Automated deep learning tool for robust colony counting in vaccine immunogenicity bioassays

BACKGROUND: The microcolony multiplex opsonophagocytic killing assay (mMOPA) is a critical functional assay for evaluating pneumococcal vaccine immunogenicity; however, its colony-counting readout is scientist-dependent and time‑consuming. RESEARCH DESIGN AND METHODS: In this study, we developed and optimized a deep learning model (DLM) colony-counting tool by adapting the Segment Anything (SAM) foundation model and evaluated its performance using historical assay qualification datasets spanning 12 pneumococcal serotypes, with additional benchmarking against blinded human expert counts. RESULTS: The DLM-derived counts demonstrated strong correlation and agreement with both historical CTL ImmunoSpot® counts and expert counts. Notably, DLM outperformed CTL for a challenging serotype and produced comparable opsonophagocytic titers with those generated using CTL counts within the representative samples evaluated. These results establish the DLM approach as fit for purpose for mMOPA readouts. CONCLUSIONS: Overall, this work removes a key throughput bottleneck in mMOPA, enables a "walkaway" colony-counting workflow supporting improved operational efficiency and scalability for large clinical studies, and highlights the broader value of integrating bioanalytical expertise with modern computer vision to advance scalable, image-based bioassays in vaccine and drug development.