Pathogens are becoming increasingly drug resistant, yet drug discovery methods have failed to produce new classes of antimicrobials for decades, thus there is an urgent need to identify effective therapies from existing FDA approved drugs. Multi-drug regimens are currently used to fight antibiotic resistance, but they are often chosen empirically, leading to suboptimal treatment outcomes and the spread of resistance.

To create new multi-drug treatment plans, computational tools are needed to narrow the vast sample size of FDA approved drugs in combination. Current computational methods rely on costly experimental data and black-box algorithms. Our model replaces omics data with drug-protein binding affinity calculations to predict effective drug combinations for Escherichia coli. Initial performance assessments show our model performs as well as models that require omics inputs. Molecular docking and neural networks were used to calculate an affinity between 59 drugs and 1499 proteins in E. coli. These drug-protein interactions were then used as features in the ML model.

During model construction the biochemical principles behind drug mechanisms of action were investigated by examining the extensive set of drug-protein interaction calculations as well as three omics studies covering chemogenomics, transcriptomics, and metabolomics. We have optimized a complex system of molecular-scale, protein-drug interactions with macro-scale, drug-drug interactions data to quickly predict drug therapies that could be used to treat deadly drug resistance pathogens.

Due to the flexible, multiscale, and hybrid nature of our model, many combinations, infeasible to interrogate via physical experiments due to cost and time, could be examined. Secondly, the model enables exploration of the underlying biological and chemical factors that influence drug mechanisms of action for better design of drug combination therapy. Our predictive model combines machine learning, deep learning, and physics-based molecular docking, which could impact and inspire future hybrid methodologies in AI and Data Science.

Changes resulting from a change in a single amino acid in a protein can be either disease causing or benign, due in large part to protein stability and protein binding with other proteins, nucleic acids, or small molecule ligands. We developed a database, the Annotated Database of Disease RElated Structures and Sequences (ADDRESS), mapping human genetic mutations to protein structures in the Protein Data Bank (PDB).

We found that mutations that shift the equilibrium more towards the unfolded (non-native) state are more often disease causing on average than those that approximately retain the stability. Interestingly, the threshold at which mutations become pathogenic is substantially less than the average stability of proteins in general, perhaps indicating the importance of cellular kinetics in a system where proteins are constantly degraded and misfolded.

We built decision trees inclusive of various topology relations and found that the cross relation was especially indicative of whether the mutation causes disease, in the case of non-essential proteins with low stability change. We also found that, in the case of treatability of a set of lysosomal storage disorders, stability change, binding to ligand, and an aspect of topology likely related to kinetics of the system were important in indicating whether a drug was effective. Incorporation of binding and aggregation propensity will build upon the current database.

Chromatin architecture, a key regulator of gene expression, can be inferred using chromatin contact data from chromosome conformation capture or Hi-C technology. However, classical Hi-C does not preserve multi-way contacts. Here we use long sequencing reads to map genome-wide multi-way contacts and investigate higher order chromatin organization in the human genome. Multiway chromatin contact data captured with Pore-C technology contains structural information beyond the pairwise data captured with traditional Hi-C. This allows for more precise representation of chromatin architecture and lends itself to efficient representation by hypergraphs to capture this higher order network structure.

We use hypergraph theory for data representation and analysis, and quantify higher order structures in neonatal fibroblasts, biopsied adult fibroblasts, and B lymphocytes. Hypergraphs and tensors are natural representations of the contact structure in the genome.

Furthermore, we investigated the relationship between the multiway and pairwise data captured with Pore-C and Hi-C technology. By integrating multi-way contacts with chromatin accessibility, gene expression, and transcription factor binding, we introduce a data-driven method to identify cell type-specific transcription clusters. We provide transcription factor-mediated functional building blocks for cell identity that serve as a global signature for cell types.

Gabrielle Dotson, Can Chen, Stephen Lindsly, Anthony Cicalo, Sam Dilworth, Charles Ryan, Sivakumar Jeyarajan, Walter Meixner, Cooper Stansbury, Joshua Pickard, Nicholas Beckloff, Amit Surana, Max Wicha, Lindsey Muir, and Indika Rajapakse. “Deciphering Multi-way Interactions in the Human Genome.” Nature Communications, in Press (2022).

Joshua Pickard, Rahmey Salhm, Can Chen, Amit Surana, Indika Rajapakse. “Hypergraph Analysis Toolbox for Long Read Sequencing,” Manuscript in preparation

Autonomous vehicles represent one of the most active technologies currently being developed, with research areas addressing, among others, the modeling of the states and behavioral elements of the occupants. This paper contributes to this line of research by studying the circadian rhythm of individuals using a novel multimodal dataset of 36 subjects consisting of five information channels. These channels include visual, thermal, physiological, linguistic, and background data.

Moreover, we propose a framework to explore whether the circadian rhythm can be modeled without continuous monitoring and investigate the hypothesis that multimodal features have a greater propensity for improved performance using data points specific to certain times during the day. Our analysis shows that multimodal fusion can lead to an accuracy of up to 77% on identifying energized and enervated states of the participants. Our findings highlight the validity of our hypothesis and present a novel approach for future research.

Summary: Antibiotic resistance is becoming a significant public health concern worldwide, with few novel treatments being discovered. Drug combination therapy is a promising solution against antibiotic resistance. But, the search for effective drug combinations within a vast combinatorial space is time- and resource-intensive. We developed an ML-based algorithm that, (1) predicts effective multi-drug therapies, (2) utilizes multi-omics datasets for uncovering complex drug-drug interactions, (3) overcomes the need for high-cost experimental datasets, (4) provides the lucid interpretation of model predictions, and (5) accounts for drug toxicity profiles to design safer treatments. Current approaches cannot address more than one of the above concerns simultaneously.

Methods: The approach involves a three-step process, (1) generating feature profiles for individual drug treatments using multi-omics data such as metabolomics, proteomics, structural profiles, etc., followed by (2) preprocessing to compute joint profiles for a combination accounting for similarity and uniqueness among drugs in the treatment, and (3) feeding the information to the ML algorithm for model development, and evaluating performance.

Results: Performance for the approach was evaluated on several drug combination datasets, two-way interactions (R=0.6015***), two-way interactions in Glycerol (R=0.5178***), three-way interactions (R=0.4515***), sequential interactions (R=0.4337***) [*** indicates p-value < 1e-3]. The trained model was interpreted using the Testing with Activation Concept Vectors (TCAV) approach, which concludes that subsystems like Pyruvate metabolism, TCA cycle, and Oxidative Phosphorylation play an important role in predictions made by the model. Additionally, the trained model was fine-tuned to predict the toxicity of combination therapy to ensure the safety of the treatment.

Impact: As novel treatments are not readily discovered, it makes it crucial to design treatments using approved FDA drugs. Our developed algorithm aims to provide an innovative and unique perspective on utilizing machine learning in guiding the development of multi-drug treatments using FDA drugs.

Managing file-based workflows is a cross-disciplinary headache. We highlight how tools originally developed to manage simulation data can be applied to simplify certain tasks in machine learning and data science, like generating data, selecting models, and streamlining the hyperparameter optimization of neural networks. The signac framework consists of three Python packages to help organize file-based projects, define reproducible computational workflows, and explore the data. It provides a command line and Python interface to access and manage project data as well as submit cluster jobs to high performance computing schedulers.

Signac implements a file-based database with no need to explicitly define a data schema. Signac organizes collections of parameter values as signac jobs and stores them in a flat directory structure. Using the command line or Python query interface, you can access data stored in the job directory, get job-specific file paths, and generate human-readable directory structures for sharing. This frees you from thinking about the minutiae of file organization and lets the data schema evolve with the project.

Signac-flow lets you define a computational workflow composed of operations with pre- and post-conditions. Using the command line interface, operations can be run locally or submitted to high performance computing systems and features built-in support for GreatLakes. The signac-dashboard package helps you inspect and filter jobs in a signac project. It runs a local web server and can interactively display files, videos, and images such as learning curves.

Developers and users are active in the Slack channel and happy to welcome new users. Check out signac.io for more.

Optical multi-layer thin films are widely used in optical and energy applications requiring photonic designs. Engineers often design such structures based on their physical intuition. However, solely relying on human experts can be time-consuming and may lead to sub-optimal designs, especially when the design space is large.

In this work, we frame the multi-layer optical design task as a sequence generation problem. Based on reinforcement learning, a deep sequence generation network is proposed for efficiently generating optical layer sequences. We train the deep sequence generation network with proximal policy optimization to generate multi-layer structures with desired properties. The proposed method is applied to two energy applications.

Our algorithm successfully discovered high-performance designs, outperforming structures designed by human experts and state-of-art algorithms. We believe our algorithm based on reinforcement learning can extend to many other multi-layer tasks and achieve high performance.

Moiré patterns in van der Waals heterostructures have lately received a significant focus in 2D materials research. They are open to external control through the twist-angle between the layers while providing significant impact on the band-structure and properties of the heterostructure, such as “magic-angle” superconducting graphene. While existing nanoscale measurement techniques such as transmission electron microscopy and near-field tip-enhanced microscopy are able to directly measure the Moiré patterns, these techniques are typically slow, costly, and often require sample preparation incompatible with other measurements and experiments.

We attempt to overcome these limitations by applying machine learning to the far-field data scattered through a metalens placed in the near-field region of the sample. The metalens, which consists of a collection of dipole resonators placed in the near field of the sample, is able to scatter the evanescent high-spatial-frequency near-field information to detectors in the far field. Using a U-Net convolutional neural network trained according to the metalens scatterer arrangement, we are able to reconstruct the near-field pattern from the scattered far-field data.

We model the problem using a simulation of the metalens that models the interaction of the resonant dipoles with the near-field as well as the dipole-dipole interactions within the metalens. This allows us to quickly generate a training dataset of tens of thousands of near-field configurations and far-field output. Using this, we investigate the effect of different metalens designs on the near-field reconstruction.

These results pave the way for future physical implementations to allow direct single-shot measurement of Moiré lattices in heterostructures in the far-field. They may find further application in nanofabrication metrology, enabling single-shot optical measurement of subwavelength features beyond current scanning optical techniques or electron microscopy.

Metal-organic frameworks (MOFs) are the pioneering candidates for solving some of the grand challenges of our society, including clean energy, carbon dioxide capture, and water purification. The crystalline nanoporous structure of these materials is advantageous for such applications compared to other solid-state materials.

However, the stability (i.e., structural integrity) of many MOFs is compromised under different operating conditions (e.g., temperature, pressure, chemical environment). Often, stability information of MOFs under these conditions is unavailable. Determining the stability of MOFs at different physicochemical conditions is a tedious experimental exercise involving multiple characterization methods. Experimentally examining the stability of MOFs under various operating conditions is impractical for the over 100,000 already-synthesized MOFs, and a standardized computational approach to determine MOFs’ stability is not available.

Here we report a comprehensive data-driven approach to predict the thermal, chemical, and mechanical stabilities of MOFs. We combine cheminformatics and materials informatics feature engineering approaches for training machine learning (ML) models. We develop four optimized ML models for the prediction of thermal, mechanical, and solvent removal stability of MOFs. The predictive performance of our ML models for thermal and solvent removal stability are better than those reported elsewhere. In principle, our models can be used for the prediction of stability of an arbitrary MOF under different operating conditions.

Designing optical structures for generating structural colors is challenging due to the complex relationship between the optical structures and the color perceived by human eyes. Machine learning-based approaches have been developed to expedite this design process. However, existing methods solely focus on structural parameters of the optical design, which could lead to sub-optimal color generation due to the inability to optimize the selection of materials.

To address this issue, an approach Neural Particle Swarm Optimization is proposed. The proposed methods combine the mixture density networks as well as optimization and achieves high design accuracy and efficiency on two structural color design tasks; the first task is designing environmental-friendly alternatives to chrome coatings and the second task concerns reconstructing pictures with multilayer optical thin films. Several designs that could replace chrome coatings have been discovered; pictures with more than 200,000 pixels and thousands of unique colors can be accurately reconstructed in a few hours.

Cities shape people. But can people shape cities? Urban planners and designers experience a dearth of insightful scientific tools that assist them in urban research. Moreover, predominant data analysis in urban scholarship has used data to create city systems and structures that have shaped people’s access to the city and its resources – albeit inequitably.

This pattern throughout history to use data, mapping, and systematic planning to suppress the voice of the vulnerable mandates a shift in perspective. Contrary to being the “new” oil, the research argues that data has always been a resource of the powerful. Most often, data has been the key to drawing the larger pictures and connecting relationships between apparently disparate entities, resourcefully sectioned for the benefit of a few privileged groups and before the less powerful people have had a chance to get a sense for or offer reflections on the macro picture. Understandably the first known traces of data analysis and mapping were created to fight and win wars. They have been strategic attempts to achieve goals that were not always about overarching good causes.

Metricle is a counter data tool in the making that allows researchers to analyze the built landscape to retrace systemic neglect in neighborhoods from the traditional top down efforts to offer a counter-lens through which to view city design and planning. A perspective that correlates people’s needs with the existing physical landscape as opposed to performing analysis to merely establish dominion. It does this through a systematic investigation of spatial imagery in conjunction with available census and external data about the place. It uses a novel technique to co-relate and associate various related and interdependent metrics to find broad trends or anomalies in the data. These captured trends or outliers are then studied in relation to the deductions made from critical observation of spatial imagery to draw causation and speculate on recommendations.

Bicycling is a promising transport mode to make our communities more sustainable, healthier, and more equitable.Motor vehicle traffic volumes, often measured by the Annual Average Daily Traffic (AADT), have been widely used in making engineering decisions. However, little data on bicycle traffic volumes have been collected and used in most U.S. cities.

In this project, we collected bicycle traffic data using a commercial automated bike counter at two locations: a multi-use path in Dearborn and a protected bike lane in Ann Arbor. Validation studies were conducted to examine the counter accuracy using video cameras. A total of 9 weeks data collection was conducted with more than 13,000 people on bikes counted as of today (the data collection in Ann Arbor is still ongoing). Data analysis was conducted to examine the bicycling traffic patterns in terms of traffic in different time of day, day of the week, and weather conditions. The primary trip purposes (i.e., commuting, recreation) at each location can be inferred from the patterns.

In addition, an open-source, public interactive dashboard (linked here) was developed that allows other researchers, traffic engineers, city planners, and the general public to freely explore the bike traffic data. The dashboard supports selecting date ranges, traffic directions, and data resolution (e.g., daily, hourly, 15-minutes). The outcome of this work can be used to get insights of bicycle infrastructure usages and support data-driven decision-making by the city planners and community engagement.