AI-driven identification of Zika virus inhibitors

Research engaged with AI has dramatically increased due to a huge improvement of technologies in terms of cost and time-effectiveness as well as decreasing use of animals. High throughput screening (HTS) is a frequently used method in drug discovery. Here, we combined two powerful tools for AI-driven identification of Zika virus (ZIKV) inhibitors. The main complication from ZIKV is microcephaly and other related birth defects when pregnant women become infected with this virus. To date, there are no specific vaccines and antiviral drugs for the treatment of ZIKV infection. We have successfully developed an HTS (1536 well-plate format) cell viability-based zika virus cytopathic effect (CPE) assay. To evaluate an AI-driven method, we implemented the CPE assay into a miniaturized format and screened 1280 unique compounds. In addition, we incorporated an ultra-high throughput imager for 1536 well-plate HTS to rapidly obtain high-quality, brightfield images which were then analyzed by two different platforms. One platform, AVIA, uses deep convolutional neural networks (CNNs) to automate viral infectivity scoring based on CPEs long before signs of infection are visible to the human eye. The second, AutoHCS, utilizes a segmentation-free feature classification approach to develop probabilistic phenotypic profiles of antiviral compounds. Ultimately, AVIA was able to distinguish between infected and uninfected cells with high accuracy as early as 40 hours post infection, demonstrating a meaningful reduction in assay duration compared to the CPE readout. Furthermore, phenotypic profiling of 1280 compounds using AutoHCS phenotypic profiling overlaid with AVIA infectivity scores identified many compounds that showed possible anti-viral effect. Interestingly, 12 out of 20 compounds that were identified as hit compounds in the original cell viability-based CPE assay were found by AVIA and AutoHCS, demonstrating reasonable concordance and reliability of this method. Further, by comparing AVIA infectivity scores against AutoHCS morphological profiles, we were able to also eliminate compounds as false-positives without additional counterscreens or orthogonal assays. Overall, we demonstrate a robust, sensitive assay that produces high quality imaging data which results in promising cost- and time-effective HTS when combined with modern AI/ML tooling.