The field of computing has made incredible strides in how we process and analyze information. Here I consider how to take these advances in theory and apply them in novel ways to different scopes of the medical pipeline: from the cellular level, to the clinic, and finally to the broadest level of healthcare and data management between different institutions. In particular, I sought to explore how the emergent technologies of deep learning and blockchain technologies can provide practical improvements to biomedical research, with a focus in mostly the neurodegenerative space. I investigated how to improve high-content high-throughput drug screens through deep learning—moving experimentation and analysis off the bench and into the servers as a first pass to generate hypotheses. There’s a plethora of information found in these screens. Through deep learning, such information can be refined to actionable insights, and complex algorithms can make sense of data that is uninterpretable to humans. So much time and money is lost in pharmaceutical endeavors due to following up on poor drug candidates. I hypothesized that if we had better predictions to begin with, then the search space for new drugs would be optimized. Hence, I developed an in-silico approach to phenotypic drug screens, by providing a deep learning method to glean related biological signals between markers and project them into images used for hypothesis generation. This resulted in better triaging efforts for finding active compounds in an Alzheimer’s-centered drug screen. More generally, I showed that this method can apply to another phenotypic screen in a completely different biological system. I showed that deep learning can not only assist at the bench, but also be useful in the clinical research space. Again focusing on neurodegenerative efforts, I developed deep learning methods to help neuropathologists identify amyloid beta pathologies. Following the idea that there is wisdom to be found in the crowd, I showed how consensus learning and ensembling over many experts can provide a robust framework for accurate detection despite being in a space where there may be expert disagreement. These efforts provided not only automation, but also potential for standardization in the field as I provided a method that is not only accurate but also completely consistent. I showed the method’s potential for practical adoption through a prospective validation study. Finally, at the broad level of data exchange between multiple medical institutions and data centers, I introduced a blockchain-based system into the clinical network, showing that blockchain technology has relevant applications outside of merely financial problems and questions. In particular, I made advances in the clinical trial process, and performed a proof-of-concept in which I simulated a real clinical trial on a more secure, traceable, and trustworthy data management system. We showed that such a system is immune to data tampering, and also provides a useful audit trail for regulators.

Though the title of my thesis infers a unifying theme via the application of machine learning, the two projects that form the bulk of my graduate degree are frankly more disparate than they are similar. Both endeavors provide novel methods to a field where ground truth is obscure and/or limited, and both apply machine learning techniques in their methodologies. Those similarities notwithstanding, the scientific domains, technical applications, experimental designs, and overall goals remain independent. While having a thesis comprised from two independent parts may not be conventional, this is, as they say, not a bug but a feature. Working within (and occasionally across) two research domains has helped me to acquire a diverse skillset and has provided me with a better and broader understanding of machine learning practices for scientific research.

As this thesis is composed of two linked, but distinct projects, the abstract (and chapters) is divided in two. The first section details work related to large-scale predictions of purchasable chemical space, and the second summarizes a novel method for automating diagnosis of melanocytic atypia in human histopathological samples.

I. Large-Scale Predictions for Purchasable Chemical SpaceThere are now over 400 million compounds one can easily purchase from the ZINC database (zinc.docking.org). About 350 million (85%) of these compounds are affordable enough for the average academic lab to conduct a ligand discovery project. However, the molecular targets (proteins) that these purchasable compounds bind and modulate—if any—are rarely known. Fewer than 1 million compounds (<0.25%) have been reported active in a target-specific assay according to public databases such as ChEMBL. In the absence of target activity information, the process of selecting compounds for general purpose screening will often be target-naïve.

To facilitate access to new chemistry for biology, my collaborator John Irwin and I generated predictions for all purchasable compounds in ZINC at the time. I explored methods for optimizing predictive performance of compound-target associations using ChEMBL’s bioactivity dataset (version 21) as a benchmark. Comparisons on cross-validation sets of the bioactivity dataset against several methods such as multinomial naïve-bayesion classifiers revealed that the combination of the Similarity Ensemble Approach (SEA) with the maximum Tanimoto similarity to the nearest bioactive yielded the best performance. I verified the utility for several of these predictions, quantified target prediction biases inherent to the dataset, and provided thresholding suggestions to the user for controlling sensitivity and specificity of the predictions, as well as novelty of target-associations allowed.

II. Automating Diagnosis of Melanocytic Atypia: A Precursor to Melanoma in SituMelanocytic atypia, a biological precursor to Melanoma, is histopathologically challenging. Pathologist interobserver agreement for melanocytic atypia in standard (H&E) histology images is low, ranging from 33-68%, with melanoma in situ (MIS) in particular contributing to diagnostic discordance. A lack of agreement among experts presents a challenge to any supervised learning task, where the utility of a learned function depends on the accuracy and reliability of labels used.

To circumvent the issue of pathological discordance in labeling, I paired H&E histology images with contiguously cut tissue sections, immunohistochemically (IHC) stained for melanocytes. I developed a deep-learning pipeline for automating diagnosis of melanocytic atypia using this custom dataset of paired, whole slide images (WSIs) and trained convolutional neural networks to identify the presence of melanocytes in H&E sections, using information solely from paired images. Networks achieved strong performance on holdout patient datasets. For each network trained, I generated full-scale (20X magnification) high-resolution (pixel-wise) prediction heatmaps on holdout tissue sections (H&E), for pathological interpretation, and applied saliency mapping to show what networks attend to in H&E images. This pipeline aims to provide assistance to the clinical pathologist to reach better consensus regarding new MIS diagnoses in cutaneous biopsies.

Large scale biological datasets are often comprised of observations which are noisy, whichare biased by environment or process, and which represent fragments of a perpetually grow-ing, yet incomplete record of human knowledge. Changes to computational methods, data deposition and storage, and improved collection of data have the potential to mitigate some of these problems. However, no one solution works for all problems, and care must be taken to ensure that a chosen method to make predictions for small molecules will be effective.

This thesis centers itself on prediction. How do we improve screening predictions made on biased and incomplete information? How do we better represent three-dimensional com-pounds to improve predictions when molecular shape is important? What methods might best inform our ability to make predictions now and improve the next step in the future?And how can we create testable hypotheses from phenotypic observations?

Chapter 1 presents the published work “Adding Stochastic Negative Examples into Ma-chine Learning Improves Molecular Bioactivity Prediction”. To address concerns over the effect that biased molecular affinity datasets may have on the accuracy of deep learning models, this work suggests an online method where to improve prediction when a dataset is made up of more binders than non-binders. The method, SNA, samples random, unannotated compounds and assigns them as non-binders during neural network training. SNA drastically improves the ability of the network to identify false positives in a full matrix of drugs and protein binders while slightly hurting performance on a time split.

Chapter 2 encompasses published work “A Simple Representation of Three-DimensionalMolecular Structure” which presents Extended Three-Dimensional Fingerprint (E3FP). This molecular fingerprinting technique generates a fingerprint that can represent three-dimensional structure for statistical and machine learning methods. Its advantages to two-dimensional fingerprints include the ability to encode structural relationships within a molecule and aggregation of fingerprints into a molecular ensembles. The E3FP was com-pared against existing two- and three-dimensional representations, and Chapter 2 shows some cases where the method outperformed these existing techniques.

Chapter 3 provides a brief commentary on the current outlook of deep learning for pre-diction of Adsorption, Distribution, Excretion, Metabolism, and Toxicity (ADMET). It de-scribes how changes to molecular representations of molecules for deep learning have improved prediction of ADMET endpoints. It speculates why techniques like neural network fine-tuning may be falling short. And, it pushes for deep learning interpretability and error estimation to improve trust in deep learning models and to facilitate iterative improvement of models.

Finally, unpublished work inChapter 4 focuses on how to use high throughput screening data to predict and rank a set of proteins to describe the clearance of free tau associated with applications for Alzheimer’s Disease.

Cellular biology has traditionally relied upon simple, low-dimensional single-cell measurements (e.g., microscopy and flow cytometry) which fail to adequately address cellular complexity, or high-dimensional aggregative measurements (e.g., bulk RNA-sequencing) which obscure cellular heterogeneity. Single-cell genomics technologies strike the ideal balance between measurement complexity and resolution, but have historically been hampered by two technological limitations: cell-cell doublets which confound data interpretation, and scalability limitations due to high reagent costs and complex parallel sample preparation workflows.

In this dissertation, I present solutions to both of these issues. First, I describe DoubletFinder, a machine learning approach for finding cell-cell doublets in scRNA-seq data by identifying real cells with heightened similarity to in silico-generated artificial doublets. Second, I describe MULTI-seq, a method enabling pooled single-cell genomics sample processing by labeling plasma membranes with sample-specific DNA barcodes prior to cellular isolation. After describing these technologies, I demonstrate three MULTI-seq applications. First, I explore the effects of sample pooling on single-cell RNA-sequencing (scRNA-seq) data-quality. Second, I extend MULTI-seq to single-cell epigenomics assays. And third, I leverage MULTI-seq to perform the largest-ever single-cell screen-by-sequencing experiment on PBMCs. Collectively, this dissertation documents molecular and computational tools for improving single-cell genomics scalability and data interpretability, and illustrates how these improvements expand the boundaries of feasibility for single-cell genomics experiments.

In-vivo phenotypic screening in larval zebrafish has shown much promise for neuroactive drug discovery, but unleashing its full potential remains a challenge. How do we robustly quantify how drugs modulate zebrafish behavior, and in parallel, how do we unravel which targets or pathways they act by? We develop a phenotypic screening and computational pipeline to begin meeting these challenges. Starting with motion index (MI) as a readout for phenotype, and correlation distance (CD) as a measure of phenotypic similarity, we extend the similarity ensemble approach to computationally predict targets for sets of phenotypic screening hits. Using this approach, we predict an “antipsychotic” target profile for previously uncharacterized hit compounds with MI’s matching those of known antipsychotic compounds. For a novel phenotype associated with sedation and paradoxical excitation caused by anesthetics such as etomidate and propofol, we predict not only the canonical GABAergic pathway, but a novel target entirely; the serotonin-6 receptor, which we validate with both in-vitro and in-vivo experiments.

However, our initial attempts at extending this approach to other known drug classes such as stimulants and convulsants are met with unexpected challenges; we hypothesize that the MI signatures and the CD used to compare might not be robust enough for these more subtle phenotypes. And so the Deepfish project is born. We train Siamese Neural Networks (SNNs) on a highly replicated screen of 650 known neuroactive drugs to learn a custom distance metric for comparing MI. This new distance function scores higher than CD at the task of separating same-drug replicate pairs versus different-drug pairs, all while generalizing to a quality control screen done months prior. In that arena, the new distance metric gets higher classification accuracy on average, but also strikingly outperforms CD for 3 drugs with more subtle phenotypes.

Armed with a way of training robust distance metrics, we make progress on using unsupervised deep-learning approaches to find more robust representations of behavior. We discover that for computing similarities between these high-dimensional embedded fingerprints, training custom distance metrics is even more imperative. However, we see signs that overfitting is possible with the Siamese Networks on our highly-replicated dataset - both with the raw MI and high-dimensional embedded representations - so we design and perform a version of the screen with fully randomized drug layouts, which we will use to benchmark our methods in the near future.

With the growth of biological and chemical datasets and the development of novel computational techniques, applications of artificial intelligence (AI) and machine learning (ML) methods that leverage these datasets to assist experimentalists become more critical than ever. This dissertation presents an overview of commonly used AL/ML tools for molecular biology and introduces two novel tools, as well as details their specific use cases. In Chapter 1, I provide a review of traditional techniques and their machine learning counterparts for ligand- and structure-based drug discovery and protein structure elucidation and design. In Chapter 2, I introduce Metric Ion Classification (MIC), a method for determining the identity of experimentally identified waters and ions in biomolecular structures. MIC builds upon recent advancements in protein-ligand interface representations and metric learning techniques to introduce a novel classification scheme with extensive validation on a variety of experimental structures. In Chapter 3, we present Autoparty, a tool for AI-assisted human-in-the-loop molecule annotation designed to facilitate the manual assessment of virtual screening results. Autoparty uses the principles of active learning to direct chemists toward useful compounds and limit the amount of labor required when evaluating compounds. These applications do not attempt to replace existing techniques; rather, they act in service of scientists to accelerate both structure determination and drug discovery pipelines. This work broadly highlights the utility of these tools and others like them and encourages their adoption alongside classical approaches.

The spatial organization of chromosomes within the cell nucleus facilitates critical genomic processes including transcription, replication, and repair. Understanding how DNA sequence informs genome folding and how chromatin conformation instructs transcription remains a central challenge. This dissertation presents computational strategies to advance our understanding of the principles governing three-dimensional chromatin structure and their implications for gene regulation. In Chapter 2, I perform large-scale in silico mutagenesis using a deep learning model to systematically uncover DNA sequences that encode folding patterns. Chapter 3 introduces new methods to quantify differences between chromatin interaction maps, revealing that integrating simple, map-informed and feature-based strategies provides the most complete perspective on functionally relevant organizational changes. Chapter 4 introduces and applies a non-negative matrix factorization method to decompose single-cell heterogeneity in chromatin structure, linking patterns in cell subpopulations to average folding principles and transcriptional consequences discerned in bulk. Together, these computational methods revealed new biological findings: repetitive elements, sometimes lacking CTCF motifs, provide sequence grammar governing chromatin interactions and the chromatin folding in only a small minority of cells often drives population-wide signals. The work broadly highlights the potential of computational approaches, especially machine learning, to accelerate discovery in genomics. This work provides templates for future studies relating sequence, spatial dynamics, and gene regulation amidst widespread variability in the folded genome.

The identification of protein function based on biological information is an area of intense research. Here we consider a complementary technique that quantitatively groups and relates proteins based on the chemical similarity of their ligands. We began with 65,000 ligands annotated into sets for hundreds of drug targets. The similarity score between each set was calculated using ligand topology. A statistical model was developed to rank the significance of the resulting similarity scores, which were expressed as networks to map the sets together. Although these networks were connected solely by chemical similarity, biologically sensible clusters nevertheless emerged.

When we used this "Similarity Ensemble Approach" to compare drugs to target sets, unexpected links emerged. Methadone, emetine, and Imodium were predicted and experimentally found to antagonize muscarinic M3, α2 adrenergic, and neurokinin NK2 receptors, respectively. Whereas drugs are intended to be selective, at least some bind to several physiologic targets, explaining their side effects and efficacy. We thereby sought further unexpected links by comparing a collection of 3,665 FDA-approved and investigational drugs against hundreds of targets. Chemical similarities between drugs and ligand sets predicted thousands of unanticipated associations. Thirty were tested experimentally, including the antagonism of the β1 receptor by the transporter inhibitor Prozac, the inhibition of the 5-HT transporter by the ion channel drug Vadilex, and the antagonism of the histamine H4 receptor by the enzyme inhibitor Rescriptor. Overall, 23 additional novel drug-target associations were confirmed, five of which were potent (< 100 nM). The physiological relevance of one, the drug DMT on serotonergic receptors, was confirmed in a knock-out mouse. This Similarity Ensemble Approach is systematic and comprehensive, and may suggest side-effects and new indications for many drugs.

Small molecule drugs also target many core metabolic enzymes in humans and pathogens. We therefore grouped and compared drugs and metabolites by their associated targets and enzymes, mapping these associations onto existing metabolic networks. This revealed what novel territory remains for metabolic drug discovery. We calculated these networks for 385 model organisms and pathogens. Chemical similarity links between drugs and metabolites may suggest drug toxicity, routes of metabolism, and polypharmacology.