Mark A. van de Wiel’s research while affiliated with Gezond Amsterdam and other places

Cancer prognosis is often based on a set of omics covariates and a set of established clinical covariates such as age and tumor stage. Combining these two sets poses challenges. First, dimension difference: clinical covariates should be favored because they are low-dimensional and usually have stronger prognostic ability than high-dimensional omics covariates. Second, interactions: genetic profiles and their prognostic effects may vary across patient subpopulations. Last, redundancy: a (set of) gene(s) may encode similar prognostic information as a clinical covariate. To address these challenges, we combine regression trees, employing clinical covariates only, with a fusion-like penalized regression framework in the leaf nodes for the omics covariates. The fusion penalty controls the variability in genetic profiles across subpopulations. We prove that the shrinkage limit of the proposed method equals a benchmark model: a ridge regression with penalized omics covariates and unpenalized clinical covariates. Furthermore, the proposed method allows researchers to evaluate, for different subpopulations, whether the overall omics effect enhances prognosis compared to only employing clinical covariates. In an application to colorectal cancer prognosis based on established clinical covariates and 20,000+ gene expressions, we illustrate the features of our method.

Machine learning models often face challenges in medical applications due to covariate shifts, where discrepancies between training and target data distributions can decrease predictive accuracy. This paper introduces an adaptation of Classification and Regression Trees (CART) that incorporates importance weighting to address these distributional differences effectively. By assigning greater weight to training samples that closely represent the target distribution, our approach modifies the CART model to improve performance in the presence of covariate shift. We evaluate the effectiveness of this method through simulation studies and apply it to real-world medical data, showing significant improvements in predictive accuracy. The results indicate that this weighted CART approach can be valuable in medical and other fields where covariate shift poses challenges, enabling more reliable predictions across diverse data distributions.

Cancer is caused by an accumulation of somatic mutations and copy number alterations (CNAs). Besides mutations, these copy number changes are key characteristics of cancer development. Nonetheless, some tumors show hardly any CNAs, a remarkable phenomenon in oncogenesis. Head and neck squamous cell carcinomas (HNSCCs) arise by either exposure to carcinogens, or infection with the human papillomavirus (HPV). HPV-negative HNSCCs are generally characterized by many CNAs and frequent mutations in CDKN2A, TP53, FAT1, and NOTCH1. Here, we present the hallmarks of the distinct subgroup of HPV-negative HNSCC with no or few CNAs (CNA-quiet) by genetic profiling of 802 oral cavity squamous cell carcinomas (OCSCCs). In total, 73 OCSCC (9.1%) are classified as CNA-quiet and 729 as CNA-other. The CNA-quiet group is characterized by wild-type TP53, frequent CASP8 and HRAS mutations, and a less immunosuppressed tumor immune microenvironment with lower density of regulatory T cells. Patients with CNA-quiet OCSCC are older, more often women, less frequently current smokers, and have a better 5-year overall survival compared to CNA-other OCSCC. This study demonstrates that CNA-quiet OCSCC should be considered as a distinct, clinically relevant subclass. Given the clinical characteristics, the patient group with these tumors will rapidly increase in the aging population.

Combining data from various sources empowers researchers to explore innovative questions, for example those raised by conducting healthcare monitoring studies. However, the lack of a unique identifier often poses challenges. Record linkage procedures determine whether pairs of observations collected on different occasions belong to the same individual using partially identifying variables (e.g. birth year, postal code). Existing methodologies typically involve a compromise between computational efficiency and accuracy. Traditional approaches simplify this task by condensing information, yet they neglect dependencies among linkage decisions and disregard the one-to-one relationship required to establish coherent links. Modern approaches offer a comprehensive representation of the data generation process, at the expense of computational overhead and reduced flexibility. We propose a flexible method, that adapts to varying data complexities, addressing registration errors and accommodating changes of the identifying information over time. Our approach balances accuracy and scalability, estimating the linkage using a Stochastic Expectation Maximisation algorithm on a latent variable model. We illustrate the ability of our methodology to connect observations using large real data applications and demonstrate the robustness of our model to the linking variables quality in a simulation study. The proposed algorithm FlexRL is implemented and available in an open source R package.

Objectives The addition of two-way interactions is a classic problem in statistics, and comes with the challenge of quadratically increasing dimension. We aim to a) devise an estimation method that can handle this challenge and b) to aid interpretation of the resulting model by developing computational tools for quantifying variable importance. Methods Existing strategies typically overcome the dimensionality problem by only allowing interactions between relevant main effects. Building on this philosophy, and aiming for settings with moderate n to p ratio, we develop a local shrinkage model that links the shrinkage of interaction effects to the shrinkage of their corresponding main effects. In addition, we derive a new analytical formula for the Shapley value, which allows rapid assessment of individual-specific variable importance scores and their uncertainties. Results We empirically demonstrate that our approach provides accurate estimates of the model parameters and very competitive predictive accuracy. In our Bayesian framework, estimation inherently comes with inference, which facilitates variable selection. Comparisons with key competitors are provided. Large-scale cohort data are used to provide realistic illustrations and evaluations. The implementation of our method in RStan is relatively straightforward and flexible, allowing for adaptation to specific needs. Conclusions Our method is an attractive alternative for existing strategies to handle interactions in epidemiological and/or clinical studies, as its linked local shrinkage can improve parameter accuracy, prediction and variable selection. Moreover, it provides appropriate inference and interpretation, and may compete well with less interpretable machine learners in terms of prediction.

Regression discontinuity design (RDD) is a quasi-experimental approach used to estimate the causal effects of an intervention assigned based on a cutoff criterion. RDD exploits the idea that close to the cutoff units below and above are similar; hence, they can be meaningfully compared. Consequently, the causal effect can be estimated only locally at the cutoff point. This makes the cutoff point an essential element of RDD. However, especially in medical applications, the exact cutoff location may not always be disclosed to the researcher, and even when it is, the actual location may deviate from the official one. As we illustrate on the application of RDD to the HIV treatment eligibility data, estimating the causal effect at an incorrect cutoff point leads to meaningless results. Moreover, since the cutoff criterion often acts as a guideline rather than as a strict rule, the location of the cutoff may be unclear from the data. The method we present can be applied both as an estimation and validation tool in RDD. We use a Bayesian approach to incorporate prior knowledge and uncertainty about the cutoff location in the causal effect estimation. At the same time, our Bayesian model LoTTA is fitted globally to the whole data, whereas RDD is a local, boundary point estimation problem. In this work we address a natural question that arises: how to make Bayesian inference more local to render a meaningful and powerful estimate of the treatment effect?

Background: Overweight and obesity impose a considerable individual and social burden, and the urban environments might encompass factors that contribute to obesity. Nevertheless, there is a scarcity of research that takes into account the simultaneous interaction of multiple environmental factors. Objectives: Our objective was to perform an exposome-wide association study of body mass index (BMI) in a multicohort setting of 15 studies. Methods: Studies were affiliated with the Dutch Geoscience and Health Cohort Consortium (GECCO), had different population sizes (688-141,825), and covered the entire Netherlands. Ten studies contained general population samples, others focused on specific populations including people with diabetes or impaired hearing. BMI was calculated from self-reported or measured height and weight. Associations with 69 residential neighborhood environmental factors (air pollution, noise, temperature, neighborhood socioeconomic and demographic factors, food environment, drivability, and walkability) were explored. Random forest (RF) regression addressed potential nonlinear and nonadditive associations. In the absence of formal methods for multimodel inference for RF, a rank aggregation-based meta-analytic strategy was used to summarize the results across the studies. Results: Six exposures were associated with BMI: five indicating neighborhood economic or social environments (average home values, percentage of high-income residents, average income, livability score, share of single residents) and one indicating the physical activity environment (walkability in 5-km buffer area). Living in high-income neighborhoods and neighborhoods with higher livability scores was associated with lower BMI. Nonlinear associations were observed with neighborhood home values in all studies. Lower neighborhood home values were associated with higher BMI scores but only for values up to €300,000. The directions of associations were less consistent for walkability and share of single residents. Discussion: Rank aggregation made it possible to flexibly combine the results from various studies, although between-study heterogeneity could not be estimated quantitatively based on RF models. Neighborhood social, economic, and physical environments had the strongest associations with BMI. https://doi.org/10.1289/EHP13393.

Deconvolution of bulk RNA profiles can identify hidden cell fractions and functional states in the tumor microenvironment. The malignant cells however, are commonly the most abundant cell type and impair deconvolution by profound interpatient heterogeneity. We developed a malignant cell fraction-informed RNA deconvolution method by Bayesian integration of DNA-derived malignant fraction estimates, OncoBLADE. We evaluated OncoBLADE experimentally using bulk RNA profiles with 19 CyTOF determined cell fractions from blood of 46 individuals, and in silico using 180,177 single-cell RNA profiles from 73 lung tumors. Using malignant cell fractions, OncoBLADE achieved improved accuracy in cell fraction and cell type-specific RNA profile estimation. We tested OncoBLADE on real bulk profiles of 50 adeno- and 50 squamous lung cancer samples, revealing hidden RNA profiles of malignant cells and fibroblasts that distinguished two histological subtypes. In conclusion, multimodal deconvolution utilizing bulk DNA and RNA data accurately unveils hidden RNA profiles within the tumor microenvironment.

To study the evolutionary processes that drive malignant progression of IDH-mutant astrocytomas, we performed multi-omics on a large cohort of matched initial and recurrent tumor samples. The overlay of genetic, epigenetic, transcriptomic and proteomic data, combined with single-cell analysis, have identified overlapping features associated with malignant progression. These features are derived from three molecular mechanisms and provide a rationale of the underlying biology of tumor malignancy: cell-cycling, tumor cell (de-)differentiation and remodeling of the extracellular matrix. Specifically, DNA-methylation levels decreased over time, predominantly in tumors with malignant transformation and co-occurred with poor prognostic genetic events. DNA-methylation was lifted from specific loci associated with DNA replication and was associated with an increased RNA and protein expression of cell cycling associated genes. All results were validated on samples of newly diagnosed IDH-mutant astrocytoma patients included the CATNON randomized phase 3 clinical trial. Importantly, malignant progression was hardly affected by radio- or chemotherapy, indicating that treatment does not affect the course of disease. Our results culminate in a DNA-methylation based signature for objective tumor grading.

To study the evolutionary processes that drive malignant progression of IDH-mutant astrocytomas, we performed multi-omics on a large cohort of matched initial and recurrent tumor samples. The overlay of genetic, epigenetic, transcriptomic and proteomic data, combined with single-cell analysis, have identified overlapping features associated with malignant progression. These features are derived from three molecular mechanisms and provide a rationale of the underlying biology of tumor malignancy: cell-cycling, tumor cell (de-)differentiation and remodeling of the extracellular matrix. Specifically, DNA-methylation levels decreased over time, predominantly in tumors with malignant transformation and co-occurred with poor prognostic genetic events. DNA-methylation was lifted from specific loci associated with DNA replication and was associated with an increased RNA and protein expression of cell cycling associated genes. All results were validated on samples of newly diagnosed IDH-mutant astrocytoma patients included the CATNON randomized phase 3 clinical trial. Importantly, malignant progression was hardly affected by radio- or chemotherapy, indicating that treatment does not affect the course of disease. Our results culminate in a DNA-methylation based signature for objective tumor grading.