Abstract B035: Deep learning-based spatially resolved immune clustering in the tumor microenvironment predicts distant metastasis risk in high-grade prostate cancer
Bibliographic record
Abstract
Abstract Objectives: Contemporary computational pathology biomarker approaches in prostate cancer (PCa) generally either analyze the entire digitized image (without adequate human interpretability) or focus solely on tumor regions, in essence refining Gleason grading. We aimed to identify human-interpretable histologic features inclusive of the tumor microenvironment (TME) with prognostic value in localized PCa. Methods: We retrospectively identified two independent cohorts of patients with localized PCa who underwent radical prostatectomy (with longitudinal outcomes data) and had digitized H&E-stained slides from surgical specimens. Cohort A was the discovery cohort; Cohort B served as validation. A deep learning model (CellViT) with a vision transformer encoder pretrained on histopathology images was used to segment and classify immune cell nuclei, and immune cell clusters were identified using DBSCAN. Cox regression was used to examine association between clinicopathologic/histologic features and time to distant metastasis (DM). A third cohort from The Cancer Genome Atlas (TCGA) was used to evaluate relationships between genomic/transcriptomic features (from whole-exome and bulk RNA sequencing) and digital pathology-derived immune metrics in radical prostatectomy specimens. Results: Cohort A (n=272) had median age of 63; 87% was Gleason 6-7 and 93% pT2-T3a. Median immune cell proportion was 4.3% (interquartile range [IQR] 3.0-5.9%), with median of 8.7 immune clusters per 25 mm2 (IQR 0–26.5). Cohort B (n=218) had median age of 62; 83% was Gleason 6-7 and 86% pT2-T3a. Median immune cell proportion was 2.7% (IQR 1.9-3.5%), with median of 4.7 clusters per 25 mm2 (IQR 0-12.1). In Cohort A (median follow-up 12.7 years), log-transformed immune cluster (but not immune cell proportion) was independently associated with DM for Gleason 8-10 (adjusted hazard ratio [AHR] 0.42, 95% confidence interval [CI] 0.19-0.93) but not Gleason 6-7 (AHR 1.26, 95% CI 0.78-2.05), with a significant interaction (Pint=0.019). Similarly in Cohort B (median follow-up 8.1 years), immune cluster (but not immune cell proportion) was associated with DM for Gleason 8-10 (AHR 0.60, 95% CI 0.37-0.98) but not Gleason 6-7 (AHR 1.19, 95% CI 0.74-1.91; Pint=0.043). In TCGA (n=329), high immune cluster (top 10th percentile) was not associated with mutational differences. For Gleason 8-10 samples (but not Gleason 6-7), immune cell deconvolution with CIBERSORTx revealed enrichment of CD8+ T cells (P=0.023), activated memory CD4+ T cells (P=0.014), and Tregs (p=0.004). Immune repertoire profiling with TRUST4 demonstrated increased TRB clonality (P=0.027) indicative of clonally expanded T cell populations in high-cluster samples for Gleason 8-10 (but not Gleason 6-7). Conclusions: We identified and validated spatial immune clustering in the TME as a novel, human-interpretable computational pathology biomarker prognostic of distant metastasis in high-grade PCa. Our findings underscore the potential of biologically informed artificial intelligence approaches for biomarker discovery in PCa. Citation Format: David D. Yang, Alexander J. Haas, Aya Abdelnaser, Eddy Saad, Alfred A. Barney, Jett P. Crowdis, Cora A. Ricker, Jihye Park, Mary-Ellen Taplin, Paul L. Nguyen, Martin T. King, Keyan Salari, Chin-Lee Wu, Eliezer M. Van Allen. Deep learning-based spatially resolved immune clustering in the tumor microenvironment predicts distant metastasis risk in high-grade prostate cancer [abstract]. In: Proceedings of the AACR Special Conference in Cancer Research: Artificial Intelligence and Machine Learning; 2025 Jul 10-12; Montreal, QC, Canada. Philadelphia (PA): AACR; Clin Cancer Res 2025;31(13_Suppl):Abstract nr B035.
Fetched live from OpenAlex and de-inverted. Abstracts are not stored in this database: the inverted indexes are 8.6 GB of the frame’s 9.3 GB of text, and the host has 13 GB free.
How this classification was reachedexpand
Full frame machine prediction
Teacher imitationNot calibrated prevalence, not ground truth. Human validation pending. The Gemma side is a direct model label for every work in the frame, read from the title-only record. The Codex side is a classifier learned from the 10,348 direct Codex labels and calibrated to design-weighted sample rates; fields without enough sample support carry no Codex call. Candidate is the union of the two sides; consensus is their intersection. These outputs are machine_predicted_unvalidated and are not human labels.
Distilled classifier scores by category (both heads)
| Category | Codex | Gemma |
|---|---|---|
| Metaresearch | 0.001 | 0.001 |
| Meta-epidemiology (narrow) | 0.000 | 0.000 |
| Meta-epidemiology (broad) | 0.000 | 0.000 |
| Bibliometrics | 0.000 | 0.000 |
| Science and technology studies | 0.000 | 0.000 |
| Scholarly communication | 0.001 | 0.000 |
| Open science | 0.001 | 0.001 |
| Research integrity | 0.000 | 0.000 |
| Insufficient payload (model declined to judge) | 0.003 | 0.000 |
Machine scores (provisional)
The two teacher heads of the student model, read on this work. A score orders the frame for review; it never asserts a category, and the validation status ships verbatim with every row.
Baseline scores from an immature model (maturity gate not passed, 7 training rounds). Scores rank; they never assert a category.
score_only:v0-immature-baseline · verbatim from the scoring run: score_only means the number may rank works, and no category label ships from itClassification
machine, unvalidatedMachine predicted; a candidate call from one source (direct Gemma or distilled Codex), not a consensus.
How this classification was reached, model by model and score by score, is at the end of the page under "How this classification was reached".