Evaluation of Active Learning Selection Strategies and Characterization of Informative Sequences for Sequence-to-Expression Models

DNA sequence-to-expression models have advanced rapidly, yet they still generalize poorly beyond their training distribution, limiting their use for tasks such as variant effect prediction. Active learning has improved data efficiency across many machine learning domains, but no large-scale study has benchmarked selection strategies for sequence-to-expression models using real experimental data or characterized the sequences they select. We benchmarked six active learning strategies across diverse model architectures, datasets, and configurations. All strategies outperformed random sampling, with uncertainty-based methods performing best. Most of the gains achievable through many small acquisition rounds could be matched with fewer, larger rounds, making lab-in-the-loop workflows experimentally practical. Different strategies selected substantially overlapping sets of sequences that occupied distinct regions of sequence space and were enriched for higher expression, specific dinucleotide compositions, and denser transcription factor binding sites. Nevertheless, active learning consistently outperformed selection based on these biological properties alone, indicating that informativeness is not fully captured by any single feature. Together, our results establish active learning as a critical tool for improving sequence-to-expression models, identify biological signatures of informative sequences, and lay the foundation for iterative lab-in-the-loop refinement.