Amazing Science

A new AI program observed physical phenomena and uncovered relevant variables—a necessary precursor to any physics theory.

All physical laws are described as mathematical relationships between state variables. These variables give a complete and non-redundant description of the relevant system. However, despite the prevalence of computing power and artificial intelligence, the process of identifying the hidden state variables themselves has resisted automation. Most data-driven methods for modeling physical phenomena still rely on the assumption that the relevant state variables are already known. A longstanding question is whether it is possible to identify state variables from only high-dimensional observational data.

Scientists now created a principle for determining how many state variables an observed system is likely to have, and what these variables might be. They were able to demonstrate the effectiveness of this approach using video recordings of a variety of physical dynamical systems, ranging from elastic double pendulums to fire flames. Without any prior knowledge of the underlying physics, our algorithm discovers the intrinsic dimension of the observed dynamics and identifies candidate sets of state variables.

Github repository is here

While robots and computers will probably never completely replace doctors and nurses, machine learning/deep learning and AI are transforming the healthcare industry, improving outcomes, and changing the way doctors think about providing care.

Machine learning is improving diagnostics, predicting outcomes, and just beginning to scratch the surface of personalized care.

Imagine walking in to see your doctor with an ache or pain. After listening to your symptoms, she inputs them into her computer, which pulls up the latest research she might need to know about how to diagnose and treat your problem. You have an MRI or an X-ray and a computer helps the radiologist detect any problems that could be too small for a human to see. Finally, a computer looks at your medical records and family history and compares that with the best and most recent research to suggest a treatment protocol to your doctor that is specifically tailored to your needs.

Industry analysts predict that 30 percent of providers will use cognitive analytics with patient data by 2018.  It’s all starting to happen, and the implications are exciting.

Diagnosis

CBI insights identified 22 companies developing new programs for imaging and diagnostics. This is an especially promising field into which to introduce machine learning because computers and deep learning algorithms are getting more and more adept at recognizing patterns — which, in truth, is what much of diagnostics is about.

An IBM-backed group called Pathway Genomics is developing a simple blood test to determine if early detection or prediction of certain cancers is possible. Lumiata has developed predictive analytics tools that can discover accurate insights and make predictions related to symptoms, diagnoses, procedures, and medications for individual patients or patient groups.

Treatment

IBM’s Watson has been tasked with helping oncologist make the best care decisions for their patients.  The Care Trio team has developed a three-pronged approach that helps doctors devise and understand the best care protocols for cancer patients.

The CareEdit tool helps teams create clinical practice guidelines that document the best course of treatment for different types of cancers. CareGuide uses the information from CareEdit into a “clinical decision support system” to help doctors choose the right treatment plan for an individual patient. 

MIT researchers have developed a cryptographic system that could help neural networks identify promising drug candidates in massive pharmacological datasets, while keeping the data private. Secure computation done at such a massive scale could enable broad pooling of sensitive pharmacological data for predictive drug discovery.

Datasets of drug-target interactions (DTI), which show whether candidate compounds act on target proteins, are critical in helping researchers develop new medications. Models can be trained to crunch datasets of known DTIs and then, using that information, find novel drug candidates.

In recent years, pharmaceutical firms, universities, and other entities have become open to pooling pharmacological data into larger databases that can greatly improve training of these models. Due to intellectual property matters and other privacy concerns, however, these datasets remain limited in scope. Cryptography methods to secure the data are so computationally intensive they don't scale well to datasets beyond, say, tens of thousands of DTIs, which is relatively small.

In a paper published in Science, researchers from MIT's Computer Science and Artificial Intelligence Laboratory (CSAIL) describe a neural network securely trained and tested on a dataset of more than a million DTIs. The network leverages modern cryptographic tools and optimization techniques to keep the input data private, while running quickly and efficiently at scale.

The team's experiments show the network performs faster and more accurately than existing approaches; it can process massive datasets in days, whereas other cryptographic frameworks would take months. Moreover, the network identified several novel interactions, including one between the leukemia drug imatinib and an enzyme ErbB4—mutations of which have been associated with cancer—which could have clinical significance.

"People realize they need to pool their data to greatly accelerate the drug discovery process and enable us, together, to make scientific advances in solving important human diseases, such as cancer or diabetes. But they don't have good ways of doing it," says corresponding author Bonnie Berger, the Simons Professor of Mathematics and a principal investigator at CSAIL. "With this work, we provide a way for these entities to efficiently pool and analyze their data at a very large scale."

Birdsnap is a free electronic field guide covering 500 of the most common North American bird species, available as a web site or aniPhone app. Researchers from Columbia University and the University of Maryland developed Birdsnap using computer vision and machine learning to explore new ways of identifying bird species. Birdsnap automatically discovers visually similar species and makes visual suggestions for how they can be distinguished. In addition, Birdsnap uses visual recognition technology to allow users who upload bird images to search for visually similar species. Birdsnap estimates the likelihood of seeing each species at any location and time of year based on sightings records, and uses this likelihood both to produce a custom guide to local birds for each user and to improve the accuracy of visual recognition.

The genesis of Birdsnap (and its predecessors Leafnsap and Dogsnap) was the realization that many techniques used for face recognition developed by Peter Belhumeur (Columbia University) and David Jacobs(University of Maryland) could also be applied to automatic species identification. State-of-the-art face recognition algorithms rely on methods that find correspondences between comparable parts of different faces, so that, for example, a nose is compared to a nose, and an eye to an eye. In the same way, Birdsnap detects the parts of a bird, so that it can examine the visual similarity of comparable parts of the bird.Our first electronic field guide Leafsnap, produced in collaboration with the Smithsonian Institution, was launched in May 2011.

This free iPhone app uses visual recognition software to help identify tree species from photographs of their leaves. Leafsnap currently includes the trees of the northeastern US and will soon grow to include the trees of the United Kingdom. Leafsnap has been downloaded by over a million users, and discussed extensively in the press (see Leafsnap.com, for more information). In 2012, we launched Dogsnap, an iPhone app that allows you to use visual recognition to help determine dog breeds. Dogsnap contains images and textual descriptions of over 150 breeds of dogs recognized by the American Kennel Club.

For their inspiration and advice on bird identification, we thank the UCSD Computer Vision group, especially Serge Belongie, Catherine Wah, and Grant Van Horn; the Caltech Computational Vision group, especially Pietro Perona, Peter Welinder, and Steve Branson; the alumni of these groups Ryan Farrell (now at BYU), Florian Schroff (at Google), and Takeshi Mita (at Toshiba); and the Visipedia effort.

Last year, MIT researchers presented a system that automated a crucial step in big-data analysis: the selection of a "feature set," or aspects of the data that are useful for making predictions. The researchers entered the system in several data science contests, where it outperformed most of the human competitors and took only hours instead of months to perform its analyses.

Now, in a pair of papers at the IEEE International Conference on Data Science and Advanced Analytics, the team described an approach to automating most of the rest of the process of big-data analysis—the preparation of the data for analysis and even the specification of problems that the analysis might be able to solve.

The researchers believe that, again, their systems could perform in days tasks that used to take data scientists months.

"The goal of all this is to present the interesting stuff to the data scientists so that they can more quickly address all these new data sets that are coming in," says Max Kanter MEng '15, who is first author on last year's paper and one of this year's papers.

"Data scientists want to know, 'Why don't you show me the top 10 things that I can do the best, and then I'll dig down into those?' So, these methods are shrinking the time between getting a data set and actually producing value out of it."

Both papers focus on time-varying data, which reflects observations made over time, and they assume that the goal of analysis is to produce a probabilistic model that will predict future events on the basis of current observations.

The first paper describes a general framework for analyzing time-varying data. It splits the analytic process into three stages: labeling the data, or categorizing salient data points so they can be fed to a machine-learning system; segmenting the data, or determining which time sequences of data points are relevant to which problems; and "featurizing" the data, the step performed by the system the researchers presented last year.

The second paper describes a new language for describing data-analysis problems and a set of algorithms that automatically recombine data in different ways, to determine what types of prediction problems the data might be useful for solving.

According to Kalyan Veeramachaneni, a principal research scientist at MIT's Laboratory for Information and Decision Systems and senior author on all three papers, the work grew out of his team's experience with real data-analysis problems brought to it by industry researchers.

In 2014 at the Museums and the Web conferences in Baltimore, Maryland, and Florence, Italy, a group of data scientists presented a report on how two thousand museums were relating on Twitter. After our presentation, they decided to continue their investigation in two directions: The first are the case studies. They wanted to analyze some museums' Twitter environments, e.g., how museums are relating to their followers, how they form different communities defined by underlying relationships, the growth of the network around a museum, detection of influencers, etc. They also wanted to study how to use this information in order to design better social-media strategies for an improved and more realistic assessment of the work and actions’ results on Twitter.

Case studies from London's Victoria & Albert, Turin's Palazzo Madama, and Barcelona's Center for Contemporary Culture have chosen museums with different Twitter strategies, countries, languages, and follower bases ranging from 7,000 at the time of writing this abstract (Palazzo Madama) to the 450,000 from V&A. The goal was to have different situations represented and run a useful research for museums of different sizes. The second direction nowadays is how museum professionals are interacting with each other on Twitter thanks to the use of hashtags, like #musetech, in order to share information, questions, and experiences.

Maps have long been used to show the animal kingdom’srange, regional mix, populations at risk and more. Now a new set of maps reveals the global distribution of genetic diversity.

“Without genetic diversity, species can’t evolve into new species,” says Andreia Miraldo, a population geneticist at the Natural History Museum of Denmark in Copenhagen. “It also plays a fundamental role in allowing species populations to adapt to changes in their environment.”

Miraldo and her colleagues gathered geographical coordinates for more than 92,000 records of mitochondrial DNA from 4,675 species of land mammals and amphibians. The researchers compared changes in cytochrome b, a gene often used to measure genetic diversity within a species, and then mapped the average genetic diversity for all species within roughly 150,000 square-kilometer areas.

There is no unified place where genomics researchers can search through all available raw genomic data in a way similar to OMIM for genes or Uniprot for proteins. With the recent increase in the amount of genomic data that is being produced and the ever-growing promises of precision medicine, this is becoming more and more of a problem. DNAdigest is a charity working to promote efficient sharing of human genomic data to improve the outcome of genomic research and diagnostics for the benefit of patients. Repositive, a social enterprise spin-out of DNAdigest, is building an online platform that indexes genomic data stored in repositories and thus enables researchers to search for and access a range of human genomic data sources through a single, easy-to-use interface, free of charge.

If DNA archives become a plausible method of data storage, it will be thanks to rapid advances in genetic technologies. The sequencing machines that “read out” DNA code have already become exponentially faster and cheaper; the National Institutes of Health shows costs for sequencing a 3-billion-letter genome plummeting from US $100 million in 2001 to a mere $1,000 today. However, DNA synthesis technologies required to “write” the code are much newer and less mature. Synthetic-biology companies like San Francisco’s Twist Biosciencehave begun manufacturing DNA to customers’ specifications only in the last few years, primarily serving biotechnology companies that are tweaking the genomes of microbes to trick them into making some desirable product. Manufacturing DNA for data storage could be a profitable new market, says Twist CEO Emily Leproust.

Twist sent a representative to the April meeting, and the company is also working with Microsoft on a separate experiment in DNA storage, in which it synthesized 10 million strands of DNA to encode Microsoft’s test file. Leproust says Microsoft and the other tech companies are currently trying to determine “what kind of R&D has to be done to make a viable commercial product.” To make a product that’s competitive with magnetic tape for long-term storage, Leproust estimates that the cost of DNA synthesis must fall to 1/10,000 of today’s price. “That is hard,” she says mildly. But, she adds, her industry can take inspiration from semiconductor manufacturing, where costs have dropped far more dramatically. And just last month, an influential group of geneticists proposed an international effort to reduce the cost of DNA synthesis, suggesting that $100 million could launch the project nicely.

A large hard drive for a personal computer might have a terabyte of storage capacity. Now imagine filling one of those with data every day for around 161 million years, and you’ll have an idea how much data will be “created, captured, copied and consumed” in 2020 alone, according to market intelligence firm International Data Corporation (IDC).

This month, IDC released its updated Global DataSphere forecast, predicting more than 59 zettabytes of data creation and consumption within the year and a 26% compound growth rate in data growth through 2024. This massive amount of data – nearly double the data used in 2018 – will be enough that the next three years of data creation and consumption will eclipse that of the previous 30 years. These projections are roughly in line with IDC’s estimates from late 2018, which forecasted ~50 zettabytes of data use in 2020 and 175 zettabytes of data use by 2025. 

Perhaps most notably, the forecast shows a dramatic uptick in data consumption due to the boom in work-from-home activity caused by COVID-19. (This uptick is not shared by data creation, which has been slightly stymied by the pandemic.) IDC expects the gap between content creation and consumption to widen over the next few years, moving from a 1:9 ratio to a 1:10 ratio by 2024.

“Growth of the Global DataSphere is driven more by the data that we consume and analyze than what we create,” said David Reinsel, senior vice president for the Global DataSphere program. “Obviously, data must be created before it can be analyzed, but the recursion rate of data – the rate at which the same data is processed again – continues to grow exponentially driving the ‘unique’ DataSphere down to 10% of the total DataSphere.”

IDC expects that productivity data will quickly expand, with total enterprise data use gaining 4% on consumer data use (which stands at around 50%). Furthermore, IDC projects drastic increases in metadata and sensor data, potentially surpassing all other data types in the near future. Overall, 40% of the DataSphere is expected to be attributable to entertainment data (like Netflix), with productivity tools also driving an increase in video data use.

“We live in an increasingly video-enabled and video-assisted world, and consume an increasing amount of entertainment video each year – these are key factors driving the growth of the Global DataSphere,” said John Rydning, research vice president for the Global DataSphere program. “At the same time, we are steadily making more productive use of the video data we capture, which is contributing to the growth of productivity data in the DataSphere.”

In the age of Big Data, it turns out that the largest, fastest growing data source lies within your cells. Quantitative biologists at the University of Illinois Urbana-Champaign and Cold Spring Harbor Laboratory, in New York, found that genomics reigns as champion over three of the biggest data domains around: astronomy, Twitter, and YouTube.

The scientists determined which would expand the fastest by evaluating acquisition, storage, distribution, and analysis of each set of data. Genomes are quantified by their chemical constructs, or base pairs. Genomics trumps other data generators because the genome sequencing rate doubles every seven months. If it maintains this rate, by 2020 more than one billion billion bases will be sequenced and stored per year, or 1 exabase. By 2025, researchers estimate the rate will be almost one zettabase, one trillion billion bases, per sequence per year. “What does it mean to have more genomes than people on the planet?”—Michael Schatz, Cold Spring Harbor Laboratory

90 percent of the genome data analyzed in the study was human. The scientists estimate that 100 million to 2 billion human genomes will be sequenced by 2025. That’s a four to five order of magnitude of growth in ten years, which far exceeds the other three data generators they studied. “For human genomics, which is the biggest driver of the whole field, the hope is that by sequencing many, many individuals, that knowledge will be obtained to help predict and cure a variety of diseases,” says University of Illinois Urbana-Champaign co-author, Gene Robinson. Before it can be useful for medicine, genomes must be coupled with other genomic data sets, including tissue information.

One reason the rate is doubling so quickly is because scientists have begun sequencing individual cells. Single-cell genome sequencing technology for cancer research can reveal mutated sequences and aid in diagnosis. Patients may have multiple single cells sequenced, and there could end up being more than 7 billion genomes sequenced.

That “is more than the population of the Earth,” says Michael Schatz, associate professor at Cold Spring Harbor Laboratory, in New York. “What does it mean to have more genomes than people on the planet?” What it means is a mountain of information must be collected, filed, and analyzed.

“Other disciplines have been really successful at these scales, like YouTube,” says Schatz. Today, YouTube users upload 300 hours of video every minute, and the researchers expect that rate to grow up to 1,700 hours per minute, or 2 exabytes of video data per year, by 2025. Google set up a seamless data-flowing infrastructure for YouTube. They provided really fast Internet, huge hard drive space, algorithms that optimized results, and a team of experienced researchers.

Advances in high-throughput sequencing are reshaping how we perceive microbial communities inhabiting the human body, with implications for therapeutic interventions. Several large-scale datasets derived from hundreds of human microbiome samples sourced from multiple studies are now publicly available.

However, idiosyncratic data processing methods between studies introduce systematic differences that confound comparative analyses. To overcome these challenges, scientists developed GUTCYC, a compendium of environmental pathway genome databases (ePGDBs) constructed from 418 assembled human microbiome datasets using METAPATHWAYS, enabling reproducible functional metagenomic annotation. They also generated metabolic network reconstructions for each metagenome using the PATHWAYTOOLS software, empowering researchers and clinicians interested in visualizing and interpreting metabolic pathways encoded by the human gut microbiome. For the first time, GUTCYC provides consistent annotations and metabolic pathway predictions, making possible comparative community analyses between health and disease states in inflammatory bowel disease, Crohn’s disease, and type 2 diabetes. GUTCYC data products are searchable online, or may be downloaded and explored locally using METAPATHWAYS and PATHWAY TOOLS.

Each year, the American Cancer Society estimates the numbers of new cancer cases and deaths that will occur in the United States in the current year and compiles the most recent data on cancer incidence, mortality, and survival. Incidence data were collected by the Surveillance, Epidemiology, and End Results Program; the National Program of Cancer Registries; and the North American Association of Central Cancer Registries.

Mortality data were collected by the National Center for Health Statistics. In 2017, 1,688,780 new cancer cases and 600,920 cancer deaths are projected to occur in the United States. For all sites combined, the cancer incidence rate is 20% higher in men than in women, while the cancer death rate is 40% higher.

However, sex disparities vary by cancer type. For example, thyroid cancer incidence rates are 3-fold higher in women than in men (21 vs 7 per 100,000 population), despite equivalent death rates (0.5 per 100,000 population), largely reflecting sex differences in the “epidemic of diagnosis.” Over the past decade of available data, the overall cancer incidence rate (2004-2013) was stable in women and declined by approximately 2% annually in men, while the cancer death rate (2005-2014) declined by about 1.5% annually in both men and women. From 1991 to 2014, the overall cancer death rate dropped 25%, translating to approximately 2,143,200 fewer cancer deaths than would have been expected if death rates had remained at their peak.

Although the cancer death rate was 15% higher in blacks than in whites in 2014, increasing access to care as a result of the Patient Protection and Affordable Care Act may expedite the narrowing racial gap; from 2010 to 2015, the proportion of blacks who were uninsured halved, from 21% to 11%, as it did for Hispanics (31% to 16%). Gains in coverage for traditionally underserved Americans will facilitate the broader application of existing cancer control knowledge across every segment of the population.

Members of the public can search a newly released database of 1,600 stars to find signs of undiscovered exoplanets. The dataset, taken over two decades by the W.M. Keck Observatory in Hawaii, comes with an open-source software package and an online tutorial.

The search for planets beyond our solar system is about to gain some new recruits. Just recently, a team that includes MIT and is led by the Carnegie Institution for Science has released the largest collection of observations made with a technique called radial velocity, to be used for hunting exoplanets. The huge dataset, taken over two decades by the W.M. Keck Observatory in Hawaii, is now available to the public, along with an open-source software package to process the data and an online tutorial.

By making the data public and user-friendly, the scientists hope to draw fresh eyes to the observations, which encompass almost 61,000 measurements of more than 1,600 nearby stars.

“This is an amazing catalog, and we realized there just aren’t enough of us on the team to be doing as much science as could come out of this dataset,” says Jennifer Burt, a Torres Postdoctoral Fellow in MIT’s Kavli Institute for Astrophysics and Space Research. “We’re trying to shift toward a more community-oriented idea of how we should do science, so that others can access the data and see something interesting.”

Burt and her colleagues have outlined some details of the newly available dataset in a paper to appear in The Astronomical Journal. After taking a look through the data themselves, the researchers have detected over 100 potential exoplanets, including one orbiting GJ 411, the fourth-closest star to our solar system. “There seems to be no shortage of exoplanets,” Burt says. “There are a ton of them out there, and there is ton of science to be done.”

Modern experiment methods in biology can generate overwhelming amounts of raw data.  To manage this, scientists must create entirely new workflows and systems capable of merging large, disparate data sets and presenting them intuitively.  Arrowland is a combined -omics visualization tool that runs on a web browser, tablet, or cellphone.  It shows functional genomics data (fluxomics, metabolomics, proteomics and transcriptomics) together on a zoomable, searchable map, similar to the street maps used for navigation. In addition to providing a coherent layout for -omics measurements, the maps in Arrowland are also an easy-to-use reference for the relationships between the reactions, genes, and proteins involved in metabolism.

“At JBEI we have leveraged the insights provided by metabolic fluxes measured through 13C Metabolic Flux Analysis to make quantitative predictions for E. coli, and we shared these fluxes on Arrowland” said Hector Garcia Martin, Director of Quantitative Metabolic Modeling at JBEI. “By making this cloud-based interactive tool publicly available, our hope is to enable application of this tool in a wide range of fields, not only bioenergy. We are open to collaborations with the broader scientific community and industry.”

Setting itself apart from other available tools, Arrowland is beneficial for its clarity and ease of use. With its unique interface and presentation method, Arrowland makes the exploration of -omics data as intuitive as possible, on a platform that does not require special hardware or configuration.  Four different types of -omics data are integrated into the same map – an ability that no competing visualization platforms share – making important correlations and progressions easy to recognize and examine.

Arrowland is open-source and currently under active development, with plans to add more maps, editing and curation features, and is integrated with an open-source modeling package that JBEI plans to release separately, called the JBEI Quantitative Metabolic Modeling Library. For the time being, a sample data set from the flux profiles published in “A Method to Constrain Genome-Scale Models with 13C Labeling Data”, PLOS Computational Biology (Garcia Martin et al, 2015) can be viewed at http://public-arrowland.jbei.org.

IBM recently announced a new IBM Watson Data Platform that combines the world’s fastest data ingestion engine touting speeds up to 100+GB/second with cloud data source, data science, and cognitive API services. IBM is also making IBM Watson Machine Learning Service more intuitive with a self-service interface.

According to Bob Picciano, Senior Vice President of IBM Analytics “Watson Data Platform applies cognitive assistance for creating machine learning models, making it far faster to get from data to insight. It also, provides one place to access machine learning services and languages, so that anyone, from an app developer to the Chief Data Officer, can collaborate seamlessly to make sense of data, ask better questions, and more effectively operationalize insight.”

For more information or a free trial of IBM Watson Data Platform, Data Science Experience, Watson APIs, or Bluemix, the following resources are useful:

• http://www.ibm.com/analytics/us/en/watson-data-platform/
• http://datascience.ibm.com/blog/shiny-a-data-scientist-best-friend/
• http://datascience.ibm.com/blog/open-beta-available-for-everyone-2/
• http://www.ibmbigdatahub.com/blog/deeper-shade-simplicity-ibm-insight-world-watson-2016
• https://apsportal.ibm.com/analytics

The death toll amounts to 373,377 lost lives.

Max Galka of Metrocosm has made some lovely maps in 2015—tracking everything from obesity trends to property values to UFO sightings—but his latest effort may be the most powerful yet. Using data from the federal Fatality Analysis Reporting System, Galka maps every single U.S. road fatality from 2004 to 2013. The death toll amounts to 373,377 lost lives.

At the national level, Galka’s map almost looks like an electricity grid stretching across America’s road network—with bright orange clusters in metro areas connected via dim red threads across remote regions. Here’s a wide view of the whole country: