Deep learning architectures such as variational autoencoders have revolutionized the analysis of transcriptomics data. However, the latent space of these variational autoencoders offers little to no interpretability. To provide further biological insights, we introduce a novel sparse Variational Autoencoder architecture, VEGA (VAE Enhanced by Gene Annotations), whose decoder wiring mirrors user-provided gene modules, providing direct interpretability to the latent variables. We demonstrate the performance of VEGA in diverse biological contexts using pathways, gene regulatory networks and cell type identities as the gene modules that define its latent space. VEGA successfully recapitulates the mechanism of cellular-specific response to treatments, the status of master regulators as well as jointly revealing the cell type and cellular state identity in developing cells. We envision the approach could serve as an explanatory biological model for development and drug treatment experiments.

While the clinical approval process is able to filter out medications whose utility does not offset their adverse drug reaction profile in humans, it is not well suited to characterizing lower frequency issues and idiosyncratic multi-drug interactions that can happen in real world diverse patient populations. With a growing abundance of real-world evidence databases containing hundreds of thousands of patient records, it is now feasible to build machine learning models that incorporate individual patient information to provide personalized adverse event predictions. In this study, we build models that integrate patient specific demographic, clinical, and genetic features (when available) with drug structure to predict adverse drug reactions. We develop an extensible graph convolutional approach to be able to integrate molecular effects from the variable number of medications a typical patient may be taking. Our model outperforms standard machine learning methods at the tasks of predicting hospitalization and death in the UK Biobank dataset yielding an R² of 0.37 and an AUC of 0.90, respectively. We believe our model has potential for evaluating new therapeutic compounds for individualized toxicities in real world diverse populations. It can also be used to prioritize medications when there are multiple options being considered for treatment.

The characteristic ionic currents of nucleotide kmers are commonly used in analyzing nanopore sequencing readouts. We present a graph convolutional network-based deep learning framework for predicting kmer characteristic ionic currents from corresponding chemical structures. We show such a framework can generalize the chemical information of the 5-methyl group from thymine to cytosine by correctly predicting 5-methylcytosine-containing DNA 6mers, thus shedding light on the de novo detection of nucleotide modifications.

Precision oncology has primarily relied on coding mutations as biomarkers of response to therapies. While transcriptome analysis can provide valuable information, incorporation into workflows has been difficult. For example, the relative rather than absolute gene expression level needs to be considered, requiring differential expression analysis across samples. However, expression programs related to the cell-of-origin and tumor microenvironment effects confound the search for cancer-specific expression changes. To address these challenges, we developed an unsupervised clustering approach for discovering differential pathway expression within cancer cohorts using gene expression measurements. The hydra approach uses a Dirichlet process mixture model to automatically detect multimodally distributed genes and expression signatures without the need for matched normal tissue. We demonstrate that the hydra approach is more sensitive than widely-used gene set enrichment approaches for detecting multimodal expression signatures. Application of the hydra analysis framework to small blue round cell tumors (including rhabdomyosarcoma, synovial sarcoma, neuroblastoma, Ewing sarcoma, and osteosarcoma) identified expression signatures associated with changes in the tumor microenvironment. The hydra approach also identified an association between ATRX deletions and elevated immune marker expression in high-risk neuroblastoma. Notably, hydra analysis of all small blue round cell tumors revealed similar subtypes, characterized by changes to infiltrating immune and stromal expression signatures. Pediatric cancers generally have few somatic mutations. To increase the number of actionable treatment leads, precision pediatric oncology initiatives also analyze tumor gene expression patterns. However, currently available approaches for gene expression data analysis in the clinical setting often use arbitrary thresholds for assessing overexpression and assume gene expression is normally distributed. These methods also rely on reference distributions of related cancer types or normal samples for assessing expression distributions. Often adequate normal samples are not available, and comparing matched cancer cohorts without accounting for subtype expression overestimates the uncertainty in the analysis. We developed a computational framework to automatically detect multimodal expression distributions within well-defined disease populations. Our analysis of small blue round cell tumors (including rhabdomyosarcoma, synovial sarcoma, neuroblastoma, Ewing sarcoma and osteosarcoma) discovered a significant number of multimodally expressed genes. Multimodally expressed genes were associated with proliferative signaling, extracellular matrix organization, and immune signaling pathways across cancer types. Expression signatures correlated with differences in patient outcomes for MYCN non-amplified neuroblastoma, osteosarcoma, and synovial sarcoma. The low mutation rate in pediatric cancers has led some to suggest that pediatric cancers are less immunogenic. However, our analysis suggests that immune infiltration can be identified across small blue round cell tumors. Thus, further research into modulating immune cells for patient benefit may be warranted.

In-silico modeling of patient clinical drug response (CDR) promises to revolutionize personalized cancer treatment. State-of-the-art CDR predictions are usually based on cancer cell line drug perturbation profiles. However, prediction performance is limited due to the inherent differences between cancer cell lines and primary tumors. In addition, current computational models generally do not leverage both chemical information of a drug and a gene expression profile of a patient during training, which could boost prediction performance. Here we develop a Patient Adapted with Chemical Embedding (PACE) dual convergence deep learning framework that a) integrates gene expression along with drug chemical structures, and b) is adapted in an unsupervised fashion by primary tumor gene expression. We show that PACE achieves better discrimination between sensitive and resistant patients compared to the state-of-the-art linear regularized method (9/12 VS 3/12 drugs with available clinical outcomes) and alternative methods.