Digital Genetic Diagnosis of Malaria Using Explainable Deep Learning
TL;DR
This study provides the first proof-of-concept that lightweight, interpretable CNNs can bridge molecular parasitology and digital pathology by enabling the inference of parasitic genetic activity from standard blood smear images, offering a scalable, low-cost diagnostic adjunct suitable for resource-limited settings and introducing a novel paradigm for digital genetic diagnosis in infectious disease pathology.
Abstract
Malaria, caused by Plasmodium spp., remains a leading cause of mortality in sub-Saharan Africa, with an estimated 282 million cases and 610,000 deaths in 2024. The var gene family primarily drives the virulence of P. falciparum, encoding the polymorphic adhesin PfEMP1, which mediates cytoadherence and immune evasion while inducing a quantifiable morphological footprint on host erythrocytes, including knob formation and altered deformability. However, current diagnostic tools remain decoupled from this genotype–phenotype nexus, and conventional convolutional neural networks treat all infected cells as a homogeneous class while their "black-box" nature obstructs clinical translation. This study aimed to develop a lightweight, interpretable CNN framework that classifies malaria-infected erythrocytes and integrates explainable AI (XAI) to semi-quantitatively infer var gene expression activity (PfEMP1 surface density) directly from cellular morphology—a concept defined as digital genetic diagnosis. A streamlined CNN architecture was trained on the NIH malaria dataset (27,558 single-cell images) for binary classification, employing feature activation mapping for interpretability and developing a novel Digital Protein Expression Density (DPED) algorithm. The model achieved an overall accuracy of 90.8% (sensitivity: 96.2%; specificity: 86.6%; ROC-AUC: 0.9724), with XAI activation maps demonstrating selective attention to intra-erythrocytic parasites and membrane regions consistent with knob architecture, while DPED maps successfully delineated high-density regions spatially correlated with predicted PfEMP1 anchorage sites, establishing a quantifiable link between routine microscopy and inferred genetic activity. This study provides the first proof-of-concept that lightweight, interpretable CNNs can bridge molecular parasitology and digital pathology by enabling the inference of parasitic genetic activity from standard blood smear images, offering a scalable, low-cost diagnostic adjunct suitable for resource-limited settings and introducing a novel paradigm for digital genetic diagnosis in infectious disease pathology.