Assessing the clinical and biological associations between multimodal artificial intelligence (MMAI) and 22-gene genomic classifier (GC) in localized prostate cancer (PCa).

Abstract

196 Background: MMAI and GC are clinically validated prognostic tools for localized PCa, and are incorporated into NCCN guidelines as treatment-decision aids. However, their clinical and biological relationships remain unclear, especially in Asian populations where comparative study is limited. We investigated the association between these genomic and AI pathology biomarkers and their respective accuracies for the prognostication of metastasis risk in patients with NCCN intermediate- (Int) to high-risk PCa. Methods: Patients with newly-diagnosed localized PCa from a single institution in Singapore were enrolled into a prospective protocol (NCT04340024), and underwent image-guided radiotherapy with or without hormonal therapy. All subjects had paired GC and MMAI scores. However, colored marker annotations were present on the H&E slides, which could negatively impact MMAI performance. Associations between GC and MMAI with metastasis-free survival (MFS) were evaluated using Cox proportional hazards models and and Area under the Receiver Operating Characteristic Curve (AUC). Pearson correlation was used to assess associations of cancer hallmark pathways. Results: Of 144 included men (142 NCCN Int/High; median follow-up 78.5 mo), GC and MMAI scores were correlated ( R = 0.55) but differed in risk stratification; GC classified 50/142 (35%) NCCN Int/High cases as GC low-, 31/142 (22%) as int-, and 61/142 (43%) as high-risk, whereas MMAI assigned 11/142 (8%), 53/142 (37%), and 78/142 (55%), respectively. Both tests were prognostic for MFS (MMAI: hazard ratio [ HR ] of 2.6 [95% CI:1.0–6.4], P =0.035; GC: HR of 3.4 [95% CI:1.5–7.6], P =0.002), although the difference in HR should be interpreted with caution, as it is attributable to a limited sample size with longer follow-up. Adding them to NCCN improved 5y MFS AUCs from 0.71 to 0.80 for both NCCN+MMAI and NCCN+GC. Biologically, MMAI-high and GC-high groups showed comparable proportions of aggressive subtypes (PAM50 Basal 51% vs 58%, P =0.419; Luminal B 46% vs 39%, P =0.377; PTEN loss 22% vs 25%; P =0.698; AR lower activity 49% vs 44%, P =0.512), though pathway enrichment diverged; GC high versus GC low tumors showed positive enrichment in E2F, oxidative phosphorylation, and mTORC1 pathways; negative enrichment in P53 and estrogen pathways, whereas MMAI high was uniquely negatively enriched for fatty acid metabolism. Conclusions: Both MMAI and GC are biomarkers that enhance NCCN risk stratification, though they likely capture distinct biological signals. Incorporating these tests with NCCN guidelines may better inform treatment escalation decisions—such as the addition of hormonal therapy for NCCN-Int or anti-androgens for NCCN-High risk disease—in localized PCa.