Nhan Do Van Thanh

My team (@SchwallerGroup) has been cooking something big, and I'm really proud of it. Excited to share "Strategy-first synthesis planning for complex natural products"! We are entering the era of agentic synthesis planning, and for the first time I see expert synthetic chemists excited about (some, not all) AI-generated routes, a stark contrast to anything we worked on before. Solving routes (the field's default) does not mean solving them elegantly or strategically. That is what we've been working on. And natural products are a space typically out of reach for other ML-based planning tools. Quick summary: SynthEx is an agentic planner: a language model writes disconnections as atom-level graph edits rather than picking from a template library. A multi-agent system runs the loop: one agent proposes competing strategies per target, each anchored on a key disconnection, another expands them into full routes, and a critic-editor pair simulates each step forward and repairs what would fail. → On 1,098 complex natural products with no reported total synthesis, a near-exhaustive AiZynthFinder run solves 13.8%. SynthEx reaches 63.9%, and the margin widens as molecules get more complex. → In blinded review, 10 expert chemists gave 1,040 ratings over 148 key steps. SynthEx's steps were indistinguishable from published human ones on feasibility, elegance and overall quality, and raters could not tell machine from human (AUC 0.48). Case studies highlight: → Okaramine M: it recovers the strategic logic of an expert route published after the model's training cut-off. → Melonine: it converges on exactly the key disconnection an expert group had committed lab effort to. That attempt failed for conformational reasons; SynthEx reaches it by a variant the same group judges more likely to succeed. → Chanoclavine to Lysergol: an open problem with no reported route. From two structures alone, it proposed a Hofmann-Löffler-Freytag sequence with a 1,6-HAT. Huge collaborative project with the Zhu, Wipf and Njardarson groups, whose feedback shaped this throughout. All the authors: Daniel Armstrong (@d_armstr )*, Xuan-Vu Nguyen* (@XuanVuNguyen18), Octavian Susanu, Gabriel Gibberd, Théo Neukomm, Taddäus Strunden, Dan Forster, Morgane Delattre, Shawn Teh, Clément Rols, John Federice, Hayden Leatherwood, Lavelle Barnes, Maarten Dobbelaere, Peter Wipf (@wipf_group), Jon Njardarson, Jieping Zhu, Philippe Schwaller (*equal) Grateful for funding from the Swiss National Science Foundation SNSF (@snsf_ch) via @NCCR_Catalysis (225147) and grant 214915; the MSCA Doctoral Networks Explainable AI for Molecules - @AiChemist_DN MSCA DN Horizon Europe and LowDataML for Sustainable Chemical Sciences; and Intel and Merck Group via AWASES. Made feasible by @Google Cloud Research Credits for the @GeminiApp Academic Program, with support from Digital Schooling. The USD 50k award I received in early 2026 let us build and run this without worrying about API costs. Backbone: Gemini 3.1 Pro. We link related reactions via the @ElsevierConnect Reaxys R&D collaboration network. -> SynthAtlas (explore all 3,243 routes, and leave comments): synthatlas.epfl.ch -> Preprint: arxiv.org/abs/2608.07454 (SynthEx method) -> Code: github.com/schwallergroup/sy… - set Watch to be notified when it lands under Apache 2.0