Harvard Division of Continuing Education

The identification of variants that occur in the human genome remains a critical step in the analysis of Next Generation Sequencing (NGS) data. Accurate and timely variant calling is essential to precision medicine which seeks to relate particular genomic variations in a patient’s genome to targetable genes, drugs, and treatments tailored to each patient. However, the analysis of these low-frequency variants requires sensitive algorithms within computational intensive pipelines. With high costs and lack of technology, most institutions resort to running bioinformatics pipelines as bash scripts on local clusters which are not only slow and cumbersome but also challenging to implement. This has in part led to reducing reproducibility and exhaustion of local data storage at such institutions. Therefore, I demonstrate how to implement bioinformatics pipelines in the cloud using open source tools like Github, Docker and Broad Institute Genome Analysis Toolkit (GATK). This fast and low-cost implementation leverages parallel execution and auto scaling within the cloud to handle the computationally intensive pipeline. I show that such pipelines produce accurate results by analyzing data from the 1000 Genomes Project. This would allow researchers to use more of their time doing research and less time configuring workflows.

The identification of variants that occur in the human genome remains a critical step in the analysis of Next Generation Sequencing (NGS) data. Accurate and timely variant calling is essential to precision medicine which seeks to relate particular genomic variations in a patient’s genome to targetable genes, drugs, and treatments tailored to each patient. However, the analysis of these low-frequency variants requires sensitive algorithms within computational intensive pipelines. With high costs and lack of technology, most institutions resort to running bioinformatics pipelines as bash scripts on local clusters which are not only slow and cumbersome but also challenging to implement. This has in part led to reducing reproducibility and exhaustion of local data storage at such institutions. Therefore, I demonstrate how to implement bioinformatics pipelines in the cloud using open source tools like Github, Docker and Broad Institute Genome Analysis Toolkit (GATK). This fast and low-cost implementation leverages parallel execution and auto scaling within the cloud to handle the computationally intensive pipeline. I show that such pipelines produce accurate results by analyzing data from the 1000 Genomes Project. This would allow researchers to use more of their time doing research and less time configuring workflows.

This project analyzes Massachusetts housing court records, testing if there is ethnic bias in who is evicted. The project starts with semi-manually scraping court records from the Massachusetts court website. Because the website disallows fully automatic scraping of records, tools were developed to facilitate and expedite the downloading of data without violating the terms and conditions for the page. With source web pages downloaded, the records were processed to extract eviction features: plaintiffs, defendants, location of eviction, and which housing court.Largely through format shifting, nearly a gigabyte of records was parsed and pared down to about 20MB stored in a usable data frame. Incomplete or unparsable records were removed. Once the raw data was processed, addresses were geocoded and turned into census tracts to add probabilities of ethnicity for each record, based on location.Additionally, last names were matched against census database records to extract the corresponding probabilities of ethnicity based on surname. These pieces of data were then combined to create a refined estimate of ethnicity for each data point. The estimated ethnicities were compared with the underlying demographics to determine if there was a meaningful difference in evictions between how many people of each ethnicity were evicted, and what should be expected based on the underlying population. Depending on how the model was tested, there were two different results. Either Caucasians are over-evicted relative to their population, or Hispanics and African Americans are over-evicted relative to their populations.

Detecting anomalies and predicting failure of industrial refrigeration equipment are paramount to guarantee a reliable supply chain and consumer safety in a variety of industries such as pharmaceutical and grocery. Failure can be predicted by performing an anomaly analysis on internal temperature data that has been collected by Internet of Things (IoT) sensors. Such an analysis involves monitoring a refrigeration unit’s defrost cycle, which is the seasonal (i.e. periodic) component of its temperature time series. In many industries, Recurrent Neural Networks (RNNs) are used to analyze and forecast time series data, and Long Short Term Memory (LSTM) cells are used to remember long term dependencies. When using deep learning tools to analyze time series data, a major challenge is modeling datasets with heterogeneous seasonal components. This thesis investigates the ability of RNNs built with LSTM cells to detect defrost events from within temperature time series that were recorded using IoT sensors. Because defrost events are seasonal and heterogeneous across time series, the successful detection of defrosts is dependent on an RNN’s ability to build a model of heterogeneous seasonality. We conducted our research in two phases. During the first phase, we generated datasets of simulated refrigeration temperature time series and used them to train and test RNNs, resulting in a 95% classification accuracy rate. During the second phase, we analyzed the challenges inherent in modeling heterogeneous seasonality in our time series datasets. We designed new experiments and modeled heterogeneous seasonality using binary datasets consisting of 1s and 0s. Using a binary dataset whose seasonal patterns were heterogeneous in both frequency and phase, RNNs achieved 100% forecasting accuracy. However, when we added confounding features to the dataset, forecasting accuracy dropped to 71% as confounding features could easily be confused with the seasonal patterns of time series data. Our results revealed that RNNs are best suited to model heterogeneous seasonality in datasets that do not contain confounding features.

In 2001, Pagh and Rodler described a single-threaded hash table design called cuckoo hashing (R. Pagh & Rodler, 2001). Since then, the rise of multi-core processors and large datasets have motivated the development and refinement of concurrent hash tables. We (a) give a brief review of relevant theory of hash functions and hash algorithms; (b) describe a multi-reader, multi-writer parallel cuckoo hash table design; and (c) describe the results of experiments with changing parameters including number of threads, failure criteria, and bucket selection strategies. We suggest multiple areas for future work, including experiments with lock granularity, varying the number of slots per bucket, and determining whether the depth-first approach of classic cuckoo hashing can have competitive performance under the proposed algorithm compared to a breadth-first approach.

In this project, I implemented a suite of software tools to demonstrate the feasibility of a speech audio processing application, i.e. automatic classification of conversational speech situations. First, I re-implemented a classical pitch tracking algorithm for speech processing - the RAPT algorithm (A Robust Algorithm for Pitch Tracking, Talkin 1995). Next, I collected and analyzed a novel corpus of podcast speech audio samples. Finally, I created an application-specific classification algorithm and evaluated its performance.

Growing attention has been drawn to energy use forecasting for smart grid applications. There has been a surge of interest in applying Gaussian process (GP) modeling to predicting building energy use. As a continuation of my previous projects, this thesis develops a web application that allows users without programming skills to predict and visualize energy demand through Gaussian process regression. The web application implements both baseline prediction and next-day prediction. This study also explores the visualization techniques that facilitate the analysis of energy consumption patterns and transform the data into informative insights. The technologies used in the web application include Flask, Heroku and Highcharts.js. This thesis presents two case studies to demonstrate the use of the web application and discusses the prediction accuracy. One case study is to predict electric energy, chilled water and steam consumption of a campus building. The second case study is to predict next-day electric energy demand of a high-tech industry area.

Technological advancements in blockchain solutions have allowed for the development of applications that were not previously viable. Many of those applications will require querying potentially vast amounts of transaction data. With a potential transaction throughput orders of magnitude higher than was possible until recently, such applications would benefit from delegating those queries to third party services with the computing power and infrastructures allowing the quick return of accurate results. The problem, however, with delegating queries, is that it creates a need for trust in a third party that blockchain technology was explicitly designed to eliminate. Such a third party query service should then be able to provide proof of the authenticity of the results it returns to its users. In this project, we build an API that provides fast and scalable random access to transaction data on the Algorand blockchain and allows queries according to a number of transaction parameters. A mathematical proof based on the principle of Zero-Knowledge Proof allows the user of the API to verify the authenticity and completeness of the queried transaction data. This work is based on the vChain framework described in the paper "vChain: Enabling Verifiable Boolean Range Queries over Blockchain Databases" by Cheng Xu, Ce Zhang, and Jianliang Xu.

Nobody likes sitting in traffic, at least the author of this work does not. In this thesis, we take the question of how can we improve traffic? We break it down and iterate on it, finally arriving at the problem of traffic optimization, via the optimization of traffic lights. This is the main driver of this research work. Traffic light control strategies have been previously classified as static and dynamic. We typically see static controllers in the fixed-cycle or actuated controllers that are widely deployed in the traffic lights of our cities; however, truly smart traffic lights are still not a widespread technology. This work tries to ground some of the research in dynamic traffic light controllers, so that smarter traffic lights become practical and ready for roll-out. We describe a Reinforcement Learning algorithm which we designed, and which only uses information that is local to the traffic light. A key element of this algorithm is the fact that we believe the inputs to it can be obtained by existing traffic intersection technologies, making its deployment feasible. We compare this algorithm against an optimal static traffic control policy, in the pursuit of understanding the strengths of both controller types. Finally, we analyze and present the results of this comparison, and illustrate a set of opportunities that may improve traffic management through different approaches, such as optimal traffic management, following the footsteps of this work or by having drivers collaborate with load balancing.

The advance of Generative Adversarial Networks (GANs) are able to create new and meaningful outputs given arbitrary encodings for three dimensional terrain generation. Using a two-stage GAN network, this novel system takes the terrain heightmap as the first processing object, and then maps the real texture image onto the heightmap according to the learned network. The synthetic terrain image results perform well and are realistic. However, improvements can be made to generator stability during training.

Myalgic Encephalomyelitis/Chronic Fatigue Syndrome (ME/CFS) is a condition that is characterized by a constellation of symptoms, including post-exertional malaise and disabling fatigue. Neither a blood test nor a single set of biomarkers exists for ME/CFS.
Dr. Jose G. Montoya, former Professor of Medicine, Infectious Diseases, and Geographic Medicine at the Stanford University Medical Center, provided us with a multidimensional data set consisting of samples from 192 ME/ CFS cases and 392 healthy controls. He eliminated participants with missing data, yielding a final sample size of 186 ME/CFS cases and 388 healthy controls. Each participant completed the MFI-20, a 20-item quality of life questionnaire, on the day of blood sample collection. Dr. Montoya measured each participant’s serum cytokine levels using a 51-multiplex array.
We divided this data into training and test sets to develop a binary predictive classifier to identify whether a study participant has ME/CFS. We tested eight different binary classification models. We fine-tuned the top three performing baseline models. We then employed the best of these three models, an elastic net generalized linear model binary classification algorithm to determine whether a study participant has ME/CFS. The model’s total accuracy is 75%, with an area under the return operator curve of 0.703, sensitivity of 0.439, and specificity of 0.93. This model is statistically significant at the α < 0.05 level with a Mcnemar's Test p-Value of 0.001. Next, we similarly evaluated eight regression prediction models for disease severity. As above, we tuned the best three of the eight baseline models. The top performer of these three models, evaluated on mean absolute error (MAE), was partial least squares regression. This two-component partial least squares model has a root mean square error (RMSE) of 25.33, a training coefficient of multiple determination R2 of 0.062, and a training MAE of 22.49, test RMSE of 25.86, and test MAE of 23.25. We successfully created a statistically significant binary classification model based upon 51 provided cytokine values in serum samples from ME/CFS patients and healthy controls. From the same cytokine values, combined with the MFI-20 survey measure of illness severity, we derived a regression and prediction model for illness severity. These findings may contribute to the development of a blood test for ME/CFS as well as the continuing hunt for disease biomarkers.

The recent growth of blockchain technology and its usage has increased the size of cryptocurrency networks. However, this increase has come at the cost of high energy consumption due to the processing power needed to maintain large cryptocurrency networks. In the largest networks, this processing power is attributed to wasted computations centered around solving a Proof of Work algorithm. There have been several attempts to address this problem and it is an area of continuing improvement. We will present a summary of proposed solutions as well as an in-depth look at a promising alternative algorithm known as Proof of Useful Work. This solution will redirect wasted computation towards useful work. We will show that this is a viable alternative to Proof of Work.