection, the AI suggests recipes that look good on paper but would fail in a real lab.

The Solution: The "Smart Kitchen Manager" (Protect*)

The authors created Protect*, which acts like a neuro-symbolic kitchen manager. It combines two things:

1. The Chef's Intuition (Neural): The AI's ability to be creative and find new paths.
2. The Rulebook (Symbolic): A strict, unbreakable set of chemical rules and logic.

Here is how it works, using our kitchen analogy:

1. The Safety Check (Automated Detection)

Before the chef starts cooking, the manager scans the ingredients (the molecule). It has a massive checklist (a database of 55+ patterns) that says, "Hey, this onion is sensitive! If we chop the carrot next to it, the onion will burn."

• In the paper: The system uses SMARTS patterns (chemical rules) to automatically find which parts of the molecule are reactive and need protection.

2. The Protective Gear (Suggesting Groups)

Once the manager spots a sensitive ingredient, it suggests the right "protective gear."

• In the kitchen: "We need to wrap this onion in foil so the heat doesn't hit it."
• In the paper: The system picks from a library of 40+ protecting groups (like TBS, Boc, etc.) that chemically cover the sensitive spot so it doesn't react.

3. The "Human-in-the-Loop" Option

Sometimes, the manager isn't sure which foil to use, or the recipe is so complex that only a human expert knows the strategy.

• The Mode: The system pauses and asks the human chemist, "I see three ways to protect this onion. Which one do you prefer?" The human picks, and the manager locks that choice in.

4. The "Hard Constraint" (The Magic Trick)

This is the most important part. Usually, if you tell a chef "Don't touch the garlic," they might forget in the middle of cooking.
Protect* doesn't just tell the chef; it physically ties the chef's hands regarding that specific ingredient.

• The Analogy: The manager puts a literal lock on the garlic jar and gives the chef a map that says, "Garlic is locked. Do not open."
• The Tech: The system creates a Protection State. It translates the chemical rule into a specific code (atom maps) and injects it directly into the AI's prompt. It's not a suggestion; it's a mathematical rule that the AI cannot break.

The Results: A Perfect Meal

The team tested this on Erythromycin B, a very difficult antibiotic to make.

• Without Protect*: The AI tried to cook the meal but kept messing up the sensitive parts. The human had to stop the process 4 times to fix the mistakes and start over.
• With Protect*: The manager locked the sensitive spots immediately. The AI cooked the whole meal perfectly in zero restarts.

Why This Matters

Think of Protect* as a bridge between human logic and AI creativity.

• Old AI was like a student who memorized recipes but didn't understand the rules of safety.
• Protect* gives that student a strict safety manual and a supervisor who ensures they follow it, without killing their creativity.

This approach doesn't just help chemists; the authors suggest it could help in other fields too, like writing computer code (where you don't want the AI to delete important lines) or editing DNA (where you don't want to accidentally cut the wrong gene). It's about giving AI a "guardrail" so it can drive fast and creatively, but never crash.

Technical Summary

1. Problem Statement

While Large Language Models (LLMs) have shown promise in chemical retrosynthesis, they suffer from a critical limitation: a lack of fine-grained control over specific molecular sites.

• The Core Issue: Unconstrained LLM generation often fails to account for chemoselectivity and regioselectivity. In complex synthesis, certain functional groups (e.g., specific hydroxyls or amines) must be "protected" (temporarily blocked) to prevent unwanted reactions.
• Current Limitations: Existing models struggle to interpret spatial-chemical constraints (e.g., "do not react at this specific oxygen atom") when presented as linear SMILES strings. This leads to:
  • Chemically plausible but practically unfeasible pathways.
  • "Hallucinations" where the model suggests reactions at sites that should be protected.
  • A high need for human intervention to correct these errors, often requiring the restart of the synthesis plan (partial re-runs).

2. Methodology: The Protect∗ Framework

The authors propose Protect∗, a neuro-symbolic framework that grounds the generative intuition of LLMs in rigorous, deterministic chemical logic. It integrates with the existing DeepRetro system (a hybrid LLM + Monte Carlo Tree Search framework).

Key Components:

1. Automated Protection Site Detection:

   • Utilizes RDKit substructure matching on canonically atom-mapped SMILES.
   • Employs a database of 55+ SMARTS patterns covering 10 categories (alcohols, phenols, amines, carbonyls, etc.) to deterministically identify reactive sites requiring protection.
   • Ensures stability across different SMILES permutations via canonical atom mapping.

2. Protecting Group (PG) Suggestion Engine:

   • Queries a curated library of 40+ characterized protecting groups (e.g., TBS, Boc, Bn, MOM).
   • Ranks suggestions based on:
     • Compatibility with the specific functional group.
     • Orthogonality (compatibility with other protecting groups already present).
     • Synthetic accessibility and cost.
     • Stability profiles (acid/base/hydrogenation resistance).

3. Dual Interaction Modes:

   • Automatic Mode: The system automatically selects the top-ranked protecting group for every detected site, enabling fully automated planning for routine syntheses.
   • Human-in-the-Loop (HITL) Mode: Presents ranked suggestions to an expert chemist, allowing strategic selection for complex syntheses requiring non-obvious orthogonal strategies.

4. Contextual Constraint Encoding (The Core Innovation):

   • ProtectionState Object: A persistent, symbolic state existing outside the neural network. It maps specific canonical atom indices to their active protection status (e.g., "Atom [O:7] is protected with TBS").
   • State Injection: This symbolic state is injected into the LLM's inference context via structured prompt engineering. Instead of relying on vague natural language instructions, the prompt explicitly lists protected atoms using their canonical map numbers.
   • Effect: This creates a "guardrail" that forces the LLM to treat specific sites as non-reactive, shifting the cognitive burden from ambiguous interpretation to deterministic, logic-constrained inference.

3. Key Contributions

• Neuro-Symbolic Architecture: Successfully bridges the gap between neural generation and symbolic validity by enforcing hard constraints without requiring expensive model fine-tuning.
• Active State Tracking: Introduces a mechanism to maintain a persistent "Protection State" across inference steps, ensuring constraints are mathematically preserved throughout the search tree.
• Deterministic Control: Moves beyond simple prompt engineering by using atom-mapped constraints to eliminate ambiguity in the LLM's understanding of molecular topology.
• Scalable Knowledge Base: Integrates a comprehensive, structured database of chemical rules (SMARTS) and protecting groups directly into the generation pipeline.

4. Results

The framework was evaluated on complex natural products known to be challenging for automated planning.

• Reduction in Human Intervention:
  • Erythromycin B: The unsteered "Vanilla DeepRetro" system required 4 partial re-runs (human interventions to fix hallucinated pathways). The Protect∗-steered system required 0 re-runs.
  • Prostaglandin E2 & Quinine: Both methods solved these without re-runs, establishing a baseline where the new method matches existing performance on simpler targets.
• Accuracy:
  • Achieved 100% Top-1 PG Accuracy in detecting reactive sites and recommending the correct protecting group for all tested molecules (Erythromycin B, Prostaglandin E2, Quinine).
• Case Study (Erythromycin B):
  • The system successfully identified secondary alcohol sites requiring protection for a late-stage macrocyclization.
  • The automatic mode independently recovered the same protecting group strategy (Ethoxy/OEt) that human experts had previously selected in HITL experiments.
  • This enabled the discovery of a novel, chemically elegant synthetic pathway that avoided the pitfalls of unprotected reactive sites.

5. Significance

• Expert-Level Autonomy: Demonstrates that neuro-symbolic systems can match or replicate expert-level strategic decisions in complex chemical synthesis, reducing the reliance on constant human oversight.
• Reliability: By grounding LLMs in symbolic logic, the system significantly reduces "hallucinations" regarding chemical feasibility, making AI-driven retrosynthesis more trustworthy for industrial and academic applications.
• Generalizability: The approach of using persistent symbolic states to constrain neural generation is not limited to chemistry. The authors suggest it could be applied to other domains requiring structured constraints, such as genetic engineering (codon marking) or code generation (architectural pattern preservation).
• Efficiency: Eliminates the need for expensive re-training of LLMs to learn chemical rules, instead leveraging existing rule-based engines to guide pre-trained models.