Phenotypic-driven drug discovery is gaining popularity due to the advances in high-content imaging and machine learning, particularly for predicting compound Mode of Action (MoA) and properties. However, reliance on biochemical assays for label acquisition introduces noise and sparsity, complicating reliability estimation in traditional discriminative models. In this work, we propose a High Content Screening Diffusion Classifier (HCS-DFC), reformulating prediction as a conditional generation task to inherently model label distributions and co-dependencies without requiring calibration datasets. By leveraging diffusion models’ ability to capture complex data distributions, HCS-DFC outperforms conformal prediction methods in reliability estimation and achieves state-of-the-art accuracy on synthetic (MNIST-based multi-task classification) and real-world cell painting datasets.