HIP: a method for high-dimensional multi-view data integration and prediction accounting for subgroup heterogeneity.

Epidemiologic and genetic studies in many complex diseases suggest subgroup disparities (e.g. by sex, race) in disease course and patient outcomes. We consider this from the standpoint of integrative analysis where we combine information from different views (e.g. genomics, proteomics, clinical data). Existing integrative analysis methods ignore the heterogeneity in subgroups, and stacking the views and accounting for subgroup heterogeneity does not model the association among the views. We propose Heterogeneity in Integration and Prediction (HIP), a statistical approach for joint association and prediction that leverages the strengths in each view to identify molecular signatures that are shared by and specific to a subgroup. We apply HIP to proteomics and gene expression data pertaining to chronic obstructive pulmonary disease (COPD) to identify proteins and genes shared by, and unique to, males and females, contributing to the variation in COPD, measured by airway wall thickness. Our COPD findings have identified proteins, genes, and pathways that are common across and specific to males and females, some implicated in COPD, while others could lead to new insights into sex differences in COPD mechanisms. HIP accounts for subgroup heterogeneity in multi-view data, ranks variables based on importance, is applicable to univariate or multivariate continuous outcomes, and incorporates covariate adjustment. With the efficient algorithms implemented using PyTorch, this method has many potential scientific applications and could enhance multiomics research in health disparities. HIP is available at https://github.com/lasandrall/HIP, a video tutorial at https://youtu.be/O6E2OLmeMDo and a Shiny Application at https://multi-viewlearn.shinyapps.io/HIP_ShinyApp/ for users with limited programming experience.

ZSWIM4 regulates embryonic patterning and BMP signaling by promoting nuclear Smad1 degradation.

The dorsoventral gradient of BMP signaling plays an essential role in embryonic patterning. Zinc Finger SWIM-Type Containing 4 (zswim4) is expressed in the Spemann-Mangold organizer at the onset of Xenopus gastrulation and is then enriched in the developing neuroectoderm at the mid-gastrula stages. Knockdown or knockout of zswim4 causes ventralization. Overexpression of zswim4 decreases, whereas knockdown of zswim4 increases the expression levels of ventrolateral mesoderm marker genes. Mechanistically, ZSWIM4 attenuates the BMP signal by reducing the protein stability of SMAD1 in the nucleus. Stable isotope labeling by amino acids in cell culture (SILAC) identifies Elongin B (ELOB) and Elongin C (ELOC) as the interaction partners of ZSWIM4. Accordingly, ZSWIM4 forms a complex with the Cul2-RING ubiquitin ligase and ELOB and ELOC, promoting the ubiquitination and degradation of SMAD1 in the nucleus. Our study identifies a novel mechanism that restricts BMP signaling in the nucleus.

Robust knowledge-guided biclustering for multi-omics data.

Biclustering is a useful method for simultaneously grouping samples and features and has been applied across various biomedical data types. However, most existing biclustering methods lack the ability to integratively analyze multi-modal data such as multi-omics data such as genome, transcriptome and epigenome. Moreover, the potential of leveraging biological knowledge represented by graphs, which has been demonstrated to be beneficial in various statistical tasks such as variable selection and prediction, remains largely untapped in the context of biclustering. To address both, we propose a novel Bayesian biclustering method called Bayesian graph-guided biclustering (BGB). Specifically, we introduce a new hierarchical sparsity-inducing prior to effectively incorporate biological graph information and establish a unified framework to model multi-view data. We develop an efficient Markov chain Monte Carlo algorithm to conduct posterior sampling and inference. Extensive simulations and real data analysis show that BGB outperforms other popular biclustering methods. Notably, BGB is robust in terms of utilizing biological knowledge and has the capability to reveal biologically meaningful information from heterogeneous multi-modal data.

Integrative analysis of multi-omics and imaging data with incorporation of biological information via structural Bayesian factor analysis.

MOTIVATION: With the rapid development of modern technologies, massive data are available for the systematic study of Alzheimer's disease (AD). Though many existing AD studies mainly focus on single-modality omics data, multi-omics datasets can provide a more comprehensive understanding of AD. To bridge this gap, we proposed a novel structural Bayesian factor analysis framework (SBFA) to extract the information shared by multi-omics data through the aggregation of genotyping data, gene expression data, neuroimaging phenotypes and prior biological network knowledge. Our approach can extract common information shared by different modalities and encourage biologically related features to be selected, guiding future AD research in a biologically meaningful way. METHOD: Our SBFA model decomposes the mean parameters of the data into a sparse factor loading matrix and a factor matrix, where the factor matrix represents the common information extracted from multi-omics and imaging data. Our framework is designed to incorporate prior biological network information. Our simulation study demonstrated that our proposed SBFA framework could achieve the best performance compared with the other state-of-the-art factor-analysis-based integrative analysis methods. RESULTS: We apply our proposed SBFA model together with several state-of-the-art factor analysis models to extract the latent common information from genotyping, gene expression and brain imaging data simultaneously from the ADNI biobank database. The latent information is then used to predict the functional activities questionnaire score, an important measurement for diagnosis of AD quantifying subjects' abilities in daily life. Our SBFA model shows the best prediction performance compared with the other factor analysis models. AVAILABILITY: Code are publicly available at https://github.com/JingxuanBao/SBFA. CONTACT: qlong@upenn.edu.

Mutational signature and prognosis in adenocarcinoma of the bladder.

Adenocarcinoma of the bladder is a rare urinary bladder carcinoma with limited therapy options due to lack of molecular characterization. Here, we aimed to reveal the mutational and transcriptomic landscapes of adenocarcinoma of the bladder and assess any relationship with prognosis. Between February 2015 and June 2021, a total of 23 patients with adenocarcinoma of the bladder were enrolled. These included 16 patients with primary bladder adenocarcinomas and seven patients with urachal adenocarcinoma. Whole exome sequencing (16 patients), whole genome sequencing (16 patients), bulk RNA sequencing (RNA-seq) (19 patients), and single-cell RNA-seq (5 patients) were conducted for the specimens. Correlation analysis, survival analysis, and t-tests were also performed. Prevalent T>A substitutions were observed among somatic mutations, and major trinucleotide contexts included 5'-CTC-3' and 5'-CTG-3'. This pattern was mainly contributed by COSMIC signature 22 related to chemical carcinogen exposure (probably aristolochic acid), which has not been reported in bladder adenocarcinoma. Moreover, genes with copy number changes were also enriched in the KEGG term 'chemical carcinogenesis'. Transcriptomic analysis suggested high immune cell infiltration and luminal-like features in the majority of samples. Interestingly, a small fraction of samples with an APOBEC-derived mutational signature exhibited a higher risk of disease progression compared with samples with only a chemical carcinogen-related signature, confirming the molecular and prognostic heterogeneity of bladder adenocarcinoma. This study presents mutational and transcriptomic landscapes of bladder adenocarcinoma, and indicates that a chemical carcinogen-related mutational signature may be related to a better prognosis compared with an APOBEC signature in adenocarcinoma of the bladder. © 2024 The Pathological Society of Great Britain and Ireland.

Heterochromatin loosening by the Oct4 linker region facilitates Klf4 binding and iPSC reprogramming.

The success of Yamanaka factor reprogramming of somatic cells into induced pluripotent stem cells suggests that some factor(s) must remodel the nuclei from a condensed state to a relaxed state. How factor-dependent chromatin opening occurs remains unclear. Using FRAP and ATAC-seq, we found that Oct4 acts as a pioneer factor that loosens heterochromatin and facilitates the binding of Klf4 and the expression of epithelial genes in early reprogramming, leading to enhanced mesenchymal-to-epithelial transition. A mutation in the Oct4 linker, L80A, which shows impaired interaction with the BAF complex component Brg1, is inactive in heterochromatin loosening. Oct4-L80A also blocks the binding of Klf4 and retards MET. Finally, vitamin C or Gadd45a could rescue the reprogramming deficiency of Oct4-L80A by enhancing chromatin opening and Klf4 binding. These studies reveal a cooperation between Oct4 and Klf4 at the chromatin level that facilitates MET at the cellular level and shed light into the research of multiple factors in cell fate determination.

Macrophage-Derived Cathepsin S Remodels the Extracellular Matrix to Promote Liver Fibrogenesis.

BACKGROUND & AIMS: Liver fibrosis is an intrinsic wound-healing response to chronic injury and the major cause of liver-related morbidity and mortality worldwide. However, no effective diagnostic or therapeutic strategies are available, owing to its poorly characterized molecular etiology. We aimed to elucidate the mechanisms underlying liver fibrogenesis. METHODS: We performed a quantitative proteomic analysis of clinical fibrotic liver samples to identify dysregulated proteins. Further analyses were performed on the sera of 164 patients with liver fibrosis. Two fibrosis mouse models and several biochemical experiments were used to elucidate liver fibrogenesis. RESULTS: We identified cathepsin S (CTSS) up-regulation as a central node for extracellular matrix remodeling in the human fibrotic liver by proteomic screening. Increased serum CTSS levels efficiently predicted liver fibrosis, even at an early stage. Secreted CTSS cleaved collagen 18A1 at its C-terminus, releasing endostatin peptide, which directly bound to and activated hepatic stellate cells via integrin α5ß1 signaling, whereas genetic ablation of Ctss remarkably suppressed liver fibrogenesis via endostatin reduction in vivo. Further studies identified macrophages as the main source of hepatic CTSS, and splenectomy effectively attenuated macrophage infiltration and CTSS expression in the fibrotic liver. Pharmacologic inhibition of CTSS ameliorated liver fibrosis progression in the mouse models. CONCLUSIONS: CTSS functions as a novel profibrotic factor by remodeling extracellular matrix proteins and may represent a promising target for the diagnosis and treatment of liver fibrosis.

Kindlin2 regulates neural crest specification via integrin-independent regulation of the FGF signaling pathway.

The focal adhesion protein Kindlin2 is essential for integrin activation, a process that is fundamental to cell-extracellular matrix adhesion. Kindlin 2 (Fermt2) is widely expressed in mouse embryos, and its absence causes lethality at the peri-implantation stage due to the failure to trigger integrin activation. The function of kindlin2 during embryogenesis has not yet been fully elucidated as a result of this early embryonic lethality. Here, we showed that kindlin2 is essential for neural crest (NC) formation in Xenopus embryos. Loss-of-function assays performed with kindlin2-specific morpholino antisense oligos (MOs) or with CRISPR/Cas9 techniques in Xenopus embryos severely inhibit the specification of the NC. Moreover, integrin-binding-deficient mutants of Kindlin2 rescued the phenotype caused by loss of kindlin2, suggesting that the function of kindlin2 during NC specification is independent of integrins. Mechanistically, we found that Kindlin2 regulates the fibroblast growth factor (FGF) pathway, and promotes the stability of FGF receptor 1. Our study reveals a novel function of Kindlin2 in regulating the FGF signaling pathway and provides mechanistic insights into the function of Kindlin2 during NC specification.

An integrative latent class model of heterogeneous data modalities for diagnosing kidney obstruction.

Radionuclide imaging plays a critical role in the diagnosis and management of kidney obstruction. However, most practicing radiologists in US hospitals have insufficient time and resources to acquire training and experience needed to interpret radionuclide images, leading to increased diagnostic errors. To tackle this problem, Emory University embarked on a study that aims to develop a computer-assisted diagnostic (CAD) tool for kidney obstruction by mining and analyzing patient data comprised of renogram curves, ordinal expert ratings on the obstruction status, pharmacokinetic variables, and demographic information. The major challenges here are the heterogeneity in data modes and the lack of gold standard for determining kidney obstruction. In this article, we develop a statistically principled CAD tool based on an integrative latent class model that leverages heterogeneous data modalities available for each patient to provide accurate prediction of kidney obstruction. Our integrative model consists of three sub-models (multilevel functional latent factor regression model, probit scalar-on-function regression model, and Gaussian mixture model), each of which is tailored to the specific data mode and depends on the unknown obstruction status (latent class). An efficient MCMC algorithm is developed to train the model and predict kidney obstruction with associated uncertainty. Extensive simulations are conducted to evaluate the performance of the proposed method. An application to an Emory renal study demonstrates the usefulness of our model as a CAD tool for kidney obstruction.

Integrating multi-omics summary data using a Mendelian randomization framework.

Mendelian randomization is a versatile tool to identify the possible causal relationship between an omics biomarker and disease outcome using genetic variants as instrumental variables. A key theme is the prioritization of genes whose omics readouts can be used as predictors of the disease outcome through analyzing GWAS and QTL summary data. However, there is a dearth of study of the best practice in probing the effects of multiple -omics biomarkers annotated to the same gene of interest. To bridge this gap, we propose powerful combination tests that integrate multiple correlated $P$-values without assuming the dependence structure between the exposures. Our extensive simulation experiments demonstrate the superiority of our proposed approach compared with existing methods that are adapted to the setting of our interest. The top hits of the analyses of multi-omics Alzheimer's disease datasets include genes ABCA7 and ATP1B1.

Impact of Skeletal Muscle Loss and Sarcopenia on Outcomes of Locally Advanced Esophageal Cancer during Neoadjuvant Chemoradiation.

BACKGROUND: The impact of changes in skeletal muscle and sarcopenia on outcomes during neoadjuvant chemoradiotherapy (NACR) for patients with esophageal cancer remains controversial. PATIENTS AND METHODS: We retrospectively analyzed the data of patients with locally advanced esophageal squamous cell cancer who received NACR followed by esophagectomy between June 2013 and December 2021. The images at third lumbar vertebra were analyzed to measure the cross-sectional area and calculate skeletal muscle index (SMI) before and after NACR. SMI less than 52.4 cm2/m2 for men and less than 38.5 cm2/m2 for women were defined as sarcopenia. The nonlinearity of the effect of percent changes in SMI (ΔSMI%) to survival outcomes was assessed by restricted cubic splines. RESULTS: Overall, data of 367 patients were analyzed. The survival outcomes between sarcopenia and non-sarcopenia groups had no significant differences before NACR. However, patients in post-NACR sarcopenia group showed poor overall survival (OS) benefit (P = 0.016) and poor disease-free survival (DFS) (P = 0.043). Severe postoperative complication rates were 11.9% in post-NACR sarcopenia group and 5.0% in post-NACR non-sarcopenia group (P = 0.019). There was a significant non-linear relationship between ΔSMI% and survival outcomes (P < 0.05 for non-linear). On the multivariable analysis of OS, ΔSMI% > 12% was the independent prognostic factor (HR 1.76, 95% CI 1.03-2.99, P = 0.039) and significant difference was also found on DFS analysis (P = 0.025). CONCLUSIONS: Patients with post-neoadjuvant chemoradiotherapy sarcopenia have worse survival and adverse short-term outcomes. Moreover, greater loss in SMI is associated with increased risks of death and disease progression during neoadjuvant chemoradiotherapy, with maximum impact noted with SMI loss greater than 12%.

Increased resected lymph node stations improved survival of esophageal squamous cell carcinoma.

BACKGROUND: Neoadjuvant chemoradiotherapy (nCRT) and surgery have been recommended as the standard treatments for locally advanced esophageal squamous cell carcinoma (ESCC). In addition, nodal metastases decreased in frequency and changed in distribution after neoadjuvant therapy. This study aimed to examine the optimal strategy for lymph node dissection (LND) in patients with ESCC who underwent nCRT. METHODS: The hazard ratios (HRs) for overall survival (OS) and disease-free survival (DFS) were calculated using the Cox proportional hazard model. To determine the minimal number of LNDs (n-LNS) or least station of LNDs (e-LNS), the Chow test was used. RESULTS: In total, 333 patients were included. The estimated cut-off values for e-LNS and n-LNS were 9 and 15, respectively. A higher number of e-LNS was significantly associated with improved OS (HR: 0.90; 95% CI 0.84-0.97, P = 0.0075) and DFS (HR: 0.012; 95% CI: 0.84-0.98, P = 0.0074). The e-LNS was a significant prognostic factor in multivariate analyses. The local recurrence rate of 23.1% in high e-LNS is much lower than the results of low e-LNS (13.3%). Comparable morbidity was found in both the e-LNS and n-LND subgroups. CONCLUSION: This cohort study revealed an association between the extent of LND and overall survival, suggesting the therapeutic value of extended lymphadenectomy during esophagectomy. Therefore, more lymph node stations being sampled leads to higher survival rates among patients who receive nCRT, and standard lymphadenectomy of at least 9 stations is strongly recommended.

Accounting for network noise in graph-guided Bayesian modeling of structured high-dimensional data.

There is a growing body of literature on knowledge-guided statistical learning methods for analysis of structured high-dimensional data (such as genomic and transcriptomic data) that can incorporate knowledge of underlying networks derived from functional genomics and functional proteomics. These methods have been shown to improve variable selection and prediction accuracy and yield more interpretable results. However, these methods typically use graphs extracted from existing databases or rely on subject matter expertise, which are known to be incomplete and may contain false edges. To address this gap, we propose a graph-guided Bayesian modeling framework to account for network noise in regression models involving structured high-dimensional predictors. Specifically, we use 2 sources of network information, including the noisy graph extracted from existing databases and the estimated graph from observed predictors in the dataset at hand, to inform the model for the true underlying network via a latent scale modeling framework. This model is coupled with the Bayesian regression model with structured high-dimensional predictors involving an adaptive structured shrinkage prior. We develop an efficient Markov chain Monte Carlo algorithm for posterior sampling. We demonstrate the advantages of our method over existing methods in simulations, and through analyses of a genomics dataset and another proteomics dataset for Alzheimer's disease.

Incorporating graph information in Bayesian factor analysis with robust and adaptive shrinkage priors.

There has been an increasing interest in decomposing high-dimensional multi-omics data into a product of low-rank and sparse matrices for the purpose of dimension reduction and feature engineering. Bayesian factor models achieve such low-dimensional representation of the original data through different sparsity-inducing priors. However, few of these models can efficiently incorporate the information encoded by the biological graphs, which has been already proven to be useful in many analysis tasks. In this work, we propose a Bayesian factor model with novel hierarchical priors, which incorporate the biological graph knowledge as a tool of identifying a group of genes functioning collaboratively. The proposed model therefore enables sparsity within networks by allowing each factor loading to be shrunk adaptively and by considering additional layers to relate individual shrinkage parameters to the underlying graph information, both of which yield a more accurate structure recovery of factor loadings. Further, this new priors overcome the phase transition phenomenon, in contrast to existing graph-incorporated approaches, so that it is robust to noisy edges that are inconsistent with the actual sparsity structure of the factor loadings. Finally, our model can handle both continuous and discrete data types. The proposed method is shown to outperform several existing factor analysis methods through simulation experiments and real data analyses.

Impact of deep muscle invasion on nodal status and survival in patients with pT2 esophageal squamous cancer.

BACKGROUND: Whether T2 esophageal squamous cell carcinoma should be subclassified remains controversial. We aimed to investigate the impact of the depth of muscularis propria invasion on nodal status and survival outcomes. METHODS: We identified patients with pT2 esophageal squamous cell carcinoma who underwent primary surgery from January 2009 to June 2017. Clinical data were extracted from prospectively maintained databases. Tumor muscularis propria invasion was stratified into superficial or deep. Binary logistic regression was used to determine risk factors for lymph node metastases. The impact of the depth of muscularis propria invasion on survival was investigated using Kaplan‒Meier analysis and a Cox proportional hazard regression model. RESULTS: A total of 750 patients from three institutes were investigated. The depth of muscularis propria invasion (odds ratio [OR]: 3.95, 95% confidence interval [CI]: 2.46-6.35; p < 0.001) was correlated with lymph node metastases using logistic regression. T substage (hazard ratio [HR]: 1.37, 95% CI: 1.05-1.79; p < 0.001) and N status (HR: 1.51, 95% CI: 1.05-2.17; p < 0.001) were independent risk factors in multivariate Cox regression analysis. The deep muscle invasion was associated with worse overall survival (HR: 1.52, 95% CI: 1.19-1.94; p = 0.001) than superficial, specifically in T2N0 patients (HR: 1.38, 95% CI: 1.08-1.94; p = 0.035). CONCLUSIONS: We found that deep muscle invasion was associated with significantly worse outcomes and recommended the substaging of pT2 esophageal squamous cell carcinoma in routine pathological examination.

Preference matrix guided sparse canonical correlation analysis for mining brain imaging genetic associations in Alzheimer's disease.

Investigating the relationship between genetic variation and phenotypic traits is a key issue in quantitative genetics. Specifically for Alzheimer's disease, the association between genetic markers and quantitative traits remains vague while, once identified, will provide valuable guidance for the study and development of genetics-based treatment approaches. Currently, to analyze the association of two modalities, sparse canonical correlation analysis (SCCA) is commonly used to compute one sparse linear combination of the variable features for each modality, giving a pair of linear combination vectors in total that maximizes the cross-correlation between the analyzed modalities. One drawback of the plain SCCA model is that the existing findings and knowledge cannot be integrated into the model as priors to help extract interesting correlations as well as identify biologically meaningful genetic and phenotypic markers. To bridge this gap, we introduce preference matrix guided SCCA (PM-SCCA) that not only takes priors encoded as a preference matrix but also maintains computational simplicity. A simulation study and a real-data experiment are conducted to investigate the effectiveness of the model. Both experiments demonstrate that the proposed PM-SCCA model can capture not only genotype-phenotype correlation but also relevant features effectively.

Preparation and epitope analysis of monoclonal antibodies against African swine fever virus DP96R protein.

BACKGROUND: Many proteins of African swine fever virus (ASFV, such as p72, p54, p30, CD2v, K205R) have been successfully expressed and characterized. However, there are few reports on the DP96R protein of ASFV, which is the virulence protein of ASFV and plays an important role in the process of host infection and invasion of ASFV. RESULTS: Firstly, the prokaryotic expression vector of DP96R gene was constructed, the prokaryotic system was used to induce the expression of DP96R protein, and monoclonal antibody was prepared by immunizing mice. Four monoclonal cells of DP96R protein were obtained by three ELISA screening and two sub-cloning; the titer of ascites antibody was up to 1:500,000, and the monoclonal antibody could specifically recognize DP96R protein. Finally, the subtypes of the four strains of monoclonal antibodies were identified and the minimum epitopes recognized by them were determined. CONCLUSION: Monoclonal antibody against ASFV DP96R protein was successfully prepared and identified, which lays a foundation for further exploration of the structure and function of DP96R protein and ASFV diagnostic technology.

Mechanistic study of transcription factor Sox18 during heart development.

Heart development is a delicate and complex process regulated by coordination of various signaling pathways. In this study, we investigated the role of sox18 in heart development by modulating Wnt/ß-Catenin signaling pathways. Our spatiotemporal expression analysis revealed that sox18 is mainly expressed in the heart, branchial arch, pharyngeal arch, spinal cord, and intersegmental vessels at the tailbud stage of Xenopus tropicalis embryo. Overexpression of sox18 in the X. tropicalis embryos causes heart edema, while loss-of-function of sox18 can change the signal of developmental heart marker gata4 at different stages, suggesting that sox18 plays an essential role in the development of the heart. Knockdown of SOX18 in human umbilical vein endothelial cells suggests a link between Sox18 and ß-CATENIN, a key regulator of the Wnt signaling pathway. Sox18 negatively regulates islet1 and tbx3, the downstream factors of Wnt/ß-Catenin signaling, during the linear heart tube formation and the heart looping stage. Taken together, our findings highlight the crucial role of Sox18 in the development of the heart via inhibiting Wnt/ß-Catenin signaling.

Mapping of lymph node metastasis from esophageal squamous cell carcinoma after neoadjuvant treatment: a prospective analysis from a high-volume institution in China.

The study aimed to describe the prevalence of lymph node metastases per lymph node station for esophageal squamous cell carcinoma (ESCC) after neoadjuvant treatment. Clinicopathological variables of ESCC patients were retrieved from the prospective database of the Surgical Esophageal Cancer Patient Registry in West China Hospital, Sichuan University. A two-field lymphadenectomy was routinely performed, and an extensive three-field lymphadenectomy was performed if cervical lymph node metastasis was suspected. According to AJCC/UICC 8, lymph node stations were investigated separately. The number of patients with metastatic lymph nodes divided by those who underwent lymph node dissection at that station was used to define the percentage of patients with lymph node metastases. Data are also separately analyzed according to the pathological response of the primary tumor, neoadjuvant treatment regimens, pretreatment tumor length, and tumor location. Between January 2019 and March 2023, 623 patients who underwent neoadjuvant therapy followed by transthoracic esophagectomy were enrolled. Lymph node metastases were found in 212 patients (34.0%) and most frequently seen in lymph nodes along the right recurrent nerve (10.1%, 58/575), paracardial station (11.4%, 67/587), and lymph nodes along the left gastric artery (10.9%, 65/597). For patients with pretreatment tumor length of >4 cm and non-pathological complete response of the primary tumor, the metastatic rate of the right lower cervical paratracheal lymph nodes is 10.9% (10/92) and 10.6% (11/104), respectively. For patients with an upper thoracic tumor, metastatic lymph nodes were most frequently seen along the right recurrent nerve (14.2%, 8/56). For patients with a middle thoracic tumor, metastatic lymph nodes were most commonly seen in the right lower cervical paratracheal lymph nodes (10.3%, 8/78), paracardial lymph nodes (10.2%, 29/285), and lymph nodes along the left gastric artery (10.4%, 30/289). For patients with a lower thoracic tumor, metastatic lymph nodes were most frequently seen in the paracardial station (14.2%, 35/247) and lymph nodes along the left gastric artery (13.1%, 33/252). The study precisely determined the distribution of lymph node metastases in ESCC after neoadjuvant treatment, which may help to optimize the extent of lymphadenectomy in the surgical management of ESCC patients after neoadjuvant therapy.

Exploring deep neural networks via layer-peeled model: Minority collapse in imbalanced training.

In this paper, we introduce the Layer-Peeled Model, a nonconvex, yet analytically tractable, optimization program, in a quest to better understand deep neural networks that are trained for a sufficiently long time. As the name suggests, this model is derived by isolating the topmost layer from the remainder of the neural network, followed by imposing certain constraints separately on the two parts of the network. We demonstrate that the Layer-Peeled Model, albeit simple, inherits many characteristics of well-trained neural networks, thereby offering an effective tool for explaining and predicting common empirical patterns of deep-learning training. First, when working on class-balanced datasets, we prove that any solution to this model forms a simplex equiangular tight frame, which, in part, explains the recently discovered phenomenon of neural collapse [V. Papyan, X. Y. Han, D. L. Donoho, Proc. Natl. Acad. Sci. U.S.A. 117, 24652-24663 (2020)]. More importantly, when moving to the imbalanced case, our analysis of the Layer-Peeled Model reveals a hitherto-unknown phenomenon that we term Minority Collapse, which fundamentally limits the performance of deep-learning models on the minority classes. In addition, we use the Layer-Peeled Model to gain insights into how to mitigate Minority Collapse. Interestingly, this phenomenon is first predicted by the Layer-Peeled Model before being confirmed by our computational experiments.