Few-Shot Bioactivity Prediction with Meta-Learning under Assay Heterogeneity

Meta-learning offers a promising approach to few-shot bioactivity prediction, but its effectiveness can be limited by heterogeneity across biological assays. We investigate the effect of increasing meta-training task heterogeneity and show that greater heterogeneity can degrade the performance of meta-learning methods. Motivated by this, we introduce an approach that explicitly accounts for assay relatedness by conditioning on auxiliary data from similar assays, enabling more effective few-shot bioactivity prediction.