Qi Wei

The growing availability of biomedical data offers vast potential to improve human health, but the complexity and lack of integration of these datasets often limit their utility. To address this, the Biomedical Data Translator Consortium has developed an open‐source knowledge graph–based system—Translator—designed to integrate, harmonize, and make inferences over diverse biomedical data sources. We announce here Translator's initial public release and provide an overview of its architecture, standards, user interface, and core features. Translator employs a scalable, federated, knowledge graph framework for the integration of clinical, genomic, pharmacological, and other biomedical knowledge sources, enabling query retrieval, inference, and hypothesis generation. Translator's user interface is designed to support the exploration of knowledge relationships and the generation of insights, without requiring deep technical expertise and gradually revealing more detailed evidence, provenance, and confidence information, as needed by a given user. To demonstrate Translator's application and impact, we highlight features of the user interface in the context of three real‐world use cases: suggesting potential therapeutics for patients with rare disease; explaining the mechanism of action of a pipeline drug; and screening and validating drug candidates in a model organism. We discuss strengths and limitations of reasoning within a largely federated system and the need for rich concept modeling and deep provenance tracking. Finally, we outline future directions for enhancing Translator's functionality and expanding its data sources. Translator represents a significant step forward in making complex biomedical knowledge more accessible and actionable, aiming to accelerate translational research and improve patient care.

As large clinical and multiomics datasets and knowledge resources accumulate, they need to be transformed into computable and actionable information to support automated reasoning. These datasets range from laboratory experiment results to electronic health records (EHRs). Barriers to accessibility and sharing of such datasets include diversity of content, size and privacy. Effective transformation of data into information requires harmonization of stakeholder goals, implementation, enforcement of standards regarding quality and completeness, and availability of resources for maintenance and updates. Systems such as the Biomedical Data Translator leverage knowledge graphs (KGs), structured and machine learning readable knowledge representation, to encode knowledge extracted through inference. We focus here on the transformation of data from multiomics datasets and EHRs into compact knowledge, represented in a KG data structure. We demonstrate this data transformation in the context of the Translator ecosystem, including clinical trials, drug approvals, cancer, wellness, and EHR data. These transformations preserve individual privacy. We provide access to the five resulting KGs through the Translator framework. We show examples of biomedical research questions supported by our KGs, and discuss issues arising from extracting biomedical knowledge from multiomics data.

Objective: We applied propensity score matching method at scale on patient records to confirm signals of known drug effects on preterm birth and detect previously unidentified potential drug effects. Materials and Methods: This was a retrospective study on women who had continuity of care at Providence St. Joseph Health (PSJH) both before and after pregnancy and delivered live births between 2013/01/01 and 2022/12/31 (n=365,075). Our exposures of interest were all outpatient medications prescribed during pregnancy. We limited our analyses to medication that met the minimal sample size (n=600). The primary outcome of interest was preterm birth. Secondary outcomes of interest were small for gestational age and low birth weight. We used propensity score matching at scale to evaluate the risk of these adverse pregnancy outcomes associated with drug exposure after adjusting for demographics, pregnancy characteristics, and comorbidities. Results: The total medication prescription rate increased from 58.5% to 75.3% ( P <0.0001) from 2013 to 2022. The prevalence rate of preterm birth was 7.7%. 175 out of 1329 prenatally prescribed outpatient medications met the minimum sample size. We identified 58 medications statistically significantly associated with the risk of preterm birth (P≤0.1; decreased: 12, increased: 46). Discussion: We narrowed down from 1329 medications to 58 medications that showed statistically significant association with the risk of preterm birth even after addressing numerous covariates through propensity score matching. Conclusion: This data-driven approach demonstrated that multiple testable hypotheses in pregnancy pharmacology can be prioritized at scale, laying the foundation for application in other pregnancy outcomes.

Despite the societal burden of chronic wounds and despite advances in image processing, automated image-based prediction of wound prognosis is not yet in routine clinical practice. While specific tissue types are known to be positive or negative prognostic indicators, image-based wound healing prediction systems that have been demonstrated to date do not (1) use information about the proportions of tissue types within the wound and (2) predict time-to-healing (most predict categorical clinical labels). In this work, we analyzed a unique dataset of time-series images of healing wounds from a controlled study in dogs, as well as human wound images that are annotated for the tissue type composition. In the context of a hybrid-learning approach (neural network segmentation and decision tree regression) for the image-based prediction of time-to-healing, we tested whether explicitly incorporating tissue type-derived features into the model would improve the accuracy for time-to-healing prediction versus not including such features. We tested four deep convolutional encoder–decoder neural network models for wound image segmentation and identified, in the context of both original wound images and an augmented wound image-set, that a SegNet-type network trained on an augmented image set has best segmentation performance. Furthermore, using three different regression algorithms, we evaluated models for predicting wound time-to-healing using features extracted from the four best-performing segmentation models. We found that XGBoost regression using features that are (i) extracted from a SegNet-type network and (ii) reduced using principal components analysis performed the best for time-to-healing prediction. We demonstrated that a neural network model can classify the regions of a wound image as one of four tissue types, and demonstrated that adding features derived from the superpixel classifier improves the performance for healing-time prediction.

Background Immune-mediated inflammatory diseases (IMIDs) are likely to complicate maternal health. However, literature data on patients with IMIDs undergoing pregnancy is scarce and often overlooks the impact of comorbidities. Methods We investigated 12 selected IMIDs: psoriasis, inflammatory bowel disease, rheumatoid arthritis, spondyloarthritis, multiple sclerosis, systemic lupus erythematosus, psoriatic arthritis, antiphospholipid syndrome, Sjögren’s syndrome, vasculitis, sarcoidosis, systemic sclerosis. We characterized patients with IMIDs prior to pregnancy (IMIDs group) based on pregnancy/maternal characteristics, comorbidities, and pre-pregnancy/prenatal immunomodulatory medications (IMMs) prescription patterns. We 1:1 propensity score matched the IMIDs cohort with people who had no IMID diagnoses prior to pregnancy (non-IMIDs cohort). Outcome measures were preterm birth (PTB), low birth weight (LBW), small for gestational age (SGA), and cesarean section. Findings The prevalence rate of pregnancy occurring with people with a previous IMID diagnosis has doubled in the past ten years. We identified 5,784 patients with IMIDs. 17% of the IMIDs group had at least one prenatal IMM prescription. Depending on the type of IMM, from 48% to 70% of the patients taking IMMs before pregnancy continued them throughout pregnancy. Patients with IMIDs had similar but slightly increased risks of PTB (Relative risk (RR)=1·1[1·0, 1·3]), LBW (RR=1·2 [1·0,1·4]), SGA (RR=1·1 [1·0,1·2]), and cesarean section (RR=1·1 [1·1,1·2]) compared to a matched cohort of people without IMIDs. Out of the 12 selected IMIDs, three for PTB, one for LBW, two for SGA, and six for cesarean section had results supporting increased risk. Interpretation The association between IMIDs and the increased risk of adverse pregnancy outcomes depend on both the nature of the IMID and the presence of comorbidities. Funding NIH

Background: COVID-19 outcomes, in the context of immune-mediated inflammatory diseases (IMIDs), are incompletely understood. Reported outcomes vary considerably depending on the patient population studied. It is essential to analyse data for a large population, while considering the effects of the pandemic time period, comorbidities, long term use of immunomodulatory medications (IMMs), and vaccination status. Methods: In this retrospective case-control study, patients of all ages with IMIDs were identified from a large U.S. healthcare system. COVID-19 infections were identified based on SARS-CoV-2 NAAT test results. Controls without IMIDs were selected from the same database. Severe outcomes were hospitalisation, mechanical ventilation (MV), and death. We analysed data from 1 March 2020 to 30 August 2022, looking separately at both pre-Omicron and Omicron predominant periods. Factors including IMID diagnoses, comorbidities, long term use of IMMs, and vaccination and booster status were analysed using multivariable logistic regression (LR) and extreme gradient boosting (XGB). Findings: Out of 2 167 656 patients tested for SARS-CoV-2, there were 290 855 with confirmed COVID-19 infection: 15 397 patients with IMIDs and 275 458 controls (patients without IMIDs). Age and most chronic comorbidities were risk factors for worse outcomes, whereas vaccination and boosters were protective. Patients with IMIDs had higher rates of hospitalisation and mortality compared with controls. However, in multivariable analyses, few IMIDs were rarely risk factors for worse outcomes. Further, asthma, psoriasis and spondyloarthritis were associated with reduced risk. Most IMMs had no significant association, but less frequently used IMM drugs were limited by sample size. XGB outperformed LR, with the AUROCs for models across different time periods and outcomes ranging from 0·77 to 0·92. Interpretation: For patients with IMIDs, as for controls, age and comorbidities were risk factors for worse COVID-19 outcomes, whereas vaccinations were protective. Most IMIDs and immunomodulatory therapies were not associated with more severe outcomes. Interestingly, asthma, psoriasis and spondyloarthritis were associated with less severe COVID-19 outcomes than those expected for the population overall. These results can help inform clinical, policy and research decisions. Funding: Pfizer, Novartis, Janssen, NIH MeSH: D001327, D000086382, D025241, D012306, D000071069

We present a simple yet novel time series imputation technique with the goal of constructing an irregular time series that is uniform across every sample in a data set. Specifically, we fix a grid defined by the midpoints of non-overlapping bins (dubbed "slices") of observation times and ensure that each sample has values for all of the features at that given time. This allows one to both impute fully missing observations to allow uniform time series classification across the entire data and, in special cases, to impute individually missing features. To do so, we slightly generalize the well-known class imbalance algorithm SMOTE \cite{smote} to allow component wise nearest neighbor interpolation that preserves correlations when there are no missing features. We visualize the method in the simplified setting of 2-dimensional uncoupled harmonic oscillators. Next, we use tSMOTE to train an Encoder/Decoder long-short term memory (LSTM) model with Logistic Regression for predicting and classifying distinct trajectories of different 2D oscillators. After illustrating the the utility of tSMOTE in this context, we use the same architecture to train a clinical model for COVID-19 disease severity on an imputed data set. Our experiments show an improvement over standard mean and median imputation techniques by allowing a wider class of patient trajectories to be recognized by the model, as well as improvement over aggregated classification models.

Background: Multiple studies have shown the utility of transcriptome-wide RNA-seq profiles as features for machine learning-based prediction of response to chemotherapy in cancer. While tumor transcriptome profiles are publicly available for thousands of tumors for many cancer types, a relatively modest number of tumor profiles are clinically annotated for response to chemotherapy. The paucity of labeled examples and the high dimension of the feature data limit performance for predicting therapeutic response using fully-supervised classification methods. Recently, multiple studies have established the utility of a deep neural network approach, the variational autoencoder (VAE), for generating meaningful latent features from original data. Here, we report the first study of a semi-supervised approach using VAE-encoded tumor transcriptome features and regularized gradient boosted decision trees (XGBoost) to predict chemotherapy drug response for five cancer types: colon, pancreatic, bladder, breast, and sarcoma.Results: We found: (1) VAE-encoding of the tumor transcriptome preserves the cancer type identity of the tumor, suggesting preservation of biologically relevant information; and (2) as a feature-set for supervised classification to predict response-to-chemotherapy, the unsupervised VAE encoding of the tumor’s gene expression profile leads to better area under the receiver operating characteristic curve and area under the precision-recall curve classification performance than the original gene expression profile or the PCA principal components or the ICA components of the gene expression profile, in four out of five cancer types that we tested.Conclusions: Given high-dimensional “omics” data, the VAE is a powerful tool for obtaining a nonlinear low-dimensional embedding; it yields features that retain biological patterns that distinguish between different types of cancer and that enable more accurate tumor transcriptome-based prediction of response to chemotherapy than would be possible using the original data or their principal components.

Motivation: Multiple studies have shown the utility of transcriptome-wide RNA-seq profiles as features for machine learning-based prediction of response to chemotherapy in cancer. While tumor transcriptome profiles are publicly available for thousands of tumors for many cancer types, a relatively modest number of tumor profiles are clinically annotated for response to chemotherapy. The paucity of labeled examples and high dimension of the feature data limit performance for predicting therapeutic response using fully-supervised classification methods. Recently, multiple studies have established the utility of a deep neural network approach, the variational autoencoder (VAE), for generating meaningful latent features from original data. Here, we report first study of a semi-supervised approach using VAE-encoded tumor transcriptome features and regularized gradient boosted decision trees (XGBoost) to predict chemotherapy drug response for five cancer types: colon adenocarcinoma, pancreatic adenocarcinoma, bladder carcinoma, sarcoma, and breast invasive carcinoma. Results: We found: (1) VAE-encoding of the tumor transcriptome preserves the cancer type identity of the tumor, suggesting preservation of biologically relevant information; and (2) as a feature-set for supervised classification to predict response-to-chemotherapy, the unsupervised VAE encoding of the tumor’s gene expression profile leads to better area under the receiver operating characteristic curve (AUROC) classification performance than either the original gene expression profile or the PCA principal components of the gene expression profile, in four out of five cancer types that we tested. Availability: github.com/ATHED/VAE_for_chemotherapy_drug_response_prediction Contact: ramseyst@oregonstate.edu Supplementary information: Supplementary data are available at Bioinformatics online.