Single-cell gene regulatory network prediction by explainable AI.

The molecular heterogeneity of cancer cells contributes to the often partial response to targeted therapies and relapse of disease due to the escape of resistant cell populations. While single-cell sequencing has started to improve our understanding of this heterogeneity, it offers a mostly descriptive view on cellular types and states. To obtain more functional insights, we propose scGeneRAI, an explainable deep learning approach that uses layer-wise relevance propagation (LRP) to infer gene regulatory networks from static single-cell RNA sequencing data for individual cells. We benchmark our method with synthetic data and apply it to single-cell RNA sequencing data of a cohort of human lung cancers. From the predicted single-cell networks our approach reveals characteristic network patterns for tumor cells and normal epithelial cells and identifies subnetworks that are observed only in (subgroups of) tumor cells of certain patients. While current state-of-the-art methods are limited by their ability to only predict average networks for cell populations, our approach facilitates the reconstruction of networks down to the level of single cells which can be utilized to characterize the heterogeneity of gene regulation within and across tumors.

Molecular characteristics and improved survival prediction in a cohort of 2023 ependymomas.

The diagnosis of ependymoma has moved from a purely histopathological review with limited prognostic value to an integrated diagnosis, relying heavily on molecular information. However, as the integrated approach is still novel and some molecular ependymoma subtypes are quite rare, few studies have correlated integrated pathology and clinical outcome, often focusing on small series of single molecular types. We collected data from 2023 ependymomas as classified by DNA methylation profiling, consisting of 1736 previously published and 287 unpublished methylation profiles. Methylation data and clinical information were correlated, and an integrated model was developed to predict progression-free survival. Patients with EPN-PFA, EPN-ZFTA, and EPN-MYCN tumors showed the worst outcome with 10-year overall survival rates of 56%, 62%, and 32%, respectively. EPN-PFA harbored chromosome 1q gains and/or 6q losses as markers for worse survival. In supratentorial EPN-ZFTA, a combined loss of CDKN2A and B indicated worse survival, whereas a single loss did not. Twelve out of 200 EPN-ZFTA (6%) were located in the posterior fossa, and these tumors relapsed or progressed even earlier than supratentorial tumors with a combined loss of CDKN2A/B. Patients with MPE and PF-SE, generally regarded as non-aggressive tumors, only had a 10-year progression-free survival of 59% and 65%, respectively. For the prediction of the 5-year progression-free survival, Kaplan-Meier estimators based on the molecular subtype, a Support Vector Machine based on methylation, and an integrated model based on clinical factors, CNV data, and predicted methylation scores achieved balanced accuracies of 66%, 68%, and 73%, respectively. Excluding samples with low prediction scores resulted in balanced accuracies of over 80%. In sum, our large-scale analysis of ependymomas provides robust information about molecular features and their clinical meaning. Our data are particularly relevant for rare and hardly explored tumor subtypes and seemingly benign variants that display higher recurrence rates than previously believed.

Transcriptomic and epigenetic dissection of spinal ependymoma (SP-EPN) identifies clinically relevant subtypes enriched for tumors with and without NF2 mutation.

Ependymomas encompass multiple clinically relevant tumor types based on localization and molecular profiles. Tumors of the methylation class "spinal ependymoma" (SP-EPN) represent the most common intramedullary neoplasms in children and adults. However, their developmental origin is ill-defined, molecular data are scarce, and the potential heterogeneity within SP-EPN remains unexplored. The only known recurrent genetic events in SP-EPN are loss of chromosome 22q and NF2 mutations, but neither types and frequency of these alterations nor their clinical relevance have been described in a large, epigenetically defined series. Transcriptomic (n = 72), epigenetic (n = 225), genetic (n = 134), and clinical data (n = 112) were integrated for a detailed molecular overview on SP-EPN. Additionally, we mapped SP-EPN transcriptomes to developmental atlases of the developing and adult spinal cord to uncover potential developmental origins of these tumors. The integration of transcriptomic ependymoma data with single-cell atlases of the spinal cord revealed that SP-EPN display the highest similarities to mature adult ependymal cells. Unsupervised hierarchical clustering of transcriptomic data together with integrated analysis of methylation profiles identified two molecular SP-EPN subtypes. Subtype A tumors primarily carried previously known germline or sporadic NF2 mutations together with 22q loss (bi-allelic NF2 loss), resulting in decreased NF2 expression. Furthermore, they more often presented as multilocular disease and demonstrated a significantly reduced progression-free survival as compared to SP-EP subtype B. In contrast, subtype B predominantly contained samples without NF2 mutation detected in sequencing together with 22q loss (monoallelic NF2 loss). These tumors showed regular NF2 expression but more extensive global copy number alterations. Based on integrated molecular profiling of a large multi-center cohort, we identified two distinct SP-EPN subtypes with important implications for genetic counseling, patient surveillance, and drug development priorities.

Artificial intelligence and pathology: From principles to practice and future applications in histomorphology and molecular profiling.

The complexity of diagnostic (surgical) pathology has increased substantially over the last decades with respect to histomorphological and molecular profiling. Pathology has steadily expanded its role in tumor diagnostics and beyond from disease entity identification via prognosis estimation to precision therapy prediction. It is therefore not surprising that pathology is among the disciplines in medicine with high expectations in the application of artificial intelligence (AI) or machine learning approaches given their capabilities to analyze complex data in a quantitative and standardized manner to further enhance scope and precision of diagnostics. While an obvious application is the analysis of histological images, recent applications for the analysis of molecular profiling data from different sources and clinical data support the notion that AI will enhance both histopathology and molecular pathology in the future. At the same time, current literature should not be misunderstood in a way that pathologists will likely be replaced by AI applications in the foreseeable future. Although AI will transform pathology in the coming years, recent studies reporting AI algorithms to diagnose cancer or predict certain molecular properties deal with relatively simple diagnostic problems that fall short of the diagnostic complexity pathologists face in clinical routine. Here, we review the pertinent literature of AI methods and their applications to pathology, and put the current achievements and what can be expected in the future in the context of the requirements for research and routine diagnostics.

Molecular refinement of pilocytic astrocytoma in adult patients.

AIM: Pilocytic astrocytomas (PA) in adults are rare and may be challenging to identify based only on histomorphology. Compared to their paediatric counterparts, they are reportedly molecularly more diverse and associated with a worse prognosis. We aimed to describe the characteristics of adult PAs more precisely by comprehensively profiling a series of 79 histologically diagnosed adult cases (≥18 years). METHODS: We performed global DNA methylation profiling and DNA and RNA panel sequencing, and integrated the results with clinical data. We further compared the molecular characteristics of adult and paediatric PAs that had a significant match to one of the established PA methylation classes in the Heidelberg brain tumour classifier. RESULTS: The mean age in our cohort was 33 years, and 43% of the tumours were located supratentorially. Based on methylation profiling, only 39% of the cases received a significant match to a PA methylation class. Sixteen per cent matched a different tumour type and 45% had a Heidelberg classifier score <0.9 with an affiliation to diverse established methylation classes in t-SNE analyses. Although the KIAA1549::BRAF fusion was found in 98% of paediatric PAs, this was true for only 27% of histologically defined and 55% of adult PAs defined by methylation profiling. CONCLUSIONS: A particularly high fraction of adult tumours with histological features of PA do not match current PA methylation classes, indicating ambiguous histology and an urgent need for molecular profiling. Moreover, even in adult PAs with a match to a PA methylation class, the distribution of genetic drivers differs significantly from their paediatric counterparts (p<0.01).

Recurrent atypical teratoid/rhabdoid tumors (AT/RT) reveal discrete features of progression on histology, epigenetics, copy number profiling, and transcriptomics.

Atypical teratoid/rhabdoid tumors (AT/RT) are the most common malignant brain tumors manifesting in infancy. They split into four molecular types. The major three (AT/RT-SHH, AT/RT-TYR, and AT/RT-MYC) all carry mutations in SMARCB1, the fourth quantitatively smaller type is characterized by SMARCA4 mutations (AT/RT-SMARCA4). Molecular characteristics of disease recurrence or metastatic spread, which go along with a particularly dismal outcome, are currently unclear. Here, we investigated tumor tissue from 26 patients affected by AT/RT to identify signatures of recurrences in comparison with matched primary tumor samples. Microscopically, AT/RT recurrences demonstrated a loss of architecture and significantly enhanced mitotic activity as compared to their related primary tumors. Based on DNA methylation profiling, primary tumor and related recurrence were grossly similar, but three out of 26 tumors belonged to a different molecular type or subtype after second surgery compared to related primary lesions. Copy number variations (CNVs) differed in six cases, showing novel gains on chromosome 1q or losses of chromosome 10 in recurrences as the most frequent alterations. To consolidate these observations, our cohort was combined with a data set of unmatched primary and recurrent AT/RT, which demonstrated chromosome 1q gain and 10 loss in 18% (n = 7) and 11% (n = 4) of the recurrences (n = 38) as compared to 7% (n = 3) and 0% (n = 0) in the primary tumors (n = 44), respectively. Similar to the observations made by DNA methylation profiling, RNA sequencing of our cohort revealed AT/RT primary tumors and matched recurrences clustering closely together. However, a number of genes showed significantly altered expression in AT/RT-SHH recurrences. Many of them are known tumor driving growth factors, involved in embryonal development and tumorigenesis, or are cell-cycle-associated. Overall, our work identifies subtle molecular changes that occur in the course of the disease and that may help define novel therapeutic targets for AT/RT recurrences.

Identification of low and very high-risk patients with non-WNT/non-SHH medulloblastoma by improved clinico-molecular stratification of the HIT2000 and I-HIT-MED cohorts.

Molecular groups of medulloblastoma (MB) are well established. Novel risk stratification parameters include Group 3/4 (non-WNT/non-SHH) methylation subgroups I-VIII or whole-chromosomal aberration (WCA) phenotypes. This study investigates the integration of clinical and molecular parameters to improve risk stratification of non-WNT/non-SHH MB. Non-WNT/non-SHH MB from the HIT2000 study and the HIT-MED registries were selected based on availability of DNA-methylation profiling data. MYC or MYCN amplification and WCA of chromosomes 7, 8, and 11 were inferred from methylation array-based copy number profiles. In total, 403 non-WNT/non-SHH MB were identified, 346/403 (86%) had a methylation class family Group 3/4 methylation score (classifier v11b6) ≥ 0.9, and 294/346 (73%) were included in the risk stratification modeling based on Group 3 or 4 score (v11b6) ≥ 0.8 and subgroup I-VIII score (mb_g34) ≥ 0.8. Group 3 MB (5y-PFS, survival estimation ± standard deviation: 41.4 ± 4.6%; 5y-OS: 48.8 ± 5.0%) showed poorer survival compared to Group 4 (5y-PFS: 68.2 ± 3.7%; 5y-OS: 84.8 ± 2.8%). Subgroups II (5y-PFS: 27.6 ± 8.2%) and III (5y-PFS: 37.5 ± 7.9%) showed the poorest and subgroup VI (5y-PFS: 76.6 ± 7.9%), VII (5y-PFS: 75.9 ± 7.2%), and VIII (5y-PFS: 66.6 ± 5.8%) the best survival. Multivariate analysis revealed subgroup in combination with WCA phenotype to best predict risk of progression and death. The integration of clinical (age, M and R status) and molecular (MYC/N, subgroup, WCA phenotype) variables identified a low-risk stratum with a 5y-PFS of 94 ± 5.7 and a very high-risk stratum with a 5y-PFS of 29 ± 6.1%. Validation in an international MB cohort confirmed the combined stratification scheme with 82.1 ± 6.0% 5y-PFS in the low and 47.5 ± 4.1% in very high-risk groups, and outperformed the clinical model. These newly identified clinico-molecular low-risk and very high-risk strata, accounting for 6%, and 21% of non-WNT/non-SHH MB patients, respectively, may improve future treatment stratification.

Machine learning models predict the primary sites of head and neck squamous cell carcinoma metastases based on DNA methylation.

In head and neck squamous cell cancers (HNSCs) that present as metastases with an unknown primary (HNSC-CUPs), the identification of a primary tumor improves therapy options and increases patient survival. However, the currently available diagnostic methods are laborious and do not offer a sufficient detection rate. Predictive machine learning models based on DNA methylation profiles have recently emerged as a promising technique for tumor classification. We applied this technique to HNSC to develop a tool that can improve the diagnostic work-up for HNSC-CUPs. On a reference cohort of 405 primary HNSC samples, we developed four classifiers based on different machine learning models [random forest (RF), neural network (NN), elastic net penalized logistic regression (LOGREG), and support vector machine (SVM)] that predict the primary site of HNSC tumors from their DNA methylation profile. The classifiers achieved high classification accuracies (RF = 83%, NN = 88%, LOGREG = SVM = 89%) on an independent cohort of 64 HNSC metastases. Further, the NN, LOGREG, and SVM models significantly outperformed p16 status as a marker for an origin in the oropharynx. In conclusion, the DNA methylation profiles of HNSC metastases are characteristic for their primary sites, and the classifiers developed in this study, which are made available to the scientific community, can provide valuable information to guide the diagnostic work-up of HNSC-CUP. © 2021 The Authors. The Journal of Pathology published by John Wiley & Sons Ltd on behalf of The Pathological Society of Great Britain and Ireland.

Clinical and molecular characterization of isolated M1 disease in pediatric medulloblastoma: experience from the German HIT-MED studies.

PURPOSE: To evaluate the clinical impact of isolated spread of medulloblastoma cells into cerebrospinal fluid without additional macroscopic metastases (M1-only). METHODS: The HIT-MED database was searched for pediatric patients with M1-only medulloblastoma diagnosed from 2000 to 2019. Corresponding clinical and molecular data was evaluated. Treatment was stratified by age and changed over time for older patients. RESULTS: 70 patients with centrally reviewed M1-only disease were identified. Clinical data was available for all and molecular data for 45/70 cases. 91% were non-WNT/non-SHH medulloblastoma (Grp3/4). 5-year PFS for 52 patients ≥ 4 years was 59.4 (± 7.1) %, receiving either upfront craniospinal irradiation (CSI) or SKK-sandwich chemotherapy (CT). Outcomes did not differ between these strategies (5-year PFS: CSI 61.7 ± 9.9%, SKK-CT 56.7 ± 6.1%). For patients < 4 years (n = 18), 5-year PFS was 50.0 (± 13.2) %. M1-persistence occurred exclusively using postoperative CT and was a strong negative predictive factor (pPFS/OS < 0.01). Patients with additional clinical or molecular high-risk (HR) characteristics had worse outcomes (5-year PFS 42.7 ± 10.6% vs. 64.0 ± 7.0%, p = 0.03). In n = 22 patients ≥ 4 years with full molecular information and without additional HR characteristics, risk classification by molecular subtyping had an effect on 5-year PFS (HR 16.7 ± 15.2%, SR 77.8 ± 13.9%; p = 0.01). CONCLUSIONS: Our results confirm that M1-only is a high-risk condition, and further underline the importance of CSF staging. Specific risk stratification of affected patients needs attention in future discussions for trials and treatment recommendations. Future patients without contraindications may benefit from upfront CSI by sparing risks related to higher cumulative CT applied in sandwich regimen.

Cauda equina paragangliomas express HOXB13.

HOXB13 is expressed in the tail bud of the developing embryo as well as in cauda equina paragangliomas and in myxopapillary ependymomas. In contrast, pheochromocytomas and paraganglioma in other locations as well as many other tumors occuring in spinal cord regions are negative.