AI-POWERED SYNTHESIS PLANNING

Synthelite's Two-Phase Planning Workflow

Synthelite employs a novel two-phase approach to generate comprehensive and feasible synthesis routes. This method combines the intrinsic chemical reasoning capabilities of LLMs with a similarity-based Monte Carlo Tree Search (MCTS) to ensure both strategic alignment and practical applicability.

In Phase 1, the LLM iteratively drafts a synthesis strategy for each retro-step, converting chemical transformations into natural language descriptions. These descriptions are then used to retrieve reaction templates from a vast, LLM-annotated database. Phase 2 refines this blueprint, sampling similar reactions and optimizing towards stock availability and alignment with LLM-proposed strategies.

Achieving Expert-Aligned Synthesis Routes

Synthelite demonstrates exceptional ability to align synthesis plans with diverse, user-specified constraints, such as preferred reaction types or specific formation stages. This adaptability is critical for chemists who often impose practical preferences on their synthesis design.

The framework consistently achieves high recall, finding valid routes that satisfy prompts, and significantly higher precision than traditional methods, indicating a more focused and relevant search. For example, Synthelite can dynamically adjust its strategy to prioritize late-stage vs. early-stage pyrazole formation based on prompt variations.

Mastering Starting Material Constraints

Beyond strategic guidance, Synthelite excels in starting material constrained synthesis, a task crucial for semi-synthesis or waste valorization. Users can specify required leaf molecules, and the LLM effectively guides the planning towards incorporating these building blocks.

This is a particularly challenging area where traditional methods like AiZynthFinder show a significant performance drop when constraints are applied. Synthelite, especially with Gemini-2.5-Pro, maintains high solve rates, showcasing the LLM's reasoning capabilities in complex, constrained environments without explicit value functions tailored to these constraints.

Ensuring Wet-Lab Feasibility and Route Quality

Synthelite significantly bridges the gap between computational planning and experimental reality by accounting for wet-lab feasibility. LLMs demonstrate fine-grained reasoning about chemoselectivity, reaction conditions, and potential side reactions.

The framework's self-evaluation mechanism, where LLMs generate feedback on previous attempts, allows for iterative refinement of synthesis strategies. This process identifies and corrects potential issues like over-alkylation, catalyst poisoning, or regioselectivity problems, leading to robust and practical routes.

Moreover, conducting multiple planning attempts combined with MCTS refinement consistently improves solve rates on benchmarks like USPTO-190, demonstrating the framework's robustness.

Conclusion & Future Directions

Synthelite presents a significant advancement in computer-aided synthesis planning by integrating LLMs' strategic reasoning and chemical knowledge with Monte Carlo Tree Search. This framework offers explainable, user-steerable, and feasibility-aware route generation, setting a new paradigm for chemist-AI collaboration.

Future work will focus on reducing dependence on closed-source LLMs and developing more efficient mechanisms for LLM-guided single-step retrosynthesis. Despite current limitations, Synthelite is a meaningful step towards bridging LLM-based reasoning with the inherently explorative nature of synthesis planning, making it a more practical and transparent computational tool.