Benchmarking a Modular Deep Learning Pipeline for High-Throughput ASC-Speck Image Classification in Drug Discovery

Abstract

Inflammasome drug screening using THP-1-ASC-GFP reporter cells produces fluorescence microscopy datasets that are difficult to assess manually in high-throughput screening. A modular four-step pipeline for automated ASC-speck classification in drug discovery is presented. The pipeline includes image acquisition, quality-controlled pre-processing, deep learning inference, and a decision output that is compatible with the Laboratory Information Management System (LIMS). Five architectures, EfficientNetB0, MobileNetV2, ResNet18, ResNet50, and ViT-B/16, were systematically benchmarked on 1,580 curated fluorescence micrographs (740 negative, 840 positive) under stratified five-fold cross-validation. EfficientNetB0 was also assessed using strict well-isolated group-stratified validation across 158 micro-wells. EfficientNetB0 provided the most favourable balance between predictive performance and computational efficiency with 89.56% accuracy, 96.47% precision, 83.57% recall, 96.35% specificity and an AUC-ROC of 0.9812 with 5.29 million parameters and 0.39 GFLOPs. Under grouped validation the accuracy was 88.15%, with little inflation of performance from replicate-level leakage. The highest AUC-ROC was achieved by ResNet18 with 0.9847 while MobileNetV2 provided a lower-complexity alternative. ViT-B/16 showed substantially lower balanced performance under the present data regime; this may reflect limited data scale and a possible mismatch between its 16-by-16 patch representation and the subcellular ASC puncta. Offline profiling showed that EfficientNetB0 processed images at 3.24 ms per image on an NVIDIA T4 GPU. Ambiguous predictions were retained for expert review using the confidence-based review band. Overall, the pipeline supports efficient primary screening with traceable outputs, although external multi-centre validation remains necessary before broader deployment. The framework also includes probability-calibration assessment (ECE = 3.42%), confidence reporting, and structured CSV/JSON export compatible with laboratory information management systems.