Virtual Molecular Pathology: Predicting Molecular Alterations and Therapeutic Targets from Routine H&E Whole-Slide Images

Molecular characterization has become indispensable for therapeutic decision-making in oncology, yet access to it remains profoundly unequal. Immunohistochemistry, FISH, and next-generation sequencing are constrained by tissue availability, cost, turnaround time, and infrastructure, and a substantial fraction of patients worldwide — particularly outside tertiary centers and in low- and middle-income settings — never receive complete molecular profiling. At the same time, the hematoxylin and eosin (H&E) stained slide, produced for essentially every solid tumor specimen, encodes morphological correlates of the underlying molecular state that human observers cannot reliably quantify.Deep learning has begun to exploit this signal. Weakly supervised whole-slide models now predict microsatellite instability, homologous recombination deficiency, HER2 status, and selected driver mutations directly from H&E, and a new generation of pathology foundation models pretrained on hundreds of thousands of slides has extended these predictions toward gene expression programs, molecular subtypes, spatial transcriptomic and proteomic maps, and virtual staining of protein targets. This shifts the framing of computational pathology from image classification toward a general morphology-to-molecule inference problem: which molecular readouts are recoverable from routine histology, at what accuracy, and under what conditions can they inform clinical action?Critical gaps remain. Reported performance frequently degrades under external validation, and cross-center or cross-population generalization — including across ancestry groups underrepresented in public cohorts — is inconsistently reported. Supervision signals are heterogeneous in quality, and the emerging use of spatial transcriptomics and high-plex spatial proteomics as dense training labels raises unresolved questions about resolution, registration, and label noise. Interpretability, calibration, prospective validation, triage-versus-replacement positioning, regulatory pathways, and reimbursement are largely unaddressed. Reporting practices are inconsistent enough that cross-study comparison is often impossible.This Research Topic invites original and synthetic work that advances the rigor and clinical translation of molecular prediction from routine histopathology. Areas of interest include, but are not limited to:• Prediction of actionable alterations, fusions, mutational signatures, MSI/dMMR, TMB, and expression-based subtypes from H&E whole-slide images• Pathology foundation models: pretraining strategies, benchmarking, domain adaptation, and parameter-efficient transfer to molecular endpoints• Virtual staining and virtual IHC, including validation against orthogonal assays and inter-observer comparison• Spatial transcriptomics and spatial proteomics as supervision signals for H&E-based inference; multimodal alignment and registration• Multimodal integration of histology with genomics, radiology, and clinical variables for treatment response and prognosis• External, multi-center, and multi-ancestry validation; batch and scanner effects; site-level confounding and leakage• Calibration, uncertainty quantification, and explainability for clinical deployment• Health-economic modeling and prospective or trial-embedded evaluation; regulatory and quality-assurance frameworks• Datasets, benchmarks, and reporting standards for morphology-to-molecule studiesWe welcome Original Research, Brief Research Reports, Methods, Systematic Reviews, Reviews, Clinical Trial reports, Data Reports, and Perspectives. Studies reporting negative or null results — endpoints that prove not to be predictable from morphology — are explicitly encouraged, as are studies from underrepresented populations and resource-constrained settings.Submissions reporting single-cohort internal cross-validation without independent external evaluation, or without transparent description of data partitioning at the patient and site level, will generally be considered out of scope unless the methodological contribution is substantial. Topic editors: XUANYE CAO, Guangsheng Pei Source: https://www.frontiersin.org/research-topics/85198/virtual-molecular-pathology-predicting-molecular-alterations-and-therapeutic-targets-from-routine-he-whole-slide-images