Teaching a computer to read the gold standard

The gold standard for diagnosing herpes is not a machine readout. It is a set of paper strips pasted onto a white sheet, with as many as 24 patient samples to a page. Each strip is read by eye by trained technicians, whose interpretations are then reconciled under supervisory review. The diagnostic test has worked this way in Seattle for four decades, and it works well.

Each strip carries a ladder of dark bands, a fingerprint of which viral proteins a patient's antibodies recognize. Every patient gets two strips, one for HSV-1 and one for HSV-2. Correctly interpreting those patterns takes years of training.

The assay, known as the HSV Western blot, has a long local history. "The HSV Western blot represents four decades of contributions from the UW and Fred Hutch," said Dr. Pavitra Roychoudhury, corresponding author on a new study bringing artificial intelligence into the workflow. She credits the pioneering development effort led by Dr. Rhoda Ashley Morrow (University of Washington) and Dr. Larry Corey (Fred Hutch), along with the laboratory scientists at UW Virology who still run the assay today. "The assay has impacted potentially hundreds of thousands of patients by providing a diagnosis to someone who may have been exposed to HSV, and driven so many clinical trials and research studies on HSV."

It is also the yardstick against which other tests are measured. Federal guidance points to the Western blot as the comparator when new HSV blood tests are developed and validated. Those results matter well beyond the lab, shaping treatment decisions, conversations with partners, and family planning.

The trouble is that a gold standard read by hand is costly, and in genuinely ambiguous cases it can be inconsistent. Reporting in the Journal of Virological Methods, a team from Fred Hutch and UW Medicine asked whether modern computer vision could help.

"Interpreting these blots requires reading a complex pattern of bands and requires years of training to read," Roychoudhury said. "The idea is not to replace the human reader but to aid in interpretation, so that (if successful), the number of required readers could be reduced from 3 to maybe 1."

Their tool, BlotDx, works in two steps, mirroring how a technician approaches interpreting the assay. Because blots are photographed one sheet at a time, BlotDx first has to find each patient's strips within the larger image. One set of models scans the photograph and draws a box around each patient's pair of strips, or traces the outline of each strip individually. A second set of models then looks at the cropped strips and determines whether each patient is positive or negative for HSV-1 and HSV-2.

Rather than building the system from scratch, the team started with computer-vision models already trained on more than a million ordinary photographs of dogs, chairs, and street scenes. They then retrained those models to recognize the patterns found in Western blots, a technique called transfer learning.

"We adapted models that were pretrained on large natural-image datasets to specific Western blot images to achieve promising classification performance," said Dr. Lucas Liu, the study's first author. From a clinical standpoint, he added, the work "demonstrates the potential of using AI for HSV Western blot analysis to lower the manual review effort and increase scalability."

The team trained BlotDx on 926 blot pairs assayed in 2016 and 2017 and photographed in 2018. Compared with the consensus interpretation of expert human readers, BlotDx called serostatus correctly 98.8% of the time for HSV-1 and 98.9% for HSV-2.

The harder test came next. The researchers pulled a second set of 185 blot pairs from samples run between 2019 and 2024 and photographed them in December 2025. These images came from newer blots run in a different laboratory, with different lighting and camera setups. The detection stage missed a few strips in the new images and had to be retrained with added variation before it found them all. Once it did, BlotDx called 97.3% of HSV-1 results and 96.2% of HSV-2 results correctly, reflecting a genuine decline for HSV-2 performance rather than noise.

The team also examined the models using saliency maps, which highlight the regions of an image that most influenced a decision. For HSV-2 calls, the models leaned heavily on the part of the strip where two of the virus's surface proteins appear, glycoprotein B and glycoprotein G. Glycoprotein G is the one that differs most between HSV-1 and HSV-2, which is why laboratory specialists treat a strong gG band as the key requirement for an HSV-2 positive result. Dr. Youyi Fong, who co-led the work explained, "The paper demonstrated a solution to an engineering problem: how to take a collection of Western blot images and train a deep learning model to effectively encode and leverage institutional knowledge."

The study was deliberately scoped as a pilot and important challenges remain. Samples with indeterminate results were excluded. These borderline cases that need repeat testing or extra workup are precisely the hardest for human readers, so real-world accuracy would likely be lower. And while those images were captured in a different laboratory, the blots themselves were produced by the same lab running the same test, so BlotDx has not yet been validated on samples processed somewhere else entirely.

The next steps include broader external validation and prospective testing inside a live clinical workflow. Dr. Liu also wants to know what the model is really seeing. Interpretability work so far, he said, offers "a first glance at which parts of the Western blots impact the model's predictions," but the biology behind those features is not yet understood.

The bigger opportunity may lie in the archive. UW Virology holds more than 60,000 saved blots from the past decade. "We want to mine these blots for information like what are the most informative bands or set of bands that really drive the classification result," Roychoudhury said. If researchers can identify the smallest set of viral proteins needed to reliably distinguish positive and negative results, those proteins could become candidates for a cheaper, faster next-generation test.

Roychoudhury sees the project as symbolic of the collaboration across the Fred Hutch/University of Washington/Seattle Children's Cancer Consortium. "Neither the original assay nor BlotDx would have happened without the contributions of individuals across Fred Hutch and the UW," she said.

This study did not receive any funding.

Fred Hutch/University of Washington/Seattle Children’s Cancer Consortium Members Dr. Keith Jerome contributed to this research.

Liu, L., Sathees, S., Sobel, A., Pepper, G., Greninger, A. L., Jerome, K. R., Fong, Y., & Roychoudhury, P. (2026). BlotDx: A deep learning tool for Western blot-based diagnostics. Journal of virological methods, 343, 115385.