Synthelite: LLM-Driven Retrosynthesis Planning

• Synthelite is a computer-aided synthesis planning framework that uses large language models to propose, evaluate, and orchestrate multi-step retrosynthetic routes.
• It integrates a semantic template search engine, blueprint planning via LLM, and Monte Carlo Tree Search to ensure precision and feasibility in reaction pathways.
• Empirical benchmarks show Synthelite delivers high success rates under strategy and starting-material constraints, effectively bridging algorithmic planning with wet-lab experimentation.

Synthelite is a computer-aided synthesis planning (CASP) framework that centers LLMs as the agent for proposing, evaluating, and orchestrating multi-step retrosynthetic transformations. Unlike prior template-based or neural policy approaches, Synthelite natively supports chemist-aligned, feasibility-aware, and prompt-steerable route generation, with high empirical success rates under both strategy and starting-material constraints. By leveraging semantic template search, LLM-driven blueprint planning, and Monte Carlo Tree Search (MCTS)-based refinement, Synthelite closes the gap between strategic algorithmic planning and wet-lab experimentation, while allowing for direct human–LLM interaction via natural language (Xuan-Vu et al., 18 Dec 2025).

1. System Architecture and Workflow

Synthelite operates as a two-phase retrosynthesis planning framework. At its core, it integrates an LLM with a template-based search engine to combine free-form, chemically informed reasoning with robust substructure transformations. The primary workflow consists of:

1. Template Search Engine: Approximately 40,000 reaction templates (SMARTS patterns, derived from AiZynthFinder) are converted by an LLM (Claude-3.7-Sonnet) into concise, one-sentence textual descriptions. These descriptions are embedded into a vector space, facilitating top-$k$ semantic retrieval for any given transformation prompt.
2. Phase 1 – Blueprint Planning via LLM:
   • An LLM (Claude-4.5-Sonnet, Gemini-2.5-Pro, or GPT-5) is iteratively prompted to (a) decide on termination, (b) outline an overall retrosynthetic strategy conditioned on user input, and (c) propose the next disconnection as a textual description.
   • The proposed transformation is matched to candidate templates using the embedding search; up to 20 candidate reactions are vetted, and the LLM selects the one most consistent with its own textual intent.
   • The greedy, stepwise loop develops the disconnection blueprint until precursors are in-stock or a stop signal is generated.
   • The drafting process is repeated three times, with self-evaluation feedback from the LLM provided to subsequent runs, guiding strategy diversity and self-correction.
3. Phase 2 – MCTS-Based Refinement:
   • For each blueprint, the approach launches a Monte Carlo Tree Search (300 iterations per blueprint), where expansion at depth $d$ relies on a logit scoring function:

$$z(r(t)\mid d) = \alpha\,\mathrm{sim}(t,q_d) + (1-\alpha)\,\frac{N(t)}{N(t)+C},$$

where $\mathrm{sim}(t,q_d)$ is the vector embedding similarity between template and LLM description, $N(t)$ is template popularity, $C=100$, and $\alpha=0.5$. - Only templates matching the same reaction site as the LLM-specified reference are investigated, and expansions are cached to reduce redundant LLM calls. - Final routes are ranked based on agreement with the Phase 1 sketch.

2. Human–LLM Interaction and Prompt Steering

A defining feature of Synthelite is its seamless expert steerability, facilitated by LLM-driven natural language prompts. This enables two primary steering paradigms:

• Strategy-constrained planning: Users can impose synthetic strategies directly, such as "Late-stage pyrazole formation," "Only break oxoisoindolinone ring," or "Perform reduction before coupling." The LLM immediately adapts retrosynthetic logic to prioritize or delay certain disconnections, protecting group strategies, or transformation orders.
• Starting-material constraints: By including specific IUPAC names or SMILES strings for required leaf building blocks, chemists can ensure that prescribed intermediates are treated as terminal leaf nodes in the route sketch. The LLM interprets these instructions, prioritizing retrosynthetic cuts and constructing the route around the constraint.

The iterative self-evaluation in Phase 1 further incorporates expert feedback and LLM reflection, enhancing both compliance and route diversity.

3. Feasibility-Aware Route Design

Synthelite harnesses the chemical knowledge embedded in LLMs for feasibility-aware synthesis planning via two synergistic mechanisms:

1. Neutral Feasibility Prompt: All routes can be conditioned with a neutral prompt that biases planning toward experimentally plausible, high-yield, low-side-reaction pathways, e.g., "Highly feasible synthesis with high overall yields; consider side reactions and avoid unnecessary steps." This reduces the frequency of exotic or impractical route suggestions.
2. Self-Evaluation and LLM Judge Scoring: After blueprint drafting, the LLM systematically reviews each step, flagging potential feasibility issues (e.g., over-alkylation, catalyst poisoning, selectivity challenges). Large-scale experiments deploy an external LLM (Gemini-2.5-Pro) to score route feasibility on a $[1,10]$ scale, with threefold scoring per route and an optimistic maximum retained for analysis.

Empirically, Synthelite routes cluster around $s=7$–$10$, significantly exceeding the $s=2$–$5$ range typical for AiZynthFinder, demonstrating high alignment with practical synthetic success criteria.

4. Quantitative Benchmarks and Performance

Synthelite was systematically evaluated on strategy-constrained and starting-material-constrained benchmarks:

Benchmark	AZF (Baseline)	Synthelite (Gemini)	Synthelite (Claude)	Synthelite (GPT-5)	Tango*
Strategy-Constrained Top-1 R.	55%	89%	--	--	--
Strategy-Constrained Top-30 R.	81%	95%	--	--	--
Precision @ top-K	<80%	~83%	~80%	~75%	--
Starting-Material Constrained	5%	95%	75%	60%	100%

• Strategy-constrained recall: Synthelite (Gemini) achieves top-1 recall of 89% and top-30 recall of 95% versus 55% and 81% for AiZynthFinder (AZF) (baseline).
• Precision: All Synthelite LLMs yield higher route compliance with prompts than AZF.
• Starting-material constraint: Synthelite (Gemini) solves 95% of targets, outperforming AZF (5%); Tango* attains 100% but via a distinct value-function-based policy.
• Feasibility scores: Synthelite routes are consistently more feasible by LLM judge assessment compared to AZF.

Ablation studies (USPTO-190) show greedy LLM-only solutions yield moderate recall (Gemini 51%, Claude 47%), improving to 69%/62% by adding iterative Phase 1 attempts, and up to ~74% with full MCTS–comparable or superior to state-of-the-art frameworks.

5. Case Studies and Route Examples

• Synthelite 1 (pyrazole installation): With the "late-stage pyrazole formation" prompt, the route concludes with a Knorr reaction; with "early-stage pyrazole formation," a 3+2 condensation places the pyrazole earlier, followed by further couplings.
• Synthelite 2 (three-ring aromatic):
  • Attempt 1: Route via halide amination and dual Suzuki couplings; self-evaluation flags over-alkylation and selectivity issues (score 5/10).
  • Attempt 2: Revision via reductive amination, Boc-protection, switching partners for coupling—delivers a stepwise improvement and judges as 10/10 feasible.

These examples illustrate both Synthelite’s alignment with natural-language strategies and its ability for reflective improvement based on chemically relevant feedback.

6. Limitations, Gaps, and Prospective Directions

Current limitations include:

• Reliance on closed-source LLMs, complicating reproducibility, latency analysis, and unbiased benchmarking.
• Coverage constraints in the reaction template library: some LLM-suggested transformations cannot be grounded to existing SMARTS templates, leading to premature truncation of search.
• Template extraction dependency: While LLM textual reasoning is broad, coverage ultimately depends on the underlying template base.
• Need for black-box LLM access: which can stymie integration or interpretation of detailed model decisions.

Future directions include the development of open-source single-step LLM policies, dynamic template induction to expand action spaces, integrating third-party reaction heuristics (such as solvent, protecting group or catalyst constraints), and reducing reliance on restricted-access LLM APIs (Xuan-Vu et al., 18 Dec 2025).

7. Context and Significance

Synthelite empirically demonstrates that LLMs can go beyond validation or one-step prediction roles. Instead, they can function as central agents orchestrating multi-step retrosynthetic planning, balancing user-intent alignment, practical feasibility, and algorithmic search efficiency. By integrating prompt steerability, self-reflection, feasibility checks, and robust search, Synthelite bridges the longstanding gap between computer-aided plan generation and experimental chemistry in the laboratory, signifying a step change in the flexibility and realism of CASP approaches (Xuan-Vu et al., 18 Dec 2025).