“In nature, nothing is perfect. Trees can be contorted, bent in weird ways and they’re still beautiful.”
- Alice Walker
On a genetic level, we humans often more resemble trees twisted in individual, complex ways than the neat diagrams presented in anatomy texts.
This makes it a whole lot more complicated to figure out how to keep people healthy and prevent, diagnose and treat disease.
Artificial intelligence, big data, and machine learning techniques have the potential to help researchers and medical professionals with this, though, as many research leaders discussed during San Francisco State University Department of Biology’s annual Personalized Medicine Conference. Held at the South San Francisco Conference Center May 31st, 2018, this gathering touched on the promises – and the complications – of data-driven medicine, individualized for particular groups of people, and ultimately, particular people.
Computers’ ability to store and analyze large sets of data can allow researchers to determine which patients are more or less likely to develop a disease or respond to a certain treatment. As keynote speaker Dr. Manuel Rivas, Assistant Professor of Biomedical Data Science at Stanford University pointed out, most individual human genes have only a weak link to a person’s health. Computerized examination of large data sets taken from many patients’ genomes might give us a better idea of which combinations of genes, acting together, would lead to greater or lesser risks of developing a condition, as his work pointed to with Type 2 diabetes.
Dr. Wyatt Clark, scientist working on research and development for BioMarin Pharmaceutical Inc., carried out research illustrating a similar idea related to Sanfilippo disease, a rare genetic condition leading to mental and physical disabilities in children. As he found, there was no one gene that would indicate the presence or absence of the disease, but several variants of different genes that seemed somehow linked to the condition. If enough of those variants appeared in a person’s genome, then that child would likely develop the disease.
Later that day, Dr. Kara Davis, Assistant Professor of Pediatrics at Stanford University, outlined her work involving developmental classification of immune B-cells in acute lymphoblastic leukemia patients. Those patients whose immune systems had more B cells whose genomes were already co-opted by the cancer at the time of diagnosis were more likely to relapse after treatment, and big data gave her the chance to identify this factor as an important indication of the future course of the disease.
Dr. Zemin Zhang, professor at the Beijing Advanced Innovation Center for Genomics at Peking University, in a panel moderated by Dr. Gini Deshpande, founder and CEO of NuMedii, discussed how analyzing the genomes of the cells in the immediate environment of a tumor can predict whether a patient will be likely to respond to immunotherapy. Saving a patient from unnecessary treatment can be good in itself, because it allows a sick person to avoid the stress, side effects, and risks of the therapy.
Dr. Taylor Jensen, Director of Research and Development at Sequenom, a LabCorp company, spoke on using big data to develop personal assays of cell-free DNA, which is DNA from a fetus that can circulate in the blood of a pregnant mother and allow for less invasive genetic tests to be done on the baby.
Some scientists, such as Dr. Gaia Andreoletti and Dr. Aashish Adhikari, Research Fellows of Computational Biology at UC Berkeley, pointed to the appeal of big-data by showing the limitations of some of our current research. From what they observed, screening of newborn babies for possible genetically linked disease using only the exome, the portion of DNA in our cells that actually contains code used to produce proteins, was not accurate enough to be used clinically. We need to consider the whole genome, including the sequences that regulate the expression of genes, to more accurately predict whether a child will develop a disease.
For some clinical applications, computer technology has already outpaced the work of human pathologists. Dr. Dexter Hadley, Assistant Professor of Pathology at the University of California – San Francisco, pointed to cases where computers could predict whether mammograms indicated the need for a biopsy for breast cancer more effectively than technicians.
Big data, and artificial intelligence-guided analysis of that data, can speed up the development of new pharmaceuticals. As Dr. Deshpande pointed out, most new drugs don’t make it through the clinical trial process because they are judged not to be effective enough. She and Dr. Khaled Sarsour, principal scientist at Genentech, talked about how personalized genetic medicine’s findings could greatly speed up drug development by giving insights early on about how many people, and which groups of people, are likely to respond to certain types of treatment.
That way researchers might be able to figure out, before investing millions of dollars and several years in creating a potential medicine, that the therapy might not be effective for large numbers of people in the demographic to which it will be targeted. Alternatively, if they find out that certain populations with certain kinds of genomes tend to respond well to one treatment or another, that indicates that scientists could pursue more treatments along those lines.
Scientists and medical professionals could ascertain some of this information on patient health and treatment effectiveness by observing patients over time, but the genomic analysis, and the big data/artificial intelligence technology, allows for faster and more efficient pattern recognition. Dr. Jeff Schrager, CTO of xCures, Inc. and adjunct professor at Stanford University, advocated for the creation of a global system, loosely modeled after air traffic control, to keep track of changing real-time observations from medical professionals about what sort of conditions patients have and from scientific researchers about what treatment approaches are being tried in order to get new medical findings right to the exact patients who need them. He says that we can completely overhaul our outdated clinical trial system within about five years, and that this is critical to effective public health.
Technology to advance our delivery of healthcare works best when coupled with humans who understand how best to work with it. Dr. Jochen Kumm, CEO of Healio Inc., discussed harnessing blockchain technology, where smaller bits of information are distributed among a larger network of computers for greater security and to access more computing power, to analyze and compare different variants of the human genome. He, and others, talked about machine learning, where computers can gradually refine their own algorithms and ‘teach’ themselves to become more accurate at tasks such as classifying and sorting data. This requires less input from human users than many people think.
Ultimately, researchers hope that computers will go beyond merely calculating, scanning and identifying data to test human-created hypotheses and start generating the hypotheses themselves based on the data.
Dr. Dragutin Petkovic, Professor of Computer Science at San Francisco State University, was heralded by conference organizers as being more cautious about AI, possibly even in disagreement with the other speakers. However, his talk more concerned how best to educate doctors and other clinicians about big data and medical AI technology, using illustrative examples and simulations so they would know how to interpret the results they received and how best to explain them to patients. This was a solid reminder that technology does not exist in a vacuum – it is part of an ecosystem in the medical world that includes human input at many levels.
Several speakers at the conference pointed to potential risks to consider when using AI technology for genome-related research. The data that we have currently is likely biased because people from some geographical areas and socio-economic groups are over-represented given their numbers in the world population. So if computers develop their sense of what is ‘normal’ based on that data, they could inherit our prejudices. As in the Alice Walker quote, people are highly individual and what we consider ‘normal,’ even something such as base body temperature, varies naturally among people. There’s more than one way to be ‘normal’ and ‘healthy’ and it can vary depending on the groups to which one belongs and one’s own experiences and environment.
Big data will definitely allow us to examine populations at a much higher level of dimensionality – grouping patients by ethnicity, gender, nationality, age, prior medical history, or any other form of classification. However, will a computer be able to figure out which sort of comparisons actually produce relevant information? There’s the joke about a new medical assistant who worked long and hard to organize a doctor’s medical files – with patient information arranged by their height! Speakers suggested that computers could learn to look at relevant factors over time – or perhaps discover new relationships among data that humans have not yet ascertained, which is the hope of much of this investment in technology.
Dr. Jochen Kumm had this retort when people told him that AI was a mysterious ‘black box’: ‘So’s the human brain! We don’t know too much about how that works, how it gets the answers we get, yet we still use it!’ He said that artificial intelligence held the potential to save lives and money so we should use it intelligently.
One success story was that of the final speaker, Dr. Michael Snyder, department chair of genetics at Stanford University. A firm believer in personalized medicine, he wears a smart watch that monitors his blood oxygen levels. This tipped him off that he was getting sick, likely with Lyme disease because he’d recently been to a tick-infested area, and he was able to ask for and receive early treatment for the condition. It took him a little bit of convincing to get his physician to go along with his smart watch, which highlights the need for a very human aspect of medicine: professionals listening to patients.
Hopefully, with the help of big data, artificial intelligence, machine learning, and a more personalized look at our individual genomes, we can come closer to understanding how medicine works best for each of us ‘tangled, twisted, and beautiful trees.’
After the conference, we mingled over refreshments and heard several graduate students discuss their work, which also incorporated technology. One group of students was developing an app connecting students to campus health services, and another was creating an app to help people determine when they were too drunk to drive. Others worked on understanding whether women experience pain differently from men.