Should We Embed in Chemistry? A Comparison of Unsupervised Transfer Learning with PCA, UMAP, and VAE on Molecular Fingerprints

Dimensionality reduction is routine in cheminformatics, with PCA and molecular fingerprints as the go-to approaches. But does moving beyond these—via modern unsupervised embedding techniques and transfer learning—provide any practical benefit for chemical classification tasks? This paper systematically benchmarks linear PCA, nonlinear UMAP, and deep learning-based VAE as embedding methods for molecular fingerprints, specifically within the toxicology domain.

The core of the study compares

• baseline classification models using uncompressed molecular fingerprints,
• versus the same models trained on lower-dimensional embeddings generated by PCA, UMAP, and VAE.

In addition, the work assesses the effect of transfer learning: learning embeddings on an external, much larger chemical compound dataset, then applying them to toxicology prediction (Tox21 tasks). The experimental design is rigorous: different embedding dimensionalities, varying external dataset sizes (from 200 to 30,000 compounds), and multiple downstream classifiers (random forests, logistic regression, k-nearest neighbors).

A few findings stand out:

• UMAP has emerged as a viable nonlinear alternative to PCA in this setting, producing meaningful compressed representations and effective clusters, particularly when data distributions are complex. That said, the performance boost relative to PCA is situation-dependent and not universally dramatic.
• VAE’s generative model usually delivers the highest classification accuracy after compression. The improvement is especially notable when embeddings are learned on large, external datasets—demonstrating that transfer learning can enhance downstream model performance if sufficient external chemical diversity is available.
• Increasing the number of embedding dimensions (2–15) generally improves accuracy up to a plateau, most pronounced for UMAP and VAE. For PCA, benefit plateaus earlier and does not respond to more diverse data in the same way.

Curiously, for PCA, the size of the external dataset barely matters: principal components seem stable even with small samples. For nonlinear embedders (UMAP, VAE), more data almost always helps, but UMAP saturates faster.

No single method dominates across all tasks, but VAE with external training gives the clearest edge when maximizing accuracy is the aim. The take-home: embedding does help, but the optimal approach and the size of the benefit depend heavily on both the downstream task and the embedding algorithm’s fit to data complexity. Unsupervised transfer learning via advanced embedders is worth considering, but don’t expect a universal win—sometimes the old baselines are harder to beat than we hope.

@article{lovric2021should,
  title = {Should We Embed in Chemistry? A Comparison of Unsupervised Transfer Learning with PCA, UMAP, and VAE on Molecular Fingerprints},
  author = {Lovric, M. and \DJ{}uricic, T. and Tran, H. T. and Hussain, H. and Laci\'{c}, E. and Rasmussen, M. A. and Kern, R.},
  year = {2021},
  journal = {Pharmaceuticals},
  volume = {14(8)},
  pages = {758},
}