Pathology

Current deep-learning H&E biomarker predictors are strongly confounded and not ready to replace molecular testing

This analysis shows that deep-learning models predicting molecular biomarkers from histology mainly learn correlations with co-occurring clinicopathologic features, not isolated biomarker effects. Prediction accuracy for a given biomarker varies markedly with the status of codependent biomarkers and clinicopathological variables. For several biomarkers, gains over what a pathologist can infer from routine features such as grade are modest. The authors state these tools should not substitute for molecular testing but may be cautiously used for triage or complementary decision-making. 1

ctDNA FGFR testing complements tissue for selecting erdafitinib in metastatic urothelial carcinoma

In this prospective multicenter study of 208 metastatic urothelial carcinoma patients, plasma ctDNA FGFR testing showed high concordance with standard tissue testing. Among 125 patients with detectable baseline ctDNA and paired tissue, FGFR status was concordant in 90%, with ctDNA sensitivity of 84% for tissue-detected alterations. ctDNA testing also identified seven additional actionable FGFR alterations not seen on tissue, supporting added yield in routine practice. Serial plasma sampling clarified FGFR status over time, and 21 patients treated with erdafitinib after testing had median progression-free survival of 7.5 months. 2

ICC genetic hierarchy robustly stratifies secondary AML and MDS/AML prognosis

This cohort of 924 MDS/AML and AML patients applied the ICC genetic hierarchy to dissect secondary AML biology and outcomes. Cases with TP53 mutation or myelodysplasia-related cytogenetic abnormality showed similar biology and prognosis, regardless of blast percentage or MDS/AML versus AML labeling. In contrast, MDS/AML with myelodysplasia-related gene mutations or NOS had distinct mutation profiles and better survival than AML, including longer median overall survival. Within each ICC-defined AML category, prior MDS history versus de novo onset did not affect prognosis, emphasizing genetics over clinical history. 3

Spatial transcriptomics distinguishes radionecrosis from true glioblastoma progression after chemoradiotherapy

This study used Xenium spatial single-cell transcriptomics on ten specimens from nine IDH-wildtype glioblastoma patients reoperated for radiologic progression. Histology classified cases into pure recurrence, pure radionecrosis, or mixed lesions, and the atlas captured complex spatial architecture in each group. Progressive glioblastoma samples contained OPC/NPC-like and proliferating tumor cells with high EGFR expression. Radionecrotic samples still harbored tumor cells but showed downregulated EGFR despite amplification and lacked proliferation markers. 4