2026-09-14
DeepRetro chains template search with validated LLM disconnections. Claude 4 Opus plus Pistachio solves 168/172 Drug Hunter targets vs 99 for Retro* and 113 for PDVN.
Retrosynthesis works backwards from a target molecule to buyable starting materials. Template libraries plus Monte Carlo tree search already handle many common drugs, but a disconnection that is missing from the template set gets pruned. Complex natural products make this worse: ring strain, protecting groups, and stereochemistry sit outside the rule tables, so the search tree grows until heuristics start cutting branches.
General-purpose LLMs can surface less common transforms from the literature, and they also invent intermediates that violate valency, fall apart, or do not exist. DeepRetro wires those two failure modes into one pipeline: run the template engine first, ask an LLM for a single-step disconnection only when that engine fails, then accept the step only if it passes chemical checks.
The pipeline first asks whether the molecule is already in a stock catalog. If not, it calls a template/MCTS solver T, an adapted AiZynthFinder trained separately on USPTO and on Pistachio 2024Q4. If T finds a complete route to stock, it stops. Only then does it query an LLM, and only for one step, not a full tree.
Proposed precursors must pass three gates: valency and allowed atoms, structural stability, and a hallucination screen. Survivors are fed back into the same pipeline recursively. Chemists can optionally regenerate a bad branch, edit a structure, or add protecting groups. A head node farms work to multiple workers so harder molecules can scale out.
The reason for single-step generation is simple. A full route in one shot lets an early hallucination poison the whole tree. A hard check after every step forces the search to land.
Single-step evaluation used 250 reactions clustered from the USPTO-50k test set by circular fingerprints, because the strongest setup costs more than $1 per molecule. With Claude 4 Opus and Pistachio:
| Method | All Correct | Any Correct | MaxFrag |
| ASKCOS | 105/250 | 118/250 | 115/250 |
| AiZynthFinder (Pistachio) | 73/250 | 83/250 | 79/250 |
| DeepRetro (Claude 4 Opus, Pistachio) | 111/250 | 137/250 | 126/250 |
All Correct barely beats ASKCOS (111 vs 105). Any Correct and MaxFrag move more. The LLM often proposes a chemically valid transform whose reagents do not match the dataset label, so strict set-matching undercounts it.
The multi-step gap is larger. On USPTO-190 in fully automatic mode, Claude 4 Opus plus Pistachio solves 183/190 (96.3%), against 177/190 for PDVN and 139/190 for Retro. On the harder Drug Hunter 172 set, the same setup solves 168/172 (97.7%), against 99 for Retro and 113 for PDVN. Open-source DeepSeek R1 plus Pistachio still reaches 165/172.
Five complex natural products ran in human-in-the-loop mode, 6 to 14 DeepRetro runs each, with chemists intervening at key steps. Ohauamine C and a tetracyclic azepine derivative are judged novel as disconnection sequences, not as new elementary reactions. Erythromycin B, reserpine, and discodermolide also get workable routes, always with human edits. Case-study routes used Claude 3.7 Sonnet and were repeated three times.
This is a model-agnostic hybrid CASP stack. Teams do not have to fine-tune a chemistry foundation model. For a medicinal-chemistry target the template engine cannot crack, the pipeline can run unattended and a chemist can patch one node. The Drug Hunter margin over Retro is the number to remember.
It is still an assistant. The case-study molecules do not go through without people, and open-source backends lag Claude 4 Opus. For groups that already ship production CASP, this is an incremental add-on with a human channel. For natural products that sit outside template coverage, it is worth a serious try.
Hallucinations are the main weakness the paper names. The LLMs are general models, not reaction-finetuned, the checkers still miss edge cases, and early errors compound under recursion. Fully automatic routes for complex natural products are not there yet. The strongest model costs more than $1 per molecule, so the 250-reaction single-step set is, in the authors' own words, a qualitative comparison. Open-source configs still trail closed models. The system still depends on USPTO and Pistachio; an LLM is not a substitute for a reaction database.
The two "novel" routes mean the disconnection sequence and intermediate scaffolds were absent from SciFinder/Reaxys, not that new reaction types were invented. DeepRetro has not solved a molecule that human chemists have not already made. Dual-use risk is stated plainly: the same stack can help plan routes to controlled substances.