Tyler O'Neal, Staff Editor HEALTH

Michigan team's predictive modeling provides a more accurate picture of how coughs disperse viruses

Scientists don't fully understand exactly how our coughs disperse virus particles into the air, but a new project led by the University of Michigan aims to provide a more accurate assessment of the dangers of close-proximity coughs and potentially lead to better barriers that can mitigate the hazards.

Previous research typically treats coughing as a single burst of air from the lungs, but that's not really how coughs work. More often, they're two or more pulses that expel droplets outward, creating swirling, turbulent airflows.

In the age of COVID-19, it's more important than ever that we understand those "multi-pulse" coughing events. They are getting a closer look from researchers at U-M and Auburn University. In addition to potential ramifications for COVID-19, it also could inform future scenarios where pathogen transmission between humans presents a major health risk. 

"There are a lot of key open questions in this space," said Jesse Capecelatro, U-M assistant professor of mechanical engineering. "There is evidence to suggest particles originating deeper within the lung carry more virus, and we want to know how those particles get dispersed during a cough.

"Do multiple pulses give rise to multiple vortex rings as the virus-laden air is expelled? And if so, what effect does that have on how the virus particles travel through space? Those are key things to know if you're trying to understand transmission and how to protect against it." Visualizations showing a snapshot of (top to bottom) single-, two-, and three-pulse coughing events. Colors show "fluid vorticity," or rotation. Vortex structures can be seen to persist closer to the mouth for two- and three-pulse cases. These swirls can accelerate particles and send them further from their source than typical cough models would predict. This is especially important because air in these later pulses comes from deeper in the lungs and, in the case of COVID-19, would likely carry higher viral load.

Using custom cough simulators connected to mannequins, as well as lasers and human subjects, researchers at Auburn University will make physical measurements of the airflow and expelled particles during simulated coughing. The mannequins will be outfitted with face masks and face shields. They'll use laser sheet imaging to allow for tracking of all particles expelled during coughing as they pass through a laser plane. This process will enable them to quantify not only the number of particles but also to track where they go, how they move with the airflow, and how fast they travel.

With that data from the high-resolution simulations and physical experiments in hand, Capecelatro and his U-M team will utilize high-performance supercomputing to generate improved predictive modeling, which will provide a more accurate picture of exactly how coughs disperse viruses.

U-M and Auburn researchers see the work as basic science that has the potential to shape how governments and communities respond to future outbreaks.

"The methods developed will be used to study how droplet-laden coughs interact with barriers such as face masks and face shields. And we'll evaluate how effective they are at blocking these flows and containing outbreaks," said Vrishank Raghav, assistant professor of aerospace engineering at Auburn.

"The new knowledge will also lead to the development of improved tools that can rapidly assess the risk of spreading infectious disease and assist in the development of public policy, such as physical distancing guidelines or indoor occupancy limits."

Yuan Yao and Kalvin Monroe, members of the Capecelatro Research Group, are contributors to the project. Capecelatro's fluid dynamics work has been in high demand this year due to the COVID-19 pandemic. He has studied particle flows on U-M buses, in hospital rooms, and in dental clinics. That work has contributed to changes in practice and policy designed to keep people safe.

For this project, the National Science Foundation has awarded the two institutions $465,000 over the next three years.

UK researchers use machine learning to rank cancer drugs in order of efficacy

Tyler O'Neal, Staff Editor HEALTH

Researchers from Queen Mary University of London have developed a machine learning algorithm that ranks drugs based on their efficacy in reducing cancer cell growth

Researchers from the Queen Mary University of London have developed a machine learning algorithm that ranks drugs based on their efficacy in reducing cancer cell growth. The approach may have the potential to advance personalized therapies in the future by allowing oncologists to select the best drugs to treat individual cancer patients.

The method, named Drug Ranking Using Machine Learning (DRUML), was published today in an academic journal and is based on machine learning analysis of data derived from the study of proteins expressed in cancer cells. Having been trained on the responses of these cells to over 400 drugs, DRUML predicts the best drug to treat a given cancer model.

Speaking of the new method, Professor Pedro Cutillas from the Queen Mary University of London, who led the study, said: "DRUML predicted drug efficacy in several cancer models and from data obtained from different laboratories and in a clinical dataset. These are exciting results because previous machine learning methods have failed to accurately predict drug responses in verification datasets, and they demonstrate the robustness and wide applicability of our method."

The research was funded by The Alan Turing Institute, Medical Research Council, Barts Charity, and Cancer Research UK.

How does DRUML work?

The team used datasets derived from proteomics (the study of proteins within cells) and phosphoproteomics (the study of how these proteins are modified) analyses of 48 leukemia, esophagus, and liver cancer cell lines as the input for DRUML to build models that may be applied to leukemia and solid tumors.

By training the models using the responses of these cells to 412 cancer drugs listed in drug response repositories, DRUML was able to produce ordered lists based on the effectiveness of the drugs to reduce cancer cell growth. The team then verified the predictive accuracy of the models using data obtained from 12 other laboratories and a clinical dataset of 36 primary acute myeloid leukemia samples.

Importantly, as new drugs are developed moving forwards, DRUML could be retrained to capture all clinically relevant cancer drugs.

Machine learning and personalized medicine

Cancers of the same type exhibit great variation in their genetic makeup and characteristics from patient to patient. In the clinic, this variation translates to patients having different responses to therapy. To address this issue, the field of personalized medicine aims to combine genetic insights with other clinical and diagnostic information to identify patterns that can allow clinicians to predict patient responses to therapies and select the most effective interventions.

The application of artificial intelligence and machine learning to biomedicine promises to aid personalized medicine and transform how cancers are diagnosed and treated in the future. This study represents a significant advancement in artificial intelligence in biomedical research and demonstrates that machine learning using proteomics and phosphoproteomics data may be an effective way of selecting the best drug to treat different cancer models.

Penn State researchers develop computational methods that allow for accurate determination of gene expression

Tyler O'Neal, Staff Editor HEALTH

More accurate measurement and interpretation of gene activities, using large volumes of sequencing data, may be possible with a new computational framework and set of algorithms currently being developed by Penn State researchers. A five-year, $1.85 million grant from the National Institutes of Health is funding the research led by Mingfu Shao, the Charles K. Etner Early Career Assistant Professor in the School of Electrical Engineering and Computer Science.

To understand how the machinery of a cell works, researchers frequently use RNA-sequencing (RNA-seq), which captures and measures the messenger RNA molecules (mRNAs) -- also called transcripts -- in cells. Because mRNAs copy and carry the genetic information of genes, measuring mRNAs is an efficient and accurate way to quantitively measure the gene activities. As such, researchers often use RNA-seq to study gene functions and cell machinery.

However, RNA-seq can only gather fragments of mRNAs, rather than full-length molecules.

"We need to computationally reconstruct the full-length sequences of the mRNAs from those short fragments," said Shao. "This is called assembly. A challenge, while also an opportunity, is that currently hundreds of thousands of RNA-seq samples have been stored in various repositories. Can we assemble them together? This is called meta-assembly. It means we want to assemble many, many samples together, instead of only one. We try to make use of the shared information across all those samples to improve the assembly accuracy." Mingfu Shao, the Charles K. Etner Early Career Assistant Professor in the School of Electrical Engineering and Computer Science

While accurate meta-assembly has not yet been achieved with current computational models, Shao has been developing a new framework to allow for many samples to be assembled at once, which would give researchers a clear understanding of the entire story told by the RNA-sequencing data.

The framework starts with multiple individual RNA-seq samples organized with a graph structure, called splice graphs. Several of these splice graphs are combined to generate another source of information called phasing paths.

"These phasing paths are very helpful in capturing the critical splicing information in individual samples," Shao said. "After merging the graph and generating the phasing path, we then decompose this combined graph into a set of paths. And each path will represent a predicted mRNA. This is the novel framework."

According to Shao, the more complete, accurate, and data-driven reconstruction of transcriptomes, which are the set of transcripts in a cell, could improve downstream RNA-seq analysis such as expression quantification and differential analysis. Researchers also would use the developed methods to study normal and diseased tissues and then identify the specific RNAs in the disease samples, which could then be used as biomarkers to help with diagnosis.

Shao said that because of the wide-spread use of transcriptomes in biomedical and biological research, he is excited about the myriad potential uses.

"Large-scale RNA-seq data has been deposited, and most of them are made publicly available to researchers," Shao said. "So, it's exciting that with our framework, we'll now have scalability. It will really enable the assembly of tens of thousands of samples at the same time. We expect that our developed methods together with existing data could have a high impact on biological and biomedical research."

In addition to meta-assembly, another direction Shao's group is exploring is the development of allele-specific assembly, allowing researchers to determine which expressed variant comes from which parent and perhaps mitigate genetic diseases.

"Current RNA-seq assembly methods don't distinguish between alleles, but maybe we could produce allele-specific assembly, for example, to tell people, 'Those mRNAs are from the maternal side, and those are from the paternal side,'" Shao said.

Shao's research also includes the reconstruction of mRNAs expressed in a single cell, instead of in tissue, using so-called single-cell RNA-sequencing data. This single-cell work is partly supported by a $400,000 grant from the National Science Foundation.