Designing a molecule from scratch is one of chemistry’s hardest puzzles — not just choosing which atoms to join, but planning the exact sequence of reactions, protecting vulnerable parts of the molecule, and avoiding dead-end routes that can waste months of lab work. A team at EPFL wants to move that tacit expertise out of senior chemists’ heads and into a language model.
Their new framework, Synthegy, was described this week in Matter by Philippe Schwaller and colleagues. Instead of asking AI to invent molecules, Synthegy uses large language models (LLMs) as smart evaluators of synthesis plans that conventional retrosynthesis software already proposes. The result: a way for chemists to type plain-English instructions and get ranked, explained synthesis routes that match strategic intent.
How it works
- A chemist enters a goal in natural language — for example, “form the pyrimidine ring in the early stages.”
- Existing retrosynthesis tools generate dozens or hundreds of candidate routes by breaking the target down into simpler pieces.
- Synthegy converts each route into text and asks an LLM to score how well it satisfies the chemist’s instruction, producing ranked routes with written explanations of the reasoning.
“As Andres M. Bran, lead author, put it: ‘When making tools for chemists, the user interface matters a lot, and previous tools relied on cumbersome filters and rules.’” The LLM-driven scoring replaces brittle rule sets with flexible, strategy-aware reasoning.
Validation and performance
- The team ran a double-blind study with 36 independent chemists who reviewed 368 pairs of routes. Synthegy’s top picks matched the chemists’ selections 71.2% of the time — roughly on par with how often expert chemists agree with each other.
- Senior researchers (professors and research scientists) agreed with Synthegy more often than PhD students, suggesting the system captures high-level strategic intuition.
- The framework was tested with multiple LLMs (including GPT-4o, Claude, and DeepSeek-r1). Gemini-2.5-pro scored highest on the benchmark, while DeepSeek-r1 emerged as a strong open-source option that can run locally.
Beyond route selection: mechanisms and feasibility
Synthegy also tackles reaction mechanism elucidation — breaking reactions into elementary electron-movement steps and asking the LLM to assess each step for chemical plausibility. On straightforward transformations such as nucleophilic substitutions, top models achieved near-perfect accuracy.
Practicals, limits and openness
- Speed/cost: evaluating 60 candidate routes takes roughly 12 minutes and costs about $2–3 in API fees.
- Current limitations: LLMs sometimes misread the directionality of reactions in their text representations, leading to incorrect feasibility calls. Smaller LMs perform close to random, and routes longer than ~20 steps are harder for the system to track coherently.
- Modularity and availability: Synthegy is designed to plug into any retrosynthesis engine and any capable LLM. The code and benchmarks are publicly available: github.com/schwallergroup/steer.
Why it matters for crypto and open science
For crypto-native readers, Synthegy’s modular, model-agnostic design and the viability of robust open-source models like DeepSeek-r1 are notable. The framework could accelerate decentralized drug discovery, collaborative R&D DAOs, and tokenized bounties for synthesis design, by lowering the expertise barrier and enabling quicker shared evaluation of synthesis plans using affordable compute. Its public codebase also makes it a candidate for community-driven improvements and deployment on private or on-chain compute stacks.
Bottom line: Synthegy doesn’t replace chemists, but it brings the kind of strategic judgment usually held by experienced practitioners into a flexible AI layer — speeding route selection, offering human-readable explanations, and opening new pathways for collaborative and decentralized chemistry workflows.
Read more AI-generated news on: undefined/news
