ANTI-VIRAL SCREEN: Next-Gen Phenotypic Profiling of Antivirals with AI-Driven Image Analysis
App Notes

ANTI-VIRAL SCREEN: Next-Gen Phenotypic Profiling of Antivirals with AI-Driven Image Analysis

By August 28, 2025October 8th, 2025No Comments

Next-Gen Phenotypic Profiling of Antivirals with AI-Driven Image Analysis

A Collaboration Between the Ultra-Fast Araceli Endeavor® Platform and ViQi’s Predictive Analytics

 

 

Introduction

In this joint application note, ViQi, Inc. and Araceli Biosciences demonstrate how combining ultra-fast, high-content imaging from the Araceli Endeavor® with ViQi’s powerful AI-driven analysis platforms—AVIA™ and AutoHCS™—can accelerate the discovery of effective antiviral compounds. By enabling rapid, live-cell, information-dense screening with automated, scalable analysis, this partnership delivers a next-generation approach to phenotypic profiling that is faster, more precise, and cost-effective. This workflow was deployed at the Herbert Wertheim UF Scripps Institute for Biomedical Innovation & Technology to identify candidate antivirals targeting the Zika virus.

The Challenge

Developing effective antiviral treatments is essential to human health, with the possibility of rapid viral evolution prioritizing fast, effective screening for antiviral compounds. Pre-COVID-19, only 90 antiviral drugs had ever been approved for use1, highlighting the need for more sophisticated basic drug discovery.

This application note details a screen to identify inhibitors against Zika virus, a mosquito-borne illness associated with neuropathy and myelitis in both adults and children (e.g. Guillain-Barré syndrome), as well as serious birth defects such as microcephaly. However, the limitations in the discovery and development of therapeutic antivirals are two-fold.

First, current strategies for measuring infectivity depend on TCID50 and plaque assays that can take up to 14 days to produce results.2 These types of endpoint assays may miss key timepoints and infection dynamics as well as introduce experimental bias through systematic errors associated with fixation and staining.

Second, large-scale screening for antiviral candidate compounds is limited by instrument speed and analytical bottlenecks. Rapid drug development depends on the expediency of identifying candidates from large library screens, but this cannot come at the expense of assay quality.

The Solution

Fast, high resolution live cell imaging, combined with automated, AI-based analysis.

ViQi, Araceli, and the UF Scripps HTS Center have collaborated to address the pain points in detecting viral infection and identifying effective and safe antiviral compounds. AVIA, an AI-based assay, uses state-of-the-art machine learning algorithms to analyze brightfield images and detect viral infectivity within hours, discarding long incubation times and multiple rounds of viral infection, which are part of the protocols for traditional assays.3 AVIA also eliminates user bias and standardizes infectivity assays in an objective manner. Because it is cloud-based, AVIA offers scalable storage and analysis infrastructure.4

When screening a large number of compounds as antivirals, fast live cell imaging combined with machine learning-based phenotyping yields an effective antiviral screen at scale. Imaging all cells per well in an entire 1536-well plate in less than 3 minutes at submicron resolution, the Araceli Endeavor high-content imager enables a multi temporal based live cell infectivity screen, preserving cell health while allowing more compounds to be screened in less time. This speed and quality is capitalized on when screening antivirals using the AutoHCS toolkit, which can precisely profile thousands of conditions in an unbiased phenotypic manner. It can detect not only reduced infectivity, but distinguish subtle indications of cellular stress. This permits researchers, like the team at UF Scripps, to eliminate antiviral candidates that cause unwanted cell stress in the first round of screening while focusing on promising therapies early in the drug discovery and development process.

Case Study Overview & Results

In search of effective antivirals against the Zika virus, UF Scripps Institute conducted a set of screens designed to use ViQi’s AI-based assays to determine and categorize degree of infectivity. The Araceli Endeavor was used for transmitted light imaging, yielding 1536-well plates in less than 3 minutes (Fig 1A,B). This quick scan time ensured cell health by minimizing time outside the incubator, enabling ultra-high- throughput live cell screening at seven timepoints across 72 hours. Capturing the entire well area with 0.27μm/pixel resolution provides reliable data to detect subtle effects, fueling sensitive machine learning (ML) analysis.

First, a single 1,536-well plate of Vero76 cells was infected with the Zika virus across 10 viral doses and imaged at 7 timepoints. This training plate was used to train AVIA to distinguish infected from uninfected cells (internally validated, >99% accuracy at 64 h.p.i, with a MOI prediction of 10.2±0.1 for the undiluted MOI of 10.0).5 External validation came from the controls in the following test plates: here, this model produced characteristic dose responses to increasing viral titer, with the infectivity curve steepening as time post infection increased, as expected (Fig 1C). As the incubation time increases, lower MOIs become detectable. We wanted to assay the antivirals at a low MOI to ensure we detect activity that prevents virus amplification. At 64h post-infection (h.p.i.), the 0.05 MOI is just below the upper plateau of the linear range, so this time point was chosen to assay the antivirals.

Second, a series of 1,280 antiviral candidates were applied to cells infected at MOI=0.05 to investigate reductions in infectivity. This screen consisted of three 1,536-well plates with compounds in triplicate, two negative control plates with only vehicle, and one positive control plate with 10 doses of known antiviral NITD- 008 applied to cells infected at MOI=0.05 (Fig 2A). Infectivity indices were predicted for images of the positive control plate using the trained AVIA model, demonstrating a reduction in infectivity as dose increases (Fig 2B, data shown for 64 h.p.i).

To further enrich phenotypic understanding of NITD-008, all doses of NITD-008, the DMSO/vehicle condition, and the untreated condition were fed as classes to train a single feature classifier in ViQi’s Automated High Content Screening (AutoHCS™) toolkit. AutoHCS uses the AI’s confusion between classes to construct a phenotypic dendrogram that describes the relationship between classes based on phenotypic similarity (Fig 2C, data shown for 64 h.p.i.). For NITD-008, three distinct phenotypic clusters emerged. One cluster (red) corresponds to an infected phenotype, with infected cells either untreated or treated with DMSO and low doses of the positive control. Another cluster (green) represents healthy cells,containing uninfected cells either untreated or treated with DMSO and an intermediate dose of NITD-008. Finally, a third cluster (blue) indicates that, at higher doses of NITD-008, off-target effects cause a phenotype that is distinct from both healthy, uninfected cells and cells infected by the Zika virus. NITD–008 has known cytotoxic effects, causing it to not pass clinical trials, which are likely the underlying cause of this distinct, high-dose phenotype. To identify antiviral candidates from the 1,280 compound screen, a single feature classifier was trained on each compound, all doses of NITD-008, and negative controls as classes. AutoHCS then generated a phenotypic dendrogram from prediction results (Fig 3A). This analysis shows five distinct clusters, which can be functionally annotated using controls. The red cluster contains infected cells untreated or treated with DMSO, indicating that all compounds in this cluster appear, phenotypically, the same as infected cells, eliminating these as successful antiviral candidates. The green cluster includes uninfected cells that were untreated or treated with DMSO and, therefore, represents healthy cells at this level of analysis: compound-treated infected cells that cluster here represent successful antiviral treatment with minimal unwanted effects. The two compounds that fall inside this cluster are both intermediate doses of the positive control, NITD-008. There are three other clusters (blue, purple, and yellow) that are phenotypically distinct from both healthy cells and infected cells. According to this relational dendrogram, the blue cluster is most phenotypically similar to the uninfected green cluster. AutoHCS strongly indicates that compounds in the blue cluster are the most likely to exhibit antiviral activity. Contrastingly, purple and yellow clusters are very phenotypically distinct from all other conditions. While treatments in this cluster may exhibit antiviral effects, the phenotype is impacted significantly, forcing these to cluster far away from healthy cells. These are very likely treatments exhibiting off-target effects that result in unwanted phenotypes.

To validate AI-based results, Cell TiterGlo (CTG) was applied to all wells following brightfield imaging on the Endeavor and luminescence was measured using a PHERAstar (BGM).6 The CTG analysis identified 20 hits (1.56% hit rate) that showed reduced infectivity (Fig 3B). All 20 of these antiviral candidates fall within the blue, yellow, and purple clusters on the AutoHCS dendrogram, indicating that AutoHCS not only validates, but expands upon CTG results. CTG is a binary analysis that does not account for off-target effects. The use of AutoHCS permits further filtering of “hit” compounds by eliminating antiviral candidates that fall into the yellow and purple clusters as derived from AI marginal probability scores. In other words, AutoHCS identifies false positives that occur within the CTG analysis.

The more promising subset of compounds (blue and green clusters) were analyzed in AutoHCS, finding 9 CTG hits clustering closest to healthy cells (Fig 3C). In this way, successful antiviral candidates can be ranked based on phenotypic similarity to healthy cells. Finally, antiviral hits can be further filtered using the AVIA inhibition assay (Fig 2A), because only images that appear phenotypically like uninfected cells receive a low infectivity score. Those with unwanted phenotypic effects will not score as similar to healthy cells. This is clearly demonstrated with four candidate hits hypothesized to be ATP-mimicking compounds that were identified as “hits” using CTG, but still received high infectivity scores from AVIA (Fig 3D).

(A) At UF Scripps, Vero76 cells were plated on 1536 well plates (563 cells/2.5 µL/well) and incubated for 24 hours at 37℃, 5% CO2. 1280 candidate antiviral compounds or vehicle were tested and plates incubated for 2 hours before application of Zika virus (MOI 0.05, 2.5 µL/well). A. At UF Scripps, Vero76 cells were plated on 1536 well plates (563 cells/2.5 µL/well) and incubated for 24 hours at 37 , 5% CO2. 1280 Plates were imaged in brightfield with one image per well using the Araceli Endeavor, capturing a majority of the well at submicron resolution candidate antiviral compounds or vehicle were tested and plates incubated for 2 hours before application of Zika virus (MOI 0.05, 2.5 µL/well). in
(A) Apparent viral infectivity phenotype rescued by application of known antiviral NITD-008 compared to DMSO alone. (B) AVIA quantification A. Apparent viral infectivity phenotype rescued by application of known antiviral NITD-008 compared to DMSO alone. B. AVIA quantification of of infectivity across 10 different concentrations demonstrates a dose response with a drop in infectivity at both 64 and 72 hours with increased infectivity across 10 different concentrations demonstrates a dose response with a drop in infectivity at both 64 and 72 hours with increased concentration, IC50=0.25 and 0.44µM, respectively. Plotted with mean of n=16 wells. (C) ViQi’s Automated High Content Screening assay concentration, IC50=0.25 and 0.44µM, respectively. Plotted with mean of n=16 wells. C. ViQi’s Automated High Content Screening assay (AutoHCS) (AutoHCS) used to phenotypically profile varying concentrations of NITD-008 at 64 hpi based on phenotypic similarity. Low concentrations used to phenotypically profile varying concentrations of NITD-008 at 64 hpi based on phenotypic similarity. Low concentrations of NITD-008 cluster of NITD-008 cluster with positive controls of virus without treatment (red), demonstrating ineffectiveness at lower doses. Notably, the highest with positive controls of virus without treatment (red), demonstrating ineffectiveness at lower doses. Notably, the highest doses of NITD-008 (blue) doses of NITD-008 (blue) clustered distinctly from negative controls (no virus or treatment), which clustered with intermediate doses of NITD- clustered distinctly from negative controls (no virus or treatment), which clustered with intermediate doses of NITD-008 (green). This demonstrates an off-target effects cluster, likely capturing the cellular stress and toxicity known to be induced at these higher doses by NITD-008. In this way, AutoHCS 008 (green). This demonstrates an off-target effects cluster, likely capturing the cellular stress and toxicity known to be induced at these higher filters antiviral hits from AVIA to an optimized list of target compounds. doses by NITD-008. In this way, AutoHCS filters antiviral hits from AVIA to an optimized list of target compounds.
(A) Demonstration of the full breadth of AutoHCS analysis. Dendrogram represents all 1280 compounds at 6µM in triplicate at 64hr. Inset shows phenotypes than healthy negative controls (green). Usually, such phenotypic distance suggests drastic and oftentimes unwanted cellular impact like stress and toxicity. We identify clusters based on controls. Green contains the uninfected control and red contains the infected control with no treatment. We identify blue and green as the clusters of interest. B. UF Scripps performed Cell titer glo (CTG) analysis at 72 hpi, indicating 20 how two clusters are represented as even further distinct from all other phenotypes than healthy negative controls (green). Usually, such “hits” where compounds show reduced infectivity. Data presented as response normalized to negative (0%) and positive (100%) controls, with hits defined as any compound with average response above 3x standard phenotypic distance suggests drastic and oftentimes unwanted cellular impact like stress and toxicity. We identify clusters based on controls. deviation greater than the average response of all samples tested (blue dashed line). C. AutoHCS phenotypic profiling analysis run on the blue and green clusters from dendrogram in A. Hits from the CTG analysis were classified into blue, yellow, and purple clusters from dendrogram in A. Because we understand the purple and yellow clusters to be quite phenotypically distant from the uninfected, healthy cells, we focus on the Green contains the uninfected control and red contains the infected control with no treatment. We identify blue and green as the clusters of blue cluster. This cluster contains 12 CTG “hit” compounds. Within the subset clustering, there is one cluster (light blue) where most of the CTG “hits” fall that is phenotypically the most similar to the “healthy” cells interest. (B) UF Scripps performed Cell titer glo (CTG) analysis at 72 hpi, indicating 20 “hits” where compounds show reduced infectivity. Data cluster (green). Interestingly, this cluster contains 9 of the original 20 CTG “hits”. These are the 9 compounds recommended to move forward with testing for antivirals. Finally, compounds within the light blue cluster that were not identified by CTG are recommended for further investigation. D. CTG vs AVIA responses for four compounds that categorized as cluster yellow (dendrogram A). They are hypothesized as ATP mimicking presented as response normalized to negative (0%) and positive (100%) controls, with hits defined as any compound with average response above 3x standard deviation greater than the average response of all samples tested (blue dashed line). (C) AutoHCS phenotypic profiling analysis run on the blue and green clusters from dendrogram in A. Hits from the CTG analysis were classified into blue, yellow, and purple clusters from dendrogram in A. Because we understand the purple and yellow clusters to be quite phenotypically distant from the uninfected, healthy cells, we focus on the blue cluster. This cluster contains 12 CTG “hit” compounds. Within the subset clustering, there is one cluster (light blue) where most of the CTG “hits” fall that is phenotypically the most similar to the “healthy” cells cluster (green). Interestingly, this cluster contains 9 of the original 20 CTG “hits”. These are the 9 compounds recommended to move forward with testing for antivirals. Finally, compounds within the light blue cluster that were not identified by CTG are recommended for further investigation. (D) CTG vs AVIA responses for four compounds that categorized as cluster yellow (dendrogram A). They are hypothesized as ATP mimicking compounds.

Benefits for Drug Discovery Teams

Speed Without Sacrificing Quality: The combination of the Araceli Endeavor imager and ViQi’s predictive AI- based tools (AVIA and AutoHCS) enables generating large amounts of high quality data quickly then analyzing the data with more sensitivity and less effort.

Unbiased, Scalable AI Analysis: The confluence of imaging and analysis technologies is essential to this antiviral screen’s success. Ultra-high throughput imaging at submicron resolution allows live cell screening at scale without sacrificing cell health or data quality, while effective machine-learning tools are essential to extract value and meaning from these images, delivering more effective candidates.

Cloud-Based Infrastructure for Efficiency and Collaboration: Managing, analyzing, and storing large volumes of data is only possible by leveraging cloud computing, which also allows for effective collaboration regardless of where the contributors are located.

More Insightful Hit Selection: Between rapid detection and quantification of viral infection and meaningful clustering of antiviral compounds based on phenotype, it is now possible to quickly and efficiently identify the best compounds to meet your set requirements such as effective antivirals that do not produce off-target effects or cell toxicity.

Conclusions

This study highlights the power of pairing the Araceli Endeavor® high-speed live-cell imaging with ViQi’s AI- based analysis to create a highly scalable, precise, and biologically rich antiviral screening platform. The approach minimizes experimental bias, reduces costs, and enables researchers to rapidly detect viral infectivity and filter out compounds with off-target effects—long before traditional endpoint assays would reveal toxicity. The cloud-based, machine learning-driven analysis performed by AVIA and AutoHCS enables not only accurate infectivity detection but also deep phenotypic profiling, unlocking insights into both desired and undesired compound effects. This end- to-end, high-throughput workflow enables faster go/no-go decisions and helps prioritize the most promising drug candidates earlier in the discovery pipeline.

Using a live cell, brightfield-based screen offers multiple advantages: more subtle effects over time can be elucidated, with lessened reagent costs and more effective outputs not reliant on a single measurement or biased by staining. ViQi’s analyses do not depend on cell death or exaggerated phenotypes – they can detect subtle changes in morphology. By employing both convolutional neural nets (CNNs) and feature classifiers, ViQi harnesses the benefits of multiple types of AIs. Ultra fast, high-content imaging not only allows productivity to be scaled up in terms of number of wells and timepoints, but also preserves cell health during imaging. Maximizing well coverage with submicron resolution further allows machine learning algorithms to extract deeper insight while minimizing variability. AVIA has been applied across a broad range of viruses, and AutoHCS has been used to identify off-target effects in several compound screens. This combination of the Araceli Endeavor’s fast, high-quality imaging with ViQis AI-based image analysis offers a novel approach where phenotypic profiling can be combined with endpoint assays to rapidly identify the most relevant compounds in a screen.

All experimental procedures were planned and executed by the HTS Center at UF Scripps located in Jupiter, FL.

References

¹De Clercq E, and Li G. Approved antiviral drugs over the past 50 years. Clin Microbiol Rev. 2016 29(3):695-747. Doi: 10.1128/CMR.00102-15. PMID: 27281742.

²Baer A, and Kehn-Hall K. Viral concentration determination through plaque assays: using traditional and novel overlay systems. J Vis Exp. 2014 (93):e52065. Doi: 10.3791/52065. PMID: 25407402; PMCID: PMC4255882.

³Dodkins R, Delaney JR, Overton T, Scholle F, Frias-De-Diego A, Crisci E, Huq N, Jordan I, Kimata JT, Findley T, and Goldberg I. A rapid, high-throughput, viral infectivity assay using automated brightfield microscopy with machine learning. SLAS Tech. 2023 28(5): 324-333. Doi: https://doi.org/10.1016/j.slast.2023.07.003. ISSN: 2472-6303.

⁴Berisha B, Mëziu E, and Shabani I. Big data analytics in Cloud computing: an overview. J Cloud Comp. 2022 11(24). Doi: https://doi.org/10.1186/s13677-022-00301-w.

⁵Results at: https://science.viqiai.cloud/web/view/00-kUqcwzK9FVPwogC2FaB3uE

⁶Promega kit G7570: www.promega.com/products/cell-health-assays/cell-viability- and-cytotoxicity-assays/celltiter_glo-luminescent-cell-viability-assay/

Contact Us

To learn more, please contact Matthew Boisvert at Matthew.Boisvert@aracelibio.com, Reese Findley at reese@viqiai.com, Dr. Ilya Goldberg at ilya@viqiai.com, Dr. Yuka Otsuka at otsuka.yu@ufl.edu, or Dr. Timothy Spicer at spicert@urfl.edu.