Synthetic Accessibility Scores in Molecules & Materials

• Synthetic accessibility scores are scalar metrics that quantify the ease of synthesizing molecules or materials using structural, heuristic, or machine-learned approaches.
• They integrate methods from graph neural networks to explicit retrosynthetic route simulations, providing actionable insights for experimental prioritization.
• In materials discovery, rank-averaged ensemble methods combining compositional and structural scores yield high predictive power, with evidence of up to 44% success in novel synthesis.

Synthetic accessibility scores quantify the ease or plausibility with which a candidate molecule or material can be synthesized in the laboratory. These scores serve as critical decision metrics for prioritizing compounds in chemical and materials discovery pipelines, aiming to bridge the gap between computational prediction and experimental feasibility. Synthetic accessibility scoring frameworks draw on diverse computational paradigms—from heuristic complexity estimation and machine-learned ranking to explicit simulation of retrosynthetic pathways—and are evaluated in both small-molecule (drug-like) and inorganic materials discovery contexts.

1. Formal Definitions and Conceptual Frameworks

Synthetic accessibility (SA) scores are typically formulated as scalar functions that map candidate chemical entities (molecules or crystals) to a real-valued or probability metric representing their synthetic difficulty or likelihood of successful laboratory realization.

• For molecules, FSscore defines a ranking function $f: \text{Molecules} \rightarrow \mathbb{R}$, learned via a pairwise preference paradigm where $FSscore(x) = \hat{s} = \varphi(\mathrm{GNN}(x))$, with $\hat{s}$ representing synthetic ease (Neeser et al., 2023).
• For inorganic crystals, the compositional and structural synthesizability scores model the conditional probability that a structure or composition is experimentally accessible, e.g.,
  • $s_c(x) = \sigma( h_c^\top f_c(x_c; \theta_c) )$ (composition-based)
  • $s_s(x) = \sigma( h_s^\top f_s(x_s; \theta_s) )$ (structure-based)
  • where $\sigma$ is the logistic sigmoid, $f_c$ and $f_s$ are neural embeddings, and $h_c$, $h_s$ are prediction heads (Prein et al., 3 Nov 2025).
• For data-driven route metrics, the round-trip score models synthetic feasibility as the ability to reconstruct the target from commercially available precursors via retrosynthesis and reaction prediction:
  • $FSscore(x) = \hat{s} = \varphi(\mathrm{GNN}(x))$0
  • where $FSscore(x) = \hat{s} = \varphi(\mathrm{GNN}(x))$1 is the retrosynthetic planner, $FSscore(x) = \hat{s} = \varphi(\mathrm{GNN}(x))$2 the forward reaction predictor, and $FSscore(x) = \hat{s} = \varphi(\mathrm{GNN}(x))$3 a fingerprint similarity measure (Liu et al., 2024).

These definitions underscore a fundamental spectrum: from rapid, structure-only surrogates to computationally intensive but more chemically specific route-based metrics.

2. Heuristic and Learning-Based Molecular Scores

Traditional SA scores, such as the Ertl–Schuffenhauer metric, utilize fragment-based approaches and molecular complexity heuristics. However, these approaches are agnostic to the connectivity of reaction knowledge and are limited in reliably discriminating complex or novel scaffolds (Liu et al., 2024). FSscore advances beyond heuristic indices by leveraging large reaction datasets and supervised learning:

• FSscore is trained on pairs of reactant–product transformations sampled from USPTO_full and CJHIF corpora, under the assumption that synthetic effort increases from reactant to product.
• Graph neural network (GNN) models with GATv2 and LineEvo layers embed molecular graphs; a scalar synthesizability is produced via a multilayer perceptron.
• Training employs a binary cross-entropy loss over pairwise human or reaction-derived preference labels.
• FSscore admits domain adaptation through active learning: uncertain or chemically focused molecule pairs (e.g., with specific chirality, ring strain, or functional types) are labeled by experts, and the model is fine-tuned to reflect nuanced chemist insight (Neeser et al., 2023).

FSscore can thus be configured for enhanced discrimination in out-of-distribution regions or specialized chemotypes by iterative expert feedback.

3. Retrosynthetic Route–Based Approaches

Route-based SA metrics explicitly model the plausibility and reachability of a molecule from available precursors using retrosynthesis algorithms and forward reaction prediction:

• The round-trip score, as defined for SDDBench, operationalizes synthesizability as the maximal Tanimoto similarity between the original molecule $FSscore(x) = \hat{s} = \varphi(\mathrm{GNN}(x))$4 and the reconstituted product $FSscore(x) = \hat{s} = \varphi(\mathrm{GNN}(x))$5 after applying a retrosynthetic planner ($FSscore(x) = \hat{s} = \varphi(\mathrm{GNN}(x))$6) and a forward reaction predictor ($FSscore(x) = \hat{s} = \varphi(\mathrm{GNN}(x))$7) along the proposed route tree:

$FSscore(x) = \hat{s} = \varphi(\mathrm{GNN}(x))$8

• The retrosynthetic planner (Neuralsym) operates as a template-based, neural-symbolic classifier trained on the full USPTO corpus. The forward model—a transformer decoder analogous to the Molecular Transformer—predicts reaction products from reactants.
• Pipeline steps:
  1. Enumerate up to $FSscore(x) = \hat{s} = \varphi(\mathrm{GNN}(x))$9 retrosynthetic routes for $\hat{s}$0.
  2. Apply forward prediction to regenerate $\hat{s}$1 from route leaves.
  3. Score using Tanimoto similarity; overall SA is determined by the best route ($\hat{s}$2 for a synthesizability threshold $\hat{s}$3).

Unlike static scoring, this methodology grounds accessibility in the demonstrated ability to traverse actionable synthetic routes, enhancing discriminative power for practical synthesis planning (Liu et al., 2024).

4. Synthetic Accessibility in Materials Discovery

Synthetic accessibility scoring has distinct requirements in computational materials science, where both composition and structure must be considered:

• Prein et al. model synthesizability for crystal structures with dual scoring:

  • Compositional score ($\hat{s}$4): based on stoichiometry vectors and engineered features (e.g., electronegativity, valence counts) via a fine-tuned MTEncoder transformer.
  • Structural score ($\hat{s}$5): based on relaxed DFT geometries, encoded as crystal graphs and processed with E(3)-equivariant GNNs (JMP model).
• Rank fusion of both outputs (using Borda count) yields a single "RankAvg" metric, with empirical guidance that $\hat{s}$6 (top $\hat{s}$70.1% of candidates) is a robust indicator of high synthesizability.
• Model training/validation leverages the Materials Project, with positive samples defined via the ICSD cross-reference (flagging experimental realization). The model achieves $\hat{s}$8 and ROC AUC $\hat{s}$9 on held-out test data.
• High RankAvg candidates drive experimental synthesis, with a one-shot success rate of 44% for predicted novel phases, demonstrating substantial translational utility in the materials domain (Prein et al., 3 Nov 2025).

5. Comparative Evaluation of Approaches

The table below synthesizes key characteristics of modern synthetic accessibility scores, as reported in the cited works:

Method	Scope	Principle	Quantitative Performance
Ertl–Schuffenhauer SA	Molecule	Fragment complexity, heuristics	ROC-AUC (drugs vs. natural): 0.83
FSscore	Molecule	GNN pairwise ranking, learnable	Acc (test): 0.905, AUC: 0.971
Round-trip Score	Molecule	Retrosynthesis + reaction prediction	Top-1 success (Pocket2Mol): 11.35%
Compositional	Material	Pretrained transformer, stoichiometry	F₁: 0.741, ROC AUC: 0.91
Structural	Material	E(3)-GNN, crystal graph	F₁: 0.773, ROC AUC: 0.93
RankAvg Ensemble	Material	Borda fusion (composition+structure)	F₁: 0.789, ROC AUC: 0.95

Performance metrics for FSscore and material accessibility scores are drawn directly from large-scale empirical benchmarks (Neeser et al., 2023, Prein et al., 3 Nov 2025). The round-trip score uniquely reflects explicit route feasibility and ranks SBDD generators by the proportion of designs with at least one high-similarity route ($s_c(x) = \sigma( h_c^\top f_c(x_c; \theta_c) )$0) (Liu et al., 2024).

6. Limitations, Integration Guidelines, and Outlook

Key limitations and guidelines for practical deployment of synthetic accessibility scores include:

• Generality and Domain Shift: Heuristic or learned models—FSscore, SA, SCScore—require fine-tuning or retraining for new chemical spaces with significant out-of-distribution risk (e.g., natural products, PROTAC scaffolds) (Neeser et al., 2023).
• Route Model Bias: Round-trip scoring can be impacted by forward model or retrosynthesis planner errors, incomplete reaction coverage, and misalignment with actual laboratory conditions (e.g., yield, cost) (Liu et al., 2024).
• Data-Driven Success: For materials, integrating rank-averaged compositional and structural probabilities with rigorous database labeling enables efficient experimental validation, with a pragmatic "top-K" selection scheme (e.g., $s_c(x) = \sigma( h_c^\top f_c(x_c; \theta_c) )$1) for high-yield synthesis targeting (Prein et al., 3 Nov 2025).
• Fine-Tuning Protocols: Active learning with expert feedback enhances discrimination for subtle attributes (chirality, ring strain) and improves reliability for RL-driven molecule and library optimization (Neeser et al., 2023).
• Pipeline Integration: Synthetic accessibility scores can be implemented as explicit ranking heads, filtering criteria, or RL rewards in generative design workflows. For round-trip methods, inference times are compatible with typical high-throughput screening on modern accelerators, making them suitable for post hoc selection (Liu et al., 2024).

A plausible implication is that future improvement of SA scores will require deeper integration of experimentally-informative features (e.g., real-world yields, reaction costs), larger and more diverse synthesis datasets, and unified protocols for simultaneous optimization of desired physical/biological properties and synthetic tractability.