TL;DR
Modern AI-driven molecular embeddings often fail to capture basic graph topology, 3D shape, or chemical similarity better than "old-school" fingerprints and descriptors. Furthermore, in ultra-large chemical libraries, these representations suffer from massive "degeneracy," where millions of structurally distinct molecules map to identical digital signatures, potentially leading to significant errors in virtual screening and dataset splitting.
What Problem Does This Paper Address?
In the world of drug discovery, how we represent a molecule to a computer is everything. We generally use two approaches: classical fingerprints (hand-crafted bit vectors based on substructures) and learned embeddings (vectors produced by deep learning models like Transformers or GNNs).
While the field has rushed toward complex AI models, their evaluation is usually limited to performance on specific benchmark tasks. We don't actually know if these models "understand" the molecule's fundamental properties—like its 2D shape or 3D volume—better than the simpler methods they are replacing. This paper systematically investigates whether these representations are actually fit for purpose across three critical axes: graph topology, 3D shape, and chemical (ADMET) similarity.
Methods
The researchers conducted a massive cross-comparison of 13 different representations:
- 4 Classical Fingerprints: ECFP4, FCFP4, Atom Pair, and Topological Torsion.
- 9 Learned Embedding Models: Including Chemical Language Models (MolBERT, ChemBERTa, MolFormer), GNNs (MolCLR, CheMeleon, MiniMol), and Graph Transformers (UniMol, COATI, MolE).
To establish "ground truth" for graph similarity, they used a normalized version of the weighted Maximum Common Edge Subgraph (MCES) distance, an NP-hard metric that measures how many bonds must be removed to make two molecules identical. They also analyzed degeneracy—the phenomenon where different molecules produce the same representation—by processing the astronomical 10.8 billion molecule Enamine REAL database.
Datasets
- Enamine REAL (Q4 2025): 10.8 billion molecules (filtered to 10.3B) used to measure representation degeneracy at scale.
- MCES Dataset: 63,437 molecules with pre-computed pairwise graph distances used to evaluate topological perception.
- ZINC: 30,000 molecules used for 3D shape and pharmacophoric (color) overlap analysis.
- Biogen ADMET Dataset: A high-quality industrial dataset used to test how well representations group molecules with similar biological properties (like clearance or solubility).
Key Insights and Ideas
- The Degeneracy Crisis: In ultra-large libraries, fingerprints are surprisingly non-unique. For FCFP4, over 40% of the 10.3B molecules are "degenerate," meaning they share a representation with at least one other structurally distinct molecule. Even the "gold standard" ECFP4 (chiral) has 225 million degenerate molecules.
- AI "Blindness" to Topology: Most learned embeddings correlate poorly with actual graph structure. Some models, like MolE, were effectively uncorrelated with the MCES ground truth. This means AI models might see two very different scaffolds as nearly identical.
- Fingerprints Still Rule 3D: Despite being 2D representations, classical fingerprints (especially Atom Pair) tracked 3D shape and color overlap more consistently than almost all learned embeddings, with the notable exception of COATI.
- The Clustering Trap: The common practice of using ECFP4-based "Butina clustering" for dataset splitting leads to "structural leakage." Highly similar molecules often end up split between training and test sets, causing researchers to overestimate how well their models actually generalize.
Key Results
- Graph Correlation: Atom Pair fingerprints were the most strongly correlated with graph topology (Spearman $\rho = -0.82$). In contrast, 5 out of 9 learned representations had correlations weaker than -0.5.
- ADMET Performance: A simple vector of 217 RDKit physicochemical descriptors (PDV) outperformed all complex AI embeddings in placing molecules with similar ADMET properties close together in representation space.
- Failure Modes: The authors identified a taxonomy of "failure modes" where fingerprints fail to distinguish common medicinal chemistry modifications, such as carbon chain elongation or ring expansion.
Limitations and Future Work
The authors acknowledge that while count-based fingerprints might reduce degeneracy compared to the bit-based ones tested here, they are not a total cure. A major limitation of current learned models is their lack of "calibration" to bond structure; they are often trained on objectives (like predicting SMILES strings) that don't force the model to learn global topology.
The paper concludes that we need a "new wave" of molecular encodings. Future work should focus on developing representations that are explicitly designed to be unique (non-degenerate) and better aligned with the physical and topological realities of molecules, rather than just optimizing for downstream task accuracy.
Summary updated