SynLlama

Generative models for chemical space routinely propose molecules that no one can actually make. A structure can score well on a docking or property objective and still be unreachable in the lab, which turns a promising screen into a dead end. SynLlama attacks this from the other direction: instead of generating a molecule and then searching for a route to it, it generates the route itself — a full synthetic pathway assembled from commercially available building blocks and robust organic reaction templates — so every output arrives with instructions for making it.

SynLlama was developed by the Head-Gordon group at UC Berkeley, with collaborators at the University of Minnesota and Contramont Research, and published in ACS Central Science in 2025. It is a supervised fine-tune of Meta's Llama 3, repurposing a general-purpose language model as a synthesis planner rather than training a bespoke architecture from scratch. That choice is the point of the paper: by inheriting a pretrained model's grasp of structured token sequences, SynLlama matches or beats specialist route-planning systems while training on one to two orders of magnitude less pathway data.

The model addresses three related tasks that medicinal chemists care about. Bottom-up synthesis planning reconstructs a route to a given target. Synthesizable analog generation produces close structural neighbors that are all makeable. Hit expansion takes a validated lead against a protein target and proposes a family of accessible variants around it. The same fine-tuned checkpoint serves all three, applied across many different targets without per-target retraining.

#Key Features

• Routes, not just structures: Every generation is a complete synthetic pathway — the building blocks to buy and the reactions to run — rather than a SMILES string that may or may not be reachable.
• Generalizes to unpurchased chemistry: Trained only on Enamine building blocks, the model still reconstructs routes using purchasable blocks it never saw, meaning its reach extends past its own training catalog.
• Data efficiency from a pretrained backbone: 2 million training routes suffice, roughly 40-fold less than Synformer's 85 million and 60-fold less than ChemProjector's 128 million.
• Analog generation for lead optimization: Given a hit compound, SynLlama proposes synthesizable neighbors that preserve predicted binding while improving synthetic accessibility.

#Technical Details

SynLlama fine-tunes two Llama 3 variants — Llama-3.2-1B and Llama-3.1-8B — with full supervised fine-tuning on synthetic pathways serialized as text. The 1B model trained on 2 million routes (SynLlama-1B-2M) is the checkpoint used for the paper's downstream experiments, chosen for faster inference at comparable cost to the 8B variant. Training routes are enumerated from roughly 230,000 Enamine building blocks combined under two curated reaction template sets of 91 and 115 reactions; the larger set defines a synthesizable space on the order of 10^30 molecules.

On 1,000-molecule reconstruction benchmarks, SynLlama recovers 69.1% of Enamine targets and 28.7% of ChEMBL targets, against 66.0% and 19.8% for Synformer, 46.2% and 13.3% for ChemProjector, and 11.0% and 5.4% for SynNet. Applied to 500 computationally designed SARS-CoV-2 main protease binders, its analogs reproduced docking scores to within 1.04 kcal/mol RMSE while improving synthetic accessibility. In a hit expansion test it rediscovered two experimentally validated inhibitors from prior work and proposed six further candidates within 1 kcal/mol of the lead.

#Applications

SynLlama is aimed at medicinal and computational chemists working between virtual screening and the bench. Its most direct use is filtering and repairing the output of upstream generative models: given an attractive but impractical structure, it returns the nearest analogs that can actually be ordered and assembled. It also supports hit expansion around a confirmed binder, route proposal for a target of interest, and construction of synthesis-aware libraries for a specific protein. Because the checkpoint is fixed and applied zero-shot to new targets and new building-block catalogs, it slots into an existing pipeline without a retraining step.

#Impact

SynLlama makes a broader argument about method choice in chemistry: a general-purpose LLM, fine-tuned carefully on a well-structured representation, can outperform architectures designed specifically for synthesis planning while consuming far less training data. The peer-reviewed ACS Central Science publication, released weights, and code covering both inference and retraining make the result reproducible. Practical limits remain — the reaction template sets are curated and finite, coverage of ChEMBL-like drug space is still under 30%, and the code carries a University of California academic license restricting use to educational, research, and not-for-profit purposes, which constrains direct commercial deployment.