Genomics are shaping the future of precision medicine in cancer by enabling treatments to be tailored to the genetic signature of an individual's tumor. The Cancer Genome Atlas (TCGA) is an effort by the National Cancer Institute to catalog the genomic alterations in 31 different human cancers and to identify new disease sub-classifications that can improve cancer treatment strategies. This talk will discuss the TCGA analysis of lower-grade glioma (LGG) brain tumors, and the important role of learning algorithms in analyzing the multifaceted genomic data produced by TCGA data. I will present a summary of the findings of the LGG Analysis Working Group, and illustrate how computational approaches can be applied to public TCGA data to investigate fundamental questions in cancers.

Understanding and manipulating the machineries of cells and tissues requires a multi-scale approach that examines details of individual gene/protein, interactions of multiple proteins, and genetic circuit networks to elucidate complex behavior of cellular and tissue states and their spatio-temporal pattern formation. We describe new development in the fundamental theory to solve the discrete chemical master equation (dCME) that can account for stochastic control of rare and small copy-number events important for determining cellular fate, including the mb-dCME algorithm for exact time-evolving probabilistic landscape computation without Gillespie simulation or Fokker-Planck/Langevin approximation. We report discoveries of complex multistable landscapes arising from simple network motifs, without requring feedback loops or cooperativities that were thought to be obligatory. We further discuss control of epigenetic states, and the robustness of wild type versus mutant cells. The important roles of spatial confinement in shaping the scaling and folding landscape of chromosome structure whose misfunction are behind 70% of known cancer will be also highlighted, with emphasis on geometric sequential Monte Carlo method for denoising and for generating physical models. At the cellular level, we describe a geometric model and an algorithm for simulating dynamic spatio-temporal pattern formation of cell populations and how the process of wound healing can be studied through simulation. It our belief that this multi-scale approach will be useful for future engineering of biological systems.

Hidden Markov models (HMM) have been widely used to analyze large-scale genomic data because of its ability to incorporate spatial correlation, in areas such as genome-wide mapping of binding sites for regulatory proteins. With advances in sequencing technology, it is now common to analyze data for multiple proteins jointly, but existing algorithms analyze data for each protein separately. However, extending the HMM framework to a multivariate form is not trivial because hidden state space increases quickly with the number of experiments and it is therefore of interest to develop a powerful yet computationally tractable algorithm. Here we present a fast inferential method for correlated hidden Markov models for multiple sequential data (series). Instead of jointly inferring hidden states for all series, we adjust the transition kernel of each series with the hidden state configuration in other correlated series and keep the hidden state inference marginal in each. Through simulation, we show that the new scheme achieves sensitivity comparable to the fully coupled HMM fit at a computational cost as low as fitting an independent HMM for each series separately. The method was applied to the analysis of histone modification data in human T cells.

Machine learning provides great potential for big data in healthcare. However, for the clinic community to trust and use algorithmic outputs, the answer must generally map back to some notion of clinical understanding. This often means that supervised or semi-supervised learning is required. However, as the volume of data grows, the scale of data labeling becomes rapidly unassailable. Moreover, even when human labels are available, the large intra- and inter-observer variance means the accuracy of any algorithm can be much lower than is comfortable and the error rates can often be unknown. I will present an approach which leverages crowd sourcing to solve both these problems simultaneously and allow the non-parametric construction of confidence intervals in machine learning predictions.

It is commonly assumed in pattern recognition that cross-validation error estimation is ‘almost unbiased’ as long as the number of folds is not too small. While this is true for random sampling, it is not true with separate sampling, where the populations are independently sampled, which is a common situation in bioinformatics. We demonstrate, via analytical and numerical methods, that classical cross-validation can have strong bias under separate sampling, depending on the difference between the sampling ratios and the true population probabilities. We propose a new separate-sampling cross-validation error estimator, and prove that it satisfies an ‘almost unbiased’ theorem similar to that of random-sampling cross-validation. We present two case studies with previously published data, which show that the results can change drastically if the correct form of cross-validation is used.

Advances in computing, imaging and genomics technology present new opportunities for investigating relationships between histology, molecular events and clinical outcomes. Whole slide imaging and image analysis methods enable the mining of massive whole-slide digital image archives to capture quantitative features describing histological characteristics of cancer tissues. These measurements represent a new dimension in tissue-based studies, and when combined with genomic and clinical endpoints, can be used to explore biologic characteristics of the tumor microenvironment and to discover new morphologic biomarkers of genetic alterations and patient outcomes. Using motivating examples from the study of glioblastoma brain tumors (GBMs), we demonstrate how public data from The Cancer Genome Atlas (TCGA) can serve as an open plat-form to conduct in silico tissue based studies that integrate existing image/genomic/clinical data resources. We show how these approaches can be used to explore the relation of the tumor microenvironment to genomic alterations and gene expression patterns and to define nuclear morphometric features that are predictive of genetic alterations and clinical outcomes. Challenges, limitations and emerging opportunities in the area of quantitative imaging and integrative analyses are also discussed.

Drug sensitivity prediction based on genomic characterization remains a significant challenge in the area of systems medicine. We will discuss multiple approaches that have been proposed for mapping genomic characterizations to drug sensitivity. In particular, we will consider ensemble-based approaches and the gain in prediction by incorporating different heterogeneous genetic characterizations. We will also discuss an alternative approach where drug sensitivity prediction is considered based on drug-protein interactions and its integration with genomic data. We will conclude the tutorial by discussing experimental approaches to validate predictions provided by the modeling techniques.