Bemis–Murcko Scaffolds

• Bemis–Murcko scaffolds are a core molecular framework comprising all ring systems and connecting linkers, with acyclic side chains removed.
• The extraction process uses graph algorithms to identify ring atoms and shortest paths, implemented in tools like RDKit for consistent scaffold mapping.
• These scaffolds enable advanced chemical space analysis, improving de novo molecular design, QSAR modeling, and property prediction in drug discovery.

A Bemis–Murcko (BM) scaffold is a graph-theoretical abstraction of a molecule representing its core framework—defined as the union of all ring systems plus linkers connecting them, with all pendant, acyclic side chains removed. Originally formulated to systematize the enumeration and comparison of compound classes in medicinal chemistry, BM scaffolds encode the “ring + linker” molecular core used to anchor similarity, diversity, and property analyses across drug, bioactive, and generative chemical spaces. This approach rigorously formalizes the notion of core chemotype, permitting quantitative analysis, machine learning, and generative modeling of chemical space at levels reflecting true structural novelty, rather than trivial side-chain decoration.

1. Formal Definition and Computational Extraction

The algorithmic extraction of a BM scaffold proceeds from the molecular graph $G = (V, E)$ (with $V$ as the atom set and $E$ as the bond set). The core procedure identifies all atoms $R(G) \subset V$ lying on at least one cycle (i.e., ring atoms, using e.g., SSSR or other cycle-finding algorithms). The BM scaffold $S(G)$ is then the induced subgraph on the set: $$V_S = R(G) \cup \left\{ v \in V \setminus R(G)\;|\;v\text{ lies on some simple path connecting two atoms in }R(G) \right\}$$ This corresponds to all ring atoms plus linker atoms that connect rings, explicitly excluding terminal substituents. An iterative algorithm for extraction marks ring atoms, finds all shortest paths among them, and accumulates the union of these path nodes; equivalently, one can prune degree-one (leaf) vertices recursively until only ring and linker atoms remain. Such definitions are directly implemented in cheminformatics toolkits such as RDKit via MurckoScaffold.GetScaffoldForMol(mol), ensuring consistent scaffold extraction from SMILES representations (Pearce et al., 28 Dec 2025).

2. Mathematical Structure and Inclusion Relations

The set of all BM scaffolds in a dataset forms a partially ordered set under the inclusion operator $\sqsubseteq$, where BM scaffold $S_i$ is included in (or isomorphic to a subgraph of) $S_j$ if there exists an injective mapping of the vertices such that all edge relationships are preserved: $$S_i \sqsubseteq S_j \iff S_i \text{ is graph-isomorphic to a subgraph of } S_j$$ The class defined by scaffold $V$0 is

$V$1

This formalism permits the organization of molecules into scaffold classes based on strict core substructure, forming explicit parent–child relationships corresponding to hierarchical chemical families (Clyde et al., 2021).

3. Scaffold Hypergraph Framework

To encapsulate all possible inclusion relationships, scaffold instances are encoded as a hypergraph $V$2, where $V$3 is the set of all unique BM scaffolds, and $V$4 is a family of hyperedges. Each compound $V$5 is associated with a hyperedge $V$6 that connects the full chain of nested scaffolds leading from the minimal subscaffold $V$7 through successive embeddings to the compound’s BM scaffold $V$8: $V$9 This construction encodes every inclusion path and supports efficient representation and traversal of the scaffold space, reflecting the chemical core hierarchy present in the data (Clyde et al., 2021).

4. Embedding Scaffolds and Compounds

Once the scaffold hypergraph $E$0 is constructed, scaffolds are embedded into a continuous $E$1-dimensional Euclidean space via a map $E$2 using a hypergraph-smoothness objective: $E$3 Minimizing this objective ensures that chemically related scaffolds (frequently co-occurring in molecule hyperedges) are placed close together in latent space. Compound vectors are then constructed as averages over their scaffold-chain embeddings: $E$4 Optionally, these embeddings can be optimized end-to-end with supervised property prediction losses, enhancing their relevance for chemical property and activity modeling (Clyde et al., 2021).

5. Applications in Molecular Design and Property Prediction

BM scaffold analysis is widely used in quantitative structure-activity relationship (QSAR) modeling, lead optimization, and generative molecular design:

• Property Prediction: Models built upon scaffold-embedding features consistently outperform classical Morgan-fingerprint and graph-neural-fingerprint baselines on benchmarks such as ESOL, FreeSolv, and BBBP. Notably, scaffold-based embeddings yield better generalization under scaffold-based train/test splits, as they encode the true core–subcore hierarchy. For example, scaffold-embedding models reduced RMSE by ≈10% on the BBBP dataset relative to the best graph convolutional networks (Clyde et al., 2021).
• De Novo Molecular Generation: In generative frameworks, BM scaffolds enable structural novelty assessment. For instance, in odorant molecule generation, every generated candidate is analyzed for scaffold novelty by extracting its BM scaffold and string-matching to training and reference databases. One report found that, using a VAE-QSAR generative pipeline, 74.4% of generated molecules possessed novel BM scaffolds not previously observed in the training or external reference sets—demonstrating exploration beyond mere substituent permutations (Pearce et al., 28 Dec 2025).

6. Metrics for Scaffold-Based Novelty and Analysis

BM scaffold extraction provides a single, canonical scaffold for each molecule. Scaffold novelty is determined via exact match to one or more reference sets, producing a mutually exclusive categorization. In the context of generative odorant discovery (Pearce et al., 28 Dec 2025), the categories and associated statistics are:

Category	Fraction of Molecules	Mean MW (Da)
Exact Memorization	5.34%	142.8
Odorant Derivatization	17.33%	181.5
Repurposing (ChemBL)	1.35%	158.3
Validated Scaffold Hop	1.54%	153.3
Uncharted Scaffold Hop	74.43%	160.8

These statistics provide evidence that the majority of generated compounds achieve genuine scaffold novelty. Analysis of physicochemical parameters within each category demonstrates that even uncharted scaffolds remain within targeted volatility and size domains, supporting both viability and functional relevance. No additional distance-based or statistical metrics are universally adopted; analysis is typically based on scaffolds as categorical features (Pearce et al., 28 Dec 2025).

7. Interpretation and Impact on the Exploration of Chemical Space

BM scaffolds operate as a structural filter, abstracting away molecular details to quantify chemical core diversity. Their use enables explicit evaluation of the extent to which generative models, optimization pipelines, or clustering algorithms move beyond established core scaffolds into unexplored structural regimes. The scaffold hypergraph and its continuous embeddings further enable the principled navigation and interpolation of scaffold space, supporting property optimization and de novo design under structural constraint. In generative odorant design, scaffold-based novelty assessment confirms that advanced VAE-QSAR pipelines discover not only derivatives of known scaffolds but also large numbers of entirely novel core frameworks. The intersection of BM scaffold formalism with hypergraph-based learning frameworks provides a robust mathematical and computational basis for understanding and expanding accessible chemical space in drug discovery, fragrance, and broader molecular design contexts (Clyde et al., 2021, Pearce et al., 28 Dec 2025).