Synthetic Accessibility Score (SAS) Overview

• Synthetic Accessibility Score (SAS) is a computational metric used to evaluate the feasibility of synthesizing small molecules using both heuristic fragment analysis and data-driven methods.
• Heuristic methods like the Ertl–Schuffenhauer score decompose molecules into fragments, but may blur distinctions in real-world synthetic tractability.
• Recent advancements integrate retrosynthetic planning and machine learning, such as the round-trip score and FSscore, to offer more precise and adaptive synthetic feasibility assessments.

Synthetic accessibility score (SAS) is a family of computational metrics used to assess the synthetic feasibility of small molecules, especially in high-throughput applications such as drug discovery, virtual screening, and molecular optimization. SAS aims to quantify, usually as a scalar, the likelihood that a given compound can be produced from commercially available starting materials using known reaction chemistry. Historically, SAS metrics have been based on heuristic fragment and complexity analysis, but recent advances leverage machine learning, data-driven retrosynthetic planning, and human-in-the-loop refinement to address the limitations of traditional scores.

1. Heuristic SAS: The Ertl–Schuffenhauer Score and Its Limitations

The Ertl–Schuffenhauer synthetic accessibility (SA) score is among the most widely adopted heuristics for estimating synthetic feasibility. SA scores are calculated from molecular structure alone, decomposing the molecule into substructural fragments and combining fragment-based "ease" contributions and explicit complexity penalties. The score is formally expressed as

$$\mathrm{SA}(m) = \sum_{f \in \text{fragments}(m)} \text{score}(f) + \text{penalty}(\text{complexity}(m))$$

where fragment scores are empirically derived from their frequency in commercial chemical catalogs and penalties increase with features such as stereocenters and ring complexity. The method is fast and scalable but does not attempt retrosynthetic analysis, nor does it validate that a synthetic route actually exists. As such, molecules with similar fragment statistics may have fundamentally different real-world synthetic tractability. Empirical evidence indicates that the SA score distributions for molecules with and without feasible retrosynthetic routes (as found by a planner) nearly overlap, demonstrating the inability of SA to reliably distinguish route-findable targets (Liu et al., 2024).

2. Data-Driven, Route-Based Metrics: The "Round-Trip" Score

Recent developments address the limitations of heuristic SAS by introducing metrics grounded in explicit, data-driven retrosynthetic planning and forward reaction prediction. The "round-trip" synthetic accessibility metric, proposed by Liu et al., integrates two components:

• Retrosynthetic Planner: A template-based, sequence-to-sequence model (Neuralsym; Segler & Waller 2017) is trained on a large reaction network constructed from USPTO-full patent data. The planner, using beam search, recursively proposes one-step disconnections and assembles multi-step routes terminating in commercial starting materials. Each route is scored by the product of the neural model's softmax probabilities for chosen templates.
• Forward Reaction Predictor: A Transformer-Decoder model trained on ~811,000 patent reactions is used to simulate the synthetic route in the forward direction. It predicts the product of each reaction step, given the reactant set, reconstructing the target molecule from starting materials.

The round-trip score $S(m)$ for a molecule $m$ is formally defined as the Tanimoto similarity between the Morgan fingerprint of $m$ and that of $m'$ (the product of simulated forward synthesis):

$$S(m) = \mathrm{Sim}_{\mathrm{Tanimoto}} \left[ m,\, f_{\Theta}(g_{\Phi}(m)) \right]$$

where $g_{\Phi}$ and $f_{\Theta}$ denote the retrosynthesis and forward prediction models, respectively. For $K$ candidate retrosynthetic routes, the final SAS is the maximum similarity among the $K$:

$S(m)$0

A high $S(m)$1 indicates a route exists by which $S(m)$2 can be planned and regenerated in silico from purchasable materials; a low $S(m)$3 suggests synthetic inaccessibility under the model (Liu et al., 2024).

3. SAS via Machine Learning: Human-in-the-Loop and FSscore

An alternative paradigm for SAS is exemplified by the Focused Synthesizability score (FSscore), which uses machine learning to learn a relative ranking of molecular synthesizability from pairwise expert preferences, tuned to specific chemical spaces (Neeser et al., 2023). FSscore is a scalar function $S(m)$4 learned from binary preferences on pairs $S(m)$5, such that:

$S(m)$6

The neural network backbone is a six-layer Graph Neural Network (GATv2 and LineEvo layers) with global pooling and MLP readout. Training minimizes binary cross-entropy loss on $S(m)$7, with $S(m)$8. Models are pre-trained on large reactant–product corpora (e.g., ~3.2M USPTO pairs), then fine-tuned with a small number of project- or domain-specific expert annotations. Monte Carlo dropout at inference provides uncertainty estimates.

FSscore excels at distinguishing fine-grained synthetic complexity, including chirality and chemical space specialization, especially after domain-specific fine-tuning. In applications, FSscore yields higher precision@k in filtering, improves hit rates in generative model optimization, and enables nuanced prioritization of proposed molecules (Neeser et al., 2023).

4. Algorithmic Workflows for Route- and Learning-Based SAS

Route-Based SAS (Round-Trip Score)

The route-based approach consists of the following stages:

1. Retrosynthetic Analysis: For the query molecule $S(m)$9, generate up to $m$0 synthetic routes using the retrosynthetic planner.
2. Forward Simulation: For each route, extract starting materials and simulate the forward reaction sequence via the trained reaction predictor, generating $m$1.
3. Similarity Evaluation: Compute the Tanimoto similarity between $m$2 and each $m$3.
4. Score Selection: Define $m$4 as the maximum similarity across the $m$5 candidate routes.

Learning-Based SAS (FSscore)

FSscore involves:

1. Pre-Training: Train the GNN model to minimize binary cross-entropy loss on pairwise reactant–product preferences extracted from large-scale reaction datasets.
2. Fine-Tuning: Select molecular pairs with high prediction uncertainty in target chemical spaces; collect expert preference labels; update later GNN layers and MLP to specialize $m$6.
3. Inference: Rank and prioritize candidate molecules by their FSscore; estimate confidence using MC dropout.

5. Comparative Performance and Application Benchmarks

SAS metrics are evaluated using metrics such as accuracy (Acc), AUC of sigmoid-scaled preference differences, precision@k, and success rates in generative loop integration.

Key quantitative findings include:

• The round-trip score $m$7 exhibits clear bimodal separation in retrosynthesis test sets: successful route-finding aligns with $m$8, while failures have $m$9 (Liu et al., 2024).
• In structure-based drug design (SBDD) benchmarks, the success rate (SSR) and top-k route quality, based on $m$0 ($m$1), vary substantially across generative models. For example, SSR for Pocket2Mol is 45.98%, while LiGAN is 9.68%.
• FSscore, when fine-tuned with as few as 20–70 labelled pairs, achieves substantial gains in out-of-distribution applicability (e.g., AUC 0.92 for natural products, AUC 0.95 for chiral discrimination) (Neeser et al., 2023).
• Integration of FSscore as a reward in reinforcement learning generative loops (e.g., with REINVENT) yields higher rates of synthesizeable molecules and more tractable reaction paths.

Metric	Route-based (Round-trip)	FSscore (Human-in-the-loop GNN)
Underlying Model	Retrosynthesis + Forward Reaction	Pairwise Ranking via GNN
Data Requirements	Reaction route database	Large paired corpus + domain expert pairs
Output	Tanimoto similarity score	Scalar preference score
Domain Adaptability	Only as good as route coverage	Directly adaptive via fine-tuning

6. Strengths, Limitations, and Future Directions

SAS methodologies exhibit distinct advantages and weaknesses:

Strengths of Route-Based SAS:

• Provides a direct in silico test of synthetic feasibility by enforcing route generation and product verification.
• Strongly correlates with actual route-finding success (large $m$2 between viable and nonviable compounds) (Liu et al., 2024).
• Clarifies failure modes that fragment heuristics cannot resolve.

Limitations of Route-Based SAS:

• Dependent on the scope and completeness of reaction data (USPTO or equivalent).
• Propagated errors from retrosynthesis models and forward predictors affect reliability.
• Search parameters (beam width, depth) can artificially limit feasible route identification.

Strengths of Learning-Based SAS:

• Tunable to specialized or emerging chemical spaces via expert annotation (Neeser et al., 2023).
• Discriminates subtle features (e.g., stereochemistry, ring strain) not captured in fragments.
• Provides uncertainty quantification.

Limitations of Learning-Based SAS:

• Pre-trained models may underperform on unfamiliar chemistries without tailored fine-tuning.
• Performance constrained by the consistency and scope of expert-provided preference data.
• Primarily ranks rather than quantitatively estimates absolute synthetic difficulty.

Anticipated Directions:

• Route-based SAS may benefit from richer proprietary reaction corpora, template-free or retrieval-augmented planners, joint training of forward and backward models, and integration of route-level uncertainty measures.
• Learning-based SAS methods may be strengthened by multi-expert labeling, richer uncertainty quantification, and the integration of retrosynthesis-derived features (Liu et al., 2024, Neeser et al., 2023).

7. Applications and Outlook

Synthetic accessibility scores are integral to:

• Prioritizing hits in virtual screening pipelines for lead optimization and hit-to-lead campaigns.
• Filtering generative model outputs in de novo design, ensuring tractable synthesis for suggested molecules.
• Informing the reward function in reinforcement learning-driven molecule generation.
• Evaluating structural modifications—such as scaffold hopping, chiral switching, or macrocycle formation—during hit expansion and optimization cycles.
• Adaptive, project-specific ranking through domain-expert-in-the-loop protocols.

A plausible implication is that, as synthetic accessibility metrics become more data-driven and context-aware, they will play an increasingly critical role in closing the design–make–test–analyze cycle of rational molecular discovery, minimizing wasted resources on synthetically intractable candidates.

Key references:

"Evaluating Molecule Synthesizability via Retrosynthetic Planning and Reaction Prediction" (Liu et al., 2024) "FSscore: A Machine Learning-based Synthetic Feasibility Score Leveraging Human Expertise" (Neeser et al., 2023)