Technical Summary: Project Ariadne

Problem Statement

Retrosynthetic planning aims to connect a target molecule to commercially available starting materials through a multistep route. While classical approaches rely on iterative single-step reaction models coupled with search algorithms (e.g., MCTS), these methods often require specialized algorithms or architectural changes to handle specific planning constraints, such as route depth or required starting materials. Direct generation methods have emerged to reframe retrosynthesis as sequence generation, but existing models typically require separate training for different planning specifications (e.g., one model for unconstrained routes, another for depth-constrained routes). This fragmentation limits the flexibility and efficiency of synthesis planning systems.

Methodology

The authors introduce Ariadne, a decoder-only transformer model that unifies target representation, optional planning constraints, and route generation into a single prompt-completion sequence.

Data Representation

Unlike previous encoder-decoder approaches (e.g., DirectMultiStep) that treat retrosynthesis as a translation task, Ariadne represents the entire synthesis planning task as a rooted S-expression within a single sequence:

• Structure: (task (spec ...) (route ...))
• Specification (spec): Contains the target molecule (query (mol ...)) and optional constraints such as route_depth or required_leaves.
• Route (route): A recursive tree structure where leaf nodes represent starting materials (leaf (mol ...)) and internal nodes represent reactions (reaction (mol ...) (children ...)).
• Tokenization: The model uses a small, S-expression-aware tokenizer where structural labels are atomic tokens and SMILES strings are tokenized character-wise.

Training Strategy

• Architecture: Ariadne employs a decoder-only architecture with pre-normalization, RMSNorm, rotary position embeddings (RoPE), and SwiGLU feed-forward blocks. The model scales primarily by increasing layer count (12L and 24L variants) while keeping hidden size (256) and attention heads (8) constant.
• Prompt Conditioning: The model is trained to generate the route portion of the sequence conditioned on the spec portion. By varying the spec block (e.g., adding depth constraints or specific starting materials), the same checkpoint can be queried with different task specifications at inference time.
• Data Augmentation: Training data includes permutations of sibling order in the route tree to ensure invariance to the order of precursors.
• Dataset: Models were trained on the PaRoutes dataset (v2026-05-12 split) preprocessed with RetroCast, utilizing both route-holdout and reaction-holdout splits to prevent data leakage.

Evaluation Framework

The authors evaluate Ariadne using the RetroCast framework, focusing on:

• Tier-0 Validity: The proportion of targets with at least one route where all reactions are chemically valid (Tier-0).
• Solv-0: The proportion of targets with at least one Tier-0-valid route that satisfies specific scoring task constraints (e.g., termination in ASKCOS Buyables, specific depth, or required starting materials).
• Top-K Accuracy: The ability to reconstruct a reference route within the top K candidates.
• Benchmarks: The primary evaluation uses the mkt-cnv-160 family (target-only, depth-constrained, and leaf-constrained variants) and ref-* benchmarks (using custom reference stocks).

Key Contributions

1. Unified Prompt-Conditioned Generation: Ariadne demonstrates that multistep route generation can be formulated as a single decoder-only task where constraints are handled via prompt fields rather than separate model training.
2. Efficiency in Constrained Planning: The model achieves state-of-the-art performance on constrained tasks (depth and required starting materials) using a single checkpoint, significantly outperforming bidirectional search planners like DESP in both accuracy and computational cost.
3. Decoupling of Generalization and Planning: The study provides diagnostic metrics (e.g., Route|Root accuracy) to separate the model's ability to select the first disconnection from its ability to complete the rest of the route, revealing that the primary bottleneck is often the initial disconnection choice.
4. Open-Source Release: The authors release the codebase and training scripts to support further research in prompt-conditioned structural route generation.

Results

• Constraint Following: On the mkt-cnv-160 benchmarks, adding prompt fields for constraints significantly improved performance.
  • Depth Constraints: Solv-0 increased from 76.9% (target-only) to 90.6% when the requested depth was included in the prompt.
  • Required-Leaf Constraints: Solv-0 increased from 50.0% to 81.2%, and Top-10 reconstruction improved from 26.2% to 37.5% when a specific starting material was prompted.
• Comparison to Search-Based Planners: On the required-leaf benchmark, Ariadne (24L) achieved 81.2% Solv-0 and 37.5% Top-10 in 24 GPU-minutes, outperforming the best DESP configuration (71.2% Solv-0, 17.5% Top-10) which required 6.8 GPU-hours (a ~17x speedup).
• Standard Reconstruction: On standard target-only tasks, Ariadne 24L is comparable to the DMS Explorer XL (50M parameters) in Top-10 accuracy (59.4% vs 57.5%) but with roughly half the inference time (23.9 min vs 47.6 min).
• Holdout Performance: Ariadne shows significant gains over MCTS on route-holdout benchmarks (Top-10: 59.4% vs 40.6%), suggesting better ability to assemble known reactions into new routes. However, on reaction-holdout benchmarks, MCTS remains stronger in Solv-0, indicating that search-based methods still hold an advantage in generalizing to unseen reaction transformations.
• Stock Conditioning Limitations: When evaluated on benchmarks requiring termination in custom reference stocks (not just commercial buyables), Ariadne's performance drops significantly compared to MCTS (e.g., 68.2% vs 85.2% Solv-0 on ref-cnv-400), highlighting a current limitation in conditioning generation on arbitrary user-defined stock sets.

Significance and Claims

The paper claims that Ariadne extends sequence generation from specialist retrosynthesis models to prompt-conditioned structural route generation. The primary significance lies in demonstrating that a single decoder-only model can handle multiple planning specifications (unconstrained, depth-constrained, and leaf-constrained) without retraining, offering a more flexible and computationally efficient alternative to search-based planners for specific constraint types.

However, the authors maintain a modest stance regarding immediate experimental utility. They explicitly state that:

• The models do not introduce Tier-1–3 route checkers (e.g., reaction plausibility, condition prediction), which remain the main bottleneck for experimental application.
• The current approach cannot yet condition generation on arbitrary user-specified stock sets, a critical requirement for practical planning involving in-house building blocks.
• The work is not intended to replace explicit search (MCTS) but rather to offer a complementary scaling axis where learned generators amortize recurring route structures for faster inference, potentially in a hybrid system where search generates trajectories for distillation.

The paper concludes that while Ariadne represents a significant step toward unified synthesis planning, further research is needed in stock-aware prompting, reliable first-disconnection selection, and standardized validation protocols before these models can be fully trusted by experimental chemists.