Fig. 4

Comparison of positive sample acquisition rates across different feature representations and acquisition functions on the ClinTox dataset , where mean and standard error is computed across 10 seeds. The plot shows cumulative toxic compound identification starting from a balanced initial set (50 positive, 50 negative). BERT-EPIG demonstrated a 2-fold improvement over Random sampling by identifying 70% of toxic compounds (gray horizontal line) in only 266 iterations compared to approximately 600 iterations for Random sampling, demonstrating better exploration of the chemical space when starting with limited labeled data