Developing Quantitative AI-histologic Signatures Associated with Prognosis and Treatment Benefit Across Cancer Types in the PLCO Trial

Principal Investigator

Name
Haochen Zhang

Degrees
Ph.D.

Institution
Valar Labs

Position Title
Clinical Data Science Lead

Email
haochen@valarlabs.com

About this CDAS Project

Study
PLCO (Learn more about this study)

Project ID
PLCOI-2054

Initial CDAS Request Approval
Jul 14, 2026

Title
Developing Quantitative AI-histologic Signatures Associated with Prognosis and Treatment Benefit Across Cancer Types in the PLCO Trial

Summary
Across solid tumors, a central challenge in oncology is distinguishing indolent disease that can be managed conservatively from aggressive disease that warrants definitive or intensified treatment. This challenge recurs in each of the cancers studied in the PLCO trial. In localized prostate cancer, clinicians must choose among active surveillance, definitive local therapy, and the addition of androgen deprivation therapy (ADT) to radiotherapy, where the survival benefit of ADT is modest in many risk groups and concentrated in a subset of patients (RTOG 9408: 10-year overall survival 62% vs. 57%; RTOG 8610), while ADT carries clinically important cardiovascular, metabolic, sexual, and skeletal toxicity. In lung and colorectal cancer, decisions about adjuvant chemotherapy after resection hinge on the residual risk of recurrence, which current staging estimates imperfectly. In ovarian cancer, prognosis and likelihood of benefit from systemic therapy vary widely among patients with similar clinical features. In each setting, biomarkers that identify poor-prognosis patients and those most likely to benefit from treatment intensification could improve how treatment is matched to disease biology.
Valar Labs is a start-up founded by Stanford researchers at the intersection of artificial intelligence (AI) and medicine. The company constructs clinically relevant biomarker signatures from morphologic features extracted from histologic slides, using AI to inform treatment decisions. Rather than taking a primarily deep learning approach, the Valar Labs Computational Histology AI (CHAI) platform involves AI-powered segmentation of cell nuclei on a diagnostic H&E slide and then AI-powered extraction of hundreds of quantitative features describing qualities of the identified tumor and tumor microenvironment (such as geometric descriptors of tumor cell nucleus size and shape, and spatial descriptors of immune cells in regard to tumor). These features are associated with outcomes of interest, with subsequent machine learning used to train a signature capable of serving as a biomarker. Because this pipeline operates on standard H&E slides and tumor-agnostic morphologic features, it generalizes naturally across cancer types. Our initial work has focused on identifying such biomarkers in pancreatic cancer (PMID: 37044094) and bladder cancer (PMID: 39383345). The bladder work is the basis for a CLIA-approved test predicting benefit from BCG in non-muscle-invasive bladder cancer that is now used in clinical practice.
We propose to use the CHAI platform to develop and validate AI-histologic signatures associated with outcomes across the cancer types represented in PLCO: prostate, lung, colorectal, and ovarian. For each cancer type, we will construct a signature that stratifies survival outcomes among affected patients and, where treatment data support it, an exploratory signature that identifies patients most likely to benefit from a specific treatment intensification (such as the addition of ADT to radiotherapy in prostate cancer, or adjuvant chemotherapy in lung and colorectal cancer). This work would draw on PLCO digitized diagnostic H&E slides, where available, with corresponding treatment, survival, demographic, and baseline screening and clinical data. PLCO data would be combined with other sources to develop and validate clinically implementable AI-histologic biomarkers of prognosis and treatment benefit for the benefit of patients.

Aims

Primary Aim: Develop and validate AI-derived histologic signatures associated with prognosis across the PLCO cancer types (prostate, lung, colorectal, and ovarian).
For each cancer type, we will apply the CHAI platform to digitized diagnostic H&E slides from PLCO to derive a quantitative histologic signature that stratifies patients by risk of adverse outcome. Following AI-powered nuclear segmentation and extraction of hundreds of morphologic and spatial features describing the tumor and its microenvironment, we will train a multivariable signature against long-term outcomes, using cancer-specific mortality as the primary endpoint and overall survival as a secondary endpoint. Models will be developed with cross-validation within PLCO and assessed by discrimination (time-dependent concordance), calibration, and risk-group separation, adjusting for established prognostic variables for each disease (such as grade, stage, and relevant clinical biomarkers). We will then confirm that each signature carries prognostic information that is independent of, and additive to, current clinicopathologic risk stratification.

Exploratory Aim: Develop AI-derived histologic signatures predictive of benefit from treatment intensification, where treatment and outcome data permit.
Within each cancer type, where treatment data support it, we will test whether a histologic signature identifies the subset of patients who derive the greatest benefit from a specific treatment intensification, for example the addition of ADT to radiotherapy in prostate cancer or adjuvant chemotherapy after resection in lung and colorectal cancer. Because treatment in PLCO reflects real-world practice rather than randomization, this aim is exploratory: we will model the interaction between the signature and treatment receipt on survival outcomes, account for measured confounders, and interpret a positive interaction as evidence of a predictive (treatment-selection) rather than purely prognostic biomarker. To strengthen inference, results will be combined with and validated against external cohorts and randomized datasets in which treatment was assigned by protocol. Signatures that flag the patients most likely to benefit could help target intensified therapy to those who gain while sparing others its associated toxicity.

Collaborators

Haochen Zhang Valar Labs
Anirudh Joshi Valar Labs
Trevor Royce Valar Labs
Viswesh Krishna Valar Labs
Gaurav Kaul Valar Labs