Abstract
Scaffold based drug discovery (SBDD) is a technique for drug discovery which pins chemical scaffolds as the framework of design. Scaffolds, or molecular frameworks, organize the design of compounds into local neighborhoods. We formalize scaffold based drug discovery into a network design. Utilizing docking data from SARS-CoV-2 virtual screening studies and JAK2 kinase assay data, we showcase how a scaffold based conception of chemical space is intuitive for design. Lastly, we highlight the utility of scaffold based networks for chemical space as a potential solution to the intractable enumeration problem of chemical space by working inductively on local neighborhoods.
Introduction
The search for novel therapeutic agents is fundamentally a "needle in a haystack" problem on a cosmic scale. With an estimated $10^{60}$ possible drug-like molecules, the chemical universe is far too vast for exhaustive exploration. Even with modern supercomputing, high-throughput virtual screening (VS) can only simulate the binding of a tiny fraction of these molecules against protein targets. Traditional VS methods often treat chemical libraries as flat, unstructured lists, forcing researchers to choose between computationally expensive exhaustive screening or random sampling that frequently misses high-potency candidates. This paper introduces the Scaffold-Induced Molecular Graph (SIMG), a framework designed to navigate chemical space inductively by moving from known "good" neighborhoods to superior ones.
The Scaffold-Induced Molecular Graph (SIMG)
The core innovation of this work is the formalization of chemical space into a searchable network based on structural "skeletons," or Bemis-Murcko scaffolds. These scaffolds consist of the core ring systems and linkers of a molecule, stripping away sidechains to reveal the underlying architecture.
In the SIMG framework, nodes represent either individual molecules or their constituent scaffolds, while edges represent inclusion relationships. This creates a hierarchy where a complex molecule is linked to its simpler scaffold predecessors (e.g., a 3-ring system linked to its 2-ring components). This organization allows the library to be viewed as a set of local neighborhoods, where structural similarity correlates with biological activity.
Search and Prediction Strategies
To navigate this graph efficiently, the researchers implemented two primary strategies:
- Random Walks with Mixing: To identify high-affinity hits, a "walker" traverses the graph starting from a random molecule. The strategy employs a mixing parameter: half the time, the walker moves to a random neighbor to explore new territory; the other half, it moves toward the neighbor with the best docking score to exploit known high-performing regions. This balance prevents the search from getting stuck in local optima while ensuring it gravitates toward "privileged" scaffolds.
- Label Propagation with Residuals: For property prediction, a hybrid machine learning approach was used. A baseline neural network (MLP) predicts docking scores, and a graph-based algorithm then "smooths" the prediction errors (residuals) across the network. This assumes that molecules sharing scaffolds will exhibit similar deviations from the baseline model, refining accuracy across the neighborhood.
Experimental Results and Validation
The SIMG framework was validated using the Mcule database (6.3 million compounds), SARS-CoV-2 Mpro docking data, and JAK2 kinase assay data.
The results demonstrate a massive efficiency boost. Using the random walk strategy on the SARS-CoV-2 dataset, the team identified the highest-scoring compound classes by docking only 5,000 molecules—just 0.1% of the total library. As shown in the optimization curves, the SIMG-guided search consistently found molecules with significantly better docking scores (reaching -13.5) compared to the best hits found via random sampling (which plateaued above -12.5) within the same computational budget.
Furthermore, biological validation on the JAK2 dataset confirmed that the graph successfully clusters biologically similar molecules. The standard deviation of activity within scaffold "communities" was roughly half that of the dataset as a whole, proving that chemical space is indeed "smooth" when viewed through the lens of scaffolds.
Conclusion
The Scaffold-Induced Molecular Graph offers a scalable solution to the intractable enumeration problem of chemical space. By working inductively on local neighborhoods, SIMG allows for the discovery of privileged structural cores and facilitates "scaffold-hopping" to find diverse drug classes. Future work will focus on incorporating electrostatic data into graph edges and integrating SIMG with generative models to automatically design new molecules within high-performing scaffold neighborhoods.
Summary updated