Paper review · October 17, 2026
Not All Unlabeled Data Are Equal: Selecting Unlabeled Data for Semi-Supervised QSAR Modeling Using Large Chemical Databases
Not All Unlabeled Data Are Equal: Unlabeled Data Selection for Semi-Supervised QSAR Modeling with Large Chemical Repositories
The authors selected over 400,000 unlabeled molecules from ChEMBL in 4 ways and fed them into semi-supervised QSAR models. Selecting a small number of chemically distant molecules gave the most stable performance gains.
Problem
QSAR (models that predict biological activity from molecular structure) has little labeled data because measuring activity requires costly experiments. Chemical repositories such as ChEMBL, by contrast, contain millions of unlabeled molecules. Whether this data can be used for semi-supervised learning, and if so which unlabeled data to select, has received little study.
Method
The authors used 10 QSAR datasets drawn from ChEMBL and a pool of more than 400,000 unlabeled molecules. They combined 3 existing semi-supervised learning algorithms (self-training, co-training, label propagation) with 4 unlabeled-data selection strategies (random, near neighbors, far neighbors, mixed sampling) and compared them with a supervised learning baseline.
Results
- The authors reported that, on all 10 datasets, at least one semi-supervised learning configuration outperformed the supervised baseline model.
- The most consistent improvements generally came from selecting a relatively small number of chemically distant molecules.
- The authors interpreted this as a sign that the composition of unlabeled data may matter more than its quantity, which runs against the common assumption that semi-supervised learning needs a lot of unlabeled data.
Limitations
This analysis is based on the abstract. Without the full text, we could not verify the size of the improvements, the evaluation metrics or the variation across datasets. According to the abstract, the experiments are limited to 10 datasets drawn from ChEMBL and 3 existing semi-supervised learning algorithms. Also, the finding that 'at least one configuration beats the baseline' may not carry over to real-world settings, where you cannot know in advance which configuration to choose.
Why it matters
In fields where experimental labels are expensive, such as early-stage drug discovery, the work points to a practical way to improve prediction models by using public chemistry databases at no extra cost. The point that 'choosing well' matters more than 'adding more' is also relevant to other data-scarce fields.
Key terms
- QSAR(Quantitative Structure-Activity Relationship)
- A model that predicts a molecule's biological activity (for example, how strongly it inhibits a target protein) from its chemical structure.
- Semi-Supervised Learning
- A way of training a model using a small amount of labeled data together with a large amount of unlabeled data.
- Self-training
- A semi-supervised learning technique in which a model assigns its own predicted labels to unlabeled data and then uses that data as training data again.
Original
Authors: Oliver Majer, Saso Dzeroski, Jurica Levatic · Posted 2026-10-15 · License arXiv 비독점 배포 라이선스 · Based on the abstract
This review is an AI-written summary of the paper. Please treat the original paper as the reference.