Generalizing Beyond Known Scaffolds in Molecular Graphs

This direction is split into other directions: scaffold generalization gap. Its last version stays readable and citable here.

Version 1 of 1131 papers162 conceptsfrozen 2026-10-04

In drug discovery, the ultimate goal is "scaffold hopping"—finding a molecule with a completely new structural core that still performs a specific biological function. However, most machine learning models suffer from a "generalization gap": they perform brilliantly on molecules similar to those they have already seen but fail when encountering novel chemical skeletons.

This failure is often hidden by standard evaluation metrics. If we split a dataset randomly, the training and test sets will likely contain molecules with the same "scaffold" (the core ring system). To truly test a model's utility for discovery, researchers like Deng Cai and Chanhui Lee emphasize scaffold-based splits, where the test set contains entirely unseen structural families.

The Problem of Spurious Correlations

Why do Graph Neural Networks (GNNs) struggle with new scaffolds? The issue lies in what the model "attends" to. In a typical training set, certain scaffolds might be statistically correlated with a property (like toxicity) simply because of how the library was synthesized, not because the scaffold itself causes the effect.

As Qitian Wu and colleagues argue in their work on Adversarial Invariant Augmentation, models often overfit to these "environmental" features—the specific molecular framework—rather than the "invariant" features, such as the functional groups that actually drive chemical behavior. When the model encounters a new scaffold (a covariate shift), these learned correlations collapse.

Disentangling Chemical Logic from Structure

To solve this, recent research focuses on disentanglement: forcing the model to separate the "essence" of a chemical property from the "noise" of its structural container.

Aming Wu proposed a framework called CFD (Concept-Enhanced Feedback Disentanglement). Instead of treating a molecule as a single monolithic graph, CFD uses an iterative feedback loop to identify "molecule-aware concepts." By separating these invariant chemical concepts from idiosyncratic graph patterns, the model becomes robust to out-of-distribution (OOD) shifts, such as changes in molecular size or core framework.

Learning from Fragments and Substructures

Another powerful intuition is that while scaffolds change, the underlying "vocabulary" of chemistry—functional groups and fragments—remains relatively stable.

  • Fragment Contrast: Yuyang Wang developed iMolCLR, which uses contrastive learning at the fragment level. By teaching the model that the same functional group should have a similar representation regardless of the scaffold it's attached to, the model learns a more generalizable "chemical common sense."
  • Expert Routing: Tianyi Jiang introduced ASE-Mol, which uses a Mixture-of-Experts (MoE) architecture. It explicitly routes different substructures to specialized "experts" that determine if a fragment is beneficial or detrimental to a property, allowing the model to assemble a prediction from generalizable parts rather than memorizing a whole-molecule template.

Contextual and Relational Logic

The most difficult OOD challenge is not just the structure of a single molecule, but how it interacts with others. In Molecular Relational Learning (MRL), the "important" part of a molecule changes depending on its partner.

Namkyeong Lee addressed this with the Conditional Graph Information Bottleneck (CGIB). CGIB identifies "core subgraphs" that are relevant to a specific interaction. Similarly, Lin Li developed ReAlignFit, which mimics the biological "Induced Fit" theory. By dynamically aligning representations during an interaction, these models move away from static pattern matching and toward a more flexible, logic-based understanding of chemical relationships.

Benchmarking the Generalization Gap

The shift toward OOD robustness has led to more rigorous benchmarking. Yongqiang Chen introduced ReactionOOD, which categorizes distribution shifts into structure, condition, and mechanism. Their findings reveal a sobering reality: shifts in the underlying chemical mechanism are far harder for AI to handle than simple structural changes. This suggests that the next frontier for molecular ML isn't just better graph representations, but architectures that can capture the fundamental physical rules governing how molecules transform.

Papers in this version

Showing 60 of 131 papers.

History & citationVersion 1 · 2026-10-04

Split into other directions · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_generalizing_beyond_known_scaffolds_in_m_01d34828,
  title = {Generalizing Beyond Known Scaffolds in Molecular Graphs},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/generalizing-beyond-known-scaffolds-in-molecular-graphs/rdr_01d348282dbf44d183c5f99eed28c3f5},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-21},
}

This exact version (version 1, frozen text):

@misc{lacuna_generalizing_beyond_known_scaffolds_in_m_01d34828_v1,
  title = {Generalizing Beyond Known Scaffolds in Molecular Graphs},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/generalizing-beyond-known-scaffolds-in-molecular-graphs/rdr_01d348282dbf44d183c5f99eed28c3f5/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-21},
}