HRDPath is a research model that looks for clues to impaired DNA repair in ovarian-cancer tissue images. It predicts a molecular characteristic of a tumour, rather than screening a symptom-free person for ovarian cancer or directly deciding which treatment that person should receive.
The work is interesting because it asks whether routinely prepared tissue images contain information usually sought through molecular testing. Its performance is also a useful example of why one impressive accuracy headline can hide several different tests of a model.
The question hidden inside the tissue
The King’s College London account explains the target: homologous recombination deficiency, or HRD, in high-grade serous tubo-ovarian cancer. HRD concerns a problem in a cell’s DNA-repair machinery and can be relevant to understanding treatment responses.
HRDPath analyses digitised tissue slides rather than directly measuring DNA. The researchers trained it using cases with molecular reference information, then examined whether patterns in the images could predict the HRD classification.
The proposed benefit is a more accessible source of information where tissue images are already available. That is a research rationale, not evidence that the model has replaced molecular testing, reduced costs or improved treatment outcomes in routine care.
A slide is not a patient
The Nature Communications study uses a patient-level approach, combining information from slides belonging to the same person. Keeping all of a person’s slides within one data split helps prevent a model being tested on familiar patient-specific patterns.
This distinction matters when reading the sample size. Hundreds of images are not hundreds of independent patients. Multiple slides can add information about one tumour while also creating opportunities for misleadingly large image counts if the unit of analysis is not stated.
The principal development dataset and the public validation datasets also differ in their molecular reference labels and composition. Calling a dataset “external” is useful information, but it does not mean every evaluation applied an unchanged model with an unchanged threshold to a representative new clinical population.
Why the performance numbers change
In the reported held-out benchmark of 29 patients, the model achieved an area under the receiver-operating-characteristic curve of 0.846. At the reported classification threshold, sensitivity was 0.692 and specificity 0.938.
Those quantities are different. The area under the curve describes discrimination across thresholds; it does not mean 84.6% of patients were correctly classified. Sensitivity and specificity describe different kinds of classification at a particular threshold.
The paper also explores combining more slides, consistency between predictions and adapting the model to other datasets. Those analyses can produce more favourable results, but they answer different questions from the original held-out benchmark.
For example, evaluating people with several available slides selects a different information setting from evaluating a person represented by one slide. Fine-tuning on examples from a new dataset is also different from applying a fully frozen model without that adaptation. Neither should be silently presented as a universal performance figure.
What a stronger clinical test would need to show
The external PTRC dataset contained many HRD-positive cases but only five HRD-negative patients. That imbalance limits how much one can learn about errors in the smaller group. The researchers themselves identify a need for broader validation.
A prospective clinical evaluation would also have a different task from this retrospective modelling study. It would need to establish how the output is used, which patients or specimens produce uncertain results, how it fits alongside other evidence and whether acting on it improves an outcome that matters.
That is an evaluation question, rather than a claim that this paper has already answered it. A model can identify meaningful associations in tissue without yet establishing that a treatment decision based on its predictions is beneficial.
HRDPath therefore offers a supported research story about information in histology and the way a model is tested. Keeping the patient unit, reference standard, slide count and adaptation beside its performance is essential to understanding how far that story currently reaches.