Artificial Intelligence is getting better than humans at detecting breast cancer
A recent study in The Lancet Oncology showed that AI found 20 percent more cancers on mammogram screens than radiologists alone.
Since the early 2000s, AI systems have eliminated more than 1.7 million jobs, and that number will only increase as AI improves. Some research estimates that by 2025, AI will eliminate more than 85 million jobs.
But for all the talk about job security, AI is also proving to be a powerful tool in healthcare—specifically, cancer detection. One recently published study has shown that, remarkably, artificial intelligence was able to detect 20 percent more cancers in imaging scans than radiologists alone.
Published in The Lancet Oncology, the study analyzed the scans of 80,000 Swedish women with a moderate hereditary risk of breast cancer who had undergone a mammogram between April 2021 and July 2022. Half of these scans were read by AI and then a radiologist to double-check the findings. The second group of scans was read by two researchers without the help of AI. (Currently, the standard of care across Europe is to have two radiologists analyze a scan before diagnosing a patient with breast cancer.)
The study showed that the AI group detected cancer in 6 out of every 1,000 scans, while the radiologists detected cancer in 5 per 1,000 scans. In other words, AI found 20 percent more cancers than the highly-trained radiologists.
Scientists have been using MRI images (like the ones pictured here) to train artificial intelligence to detect cancers earlier and with more accuracy. Here, MIT's AI system, MIRAI, looks for patterns in a patient's mammograms to detect breast cancer earlier than ever before. news.mit.edu
But even though the AI was better able to pinpoint cancer on an image, it doesn’t mean radiologists will soon be out of a job. Dr. Laura Heacock, a breast radiologist at NYU, said in an interview with CNN that radiologists do much more than simply screening mammograms, and that even well-trained technology can make errors. “These tools work best when paired with highly-trained radiologists who make the final call on your mammogram. Think of it as a tool like a stethoscope for a cardiologist.”
AI is still an emerging technology, but more and more doctors are using them to detect different cancers. For example, researchers at MIT have developed a program called MIRAI, which looks at patterns in patient mammograms across a series of scans and uses an algorithm to model a patient's risk of developing breast cancer over time. The program was "trained" with more than 200,000 breast imaging scans from Massachusetts General Hospital and has been tested on over 100,000 women in different hospitals across the world. According to MIT, MIRAI "has been shown to be more accurate in predicting the risk for developing breast cancer in the short term (over a 3-year period) compared to traditional tools." It has also been able to detect breast cancer up to five years before a patient receives a diagnosis.
The challenges for cancer-detecting AI tools now is not just accuracy. AI tools are also being challenged to perform consistently well across different ages, races, and breast density profiles, particularly given the increased risks that different women face. For example, Black women are 42 percent more likely than white women to die from breast cancer, despite having nearly the same rates of breast cancer as white women. Recently, an FDA-approved AI device for screening breast cancer has come under fire for wrongly detecting cancer in Black patients significantly more often than white patients.
As AI technology improves, radiologists will be able to accurately scan a more diverse set of patients at a larger volume than ever before, potentially saving more lives than ever.
A vaccine from Moderna could be available as early as January, Fauci recently said.
Earlier this year, biotech company Moderna broke world records for speed in vaccine development. Their researchers translated the genetic code of the coronavirus into a vaccine candidate in just 42 days.
We're about to expand our safety data in Phase II.
Phase I of the clinical trial started in Seattle on March 16th, with the already-iconic image of volunteer Jennifer Haller calmly receiving the very first dose.
Instead of traditional methods, this vaccine uses a new -- and so far unproven -- technology based on synthetic biology: It hijacks the software of life – messenger RNA – to deliver a copy of the virus's genetic sequence into cells, which, in theory, triggers the body to produce antibodies to fight off a coronavirus infection.
U.S. National Institute of Allergy and Infectious Diseases Director Anthony Fauci called the vaccine's preclinical data "impressive" and told National Geographic this week that a vaccine could be ready for general use as early as January.
The Phase I trial has dosed 45 healthy adults. Phase II trials are about to start, enrolling around 600 adults. Pivotal efficacy trials would follow soon thereafter, bankrolled in collaboration with the government office BARDA (Biomedical Advanced Research and Development Authority).
Today, the chief medical officer of Moderna, Tal Zaks, answered burning questions from the public in a webinar hosted by STAT. Here's an edited and condensed summary of his answers.
1) When will a vaccine become available?
We expect to have data in early summer about the antibody levels from our mRNA vaccine. At the same time, we can measure the antibody levels of people who have had the disease, and we should be able to measure the ability of those antibodies to prevent disease.
We will not yet know if the mRNA vaccine works to prevent disease, but we could soon talk about a potential for benefit. We don't yet know about risk. We're about to expand our safety data in Phase II.
In the summer, there is an expectation that we will be launching pivotal trials, in collaboration with government agencies that are helping fund the research. The trials would be launched with the vaccine vs. a placebo with the goal of establishing: How many cases can we show we prevented with the vaccine?
This is determined by two factors: How big is the trial? And what's the attack rate in the population we vaccinate? The challenge will be to vaccinate in the areas where the risk of infection is still high in the coming months, and we're able to vaccinate and demonstrate fewer infections compared to a placebo. If the disease is happening faster in a given area, you will be able to see an outcome faster. Potentially by the end of the year, we will have the data to say if the vaccine works.
Will that be enough for regulatory approval? The main question is: When will we cross the threshold for the anticipated benefit of a presumed vaccine to be worth the risk?
There is a distinction between approval for those who need it most, like the elderly. Their unmet need and risk/benefit is not the same as it is for younger adults.
My private opinion: I don't think it's a one-size-fits-all. It will be a more measured stance.
2) Can you speed up the testing process with challenge studies, where volunteers willingly get infected?
It's a great question and I applaud the people who ask it and I applaud those signing up to do it. I'm not sure I am a huge fan, for both practical and ethical reasons. The devil is in the details. A challenge study has to show us a vaccine can prevent not just infection but prevent disease. Otherwise, how do I know the dose in the challenge study is the right dose? If you take 100 young people, 90 of them will get mild or no disease. Ten may end up in hospital and one in the ICU.
Also, the timeline. Can it let you skip Phase II of large efficacy trial? The reality for us is that we are about to start Phase II anyway. It would be months before a challenge trial could be designed. And ethically: everybody agrees there is a risk that is not zero of having very serious disease. To justify the risk, we have to be sure the benefit is worth it - that it actually shrunk the timeline. To just give us another data point, I find it hard to accept.
This technology allows us to scale up manufacturing and production.
3) What was seen preclinically in the animal models with Moderna's mRNA vaccines?
We have taken vaccines using our technology against eight different viruses, including two flu strains. In every case, in the preclinical model, we showed we could prevent disease, and when we got to antibody levels, we got the data we wanted to see. In doses of 25-100 micrograms, that usually ends up being a sweet spot where we see an effect. It's a good place as to the expectation of what we will see in Phase I trials.
4) Why is Moderna pursuing an mRNA virus instead of a traditional inactivated virus or recombinant one? This is an untried technology.
First, speed matters in a pandemic. If you have tech that can move much quicker, that makes a difference. The reason we have broken world records is that we have invested time and effort to be ready. We're starting from a platform where it's all based on synthetic biology.
Second, it's fundamental biology - we do not need to make an elaborate vaccine or stick a new virus in an old virus, or try to make a neutralizing but not binding virus. Our technology is basically mimicking the virus. All life works on making proteins through RNA. We have a biological advantage by teaching the immune system to do the right thing.
Third, this technology allows us to scale up manufacturing and production. We as a company have always seen this ahead of us. We invested in our own manufacturing facility two years ago. We have already envisioned scale up on two dimensions. Lot size and vaccines. Vaccines is the easier piece of it. If everybody gets 100 micrograms, it's not a heck of a lot. Prior to COVID, our lead program was a CMV (Cytomegalovirus) vaccine. We had envisioned launching Phase III next year. We had been already well on the path to scale up when COVID-19 caught us by surprise. This would be millions and millions of doses, but the train tracks have been laid.
5) People tend to think of vaccines as an on-off switch -- you get a vaccine and you're protected. But efficacy can be low or high (like the flu vs. measles vaccines). How good is good enough here for protection, and could we need several doses?
Probably around 50-60 percent efficacy is good enough for preventing a significant amount of disease and decreasing the R0. We will aim higher, but it's hard to estimate what degree of efficacy to prepare for until we do the trial. (For comparison, the average flu vaccine efficacy is around 50 percent.)
We anticipate a prime boost. If our immune system has never seen a virus, you can show you're getting to a certain antibody level and then remind the immune system (with another dose). A prime boost is optimal.
My only two competitors are the virus and the clock.
6) How would mutations affect a vaccine?
Coronaviruses tend to mutate the least compared to other viruses but it's entirely possible that it mutates. The report this week about those projected mutations on the spike protein have not been predicted to alter the critical antibodies.
As we scale up manufacturing, the ability to plug in a new genetic sequence and get a new vaccine out there will be very rapid.
For flu vaccine, we don't prove efficacy every year. If we get to the same place with an mRNA vaccine, we will just change the sequence and come out with a new vaccine. The path to approval would be much faster if we leverage the totality of efficacy data like we do for flu.
7) Will there be more than one vaccine and how will they be made available?
I hope so, I don't know. The path to making these available will go through a public-private partnership. It's not your typical commercial way of deploying a vaccine. But my only two competitors are the virus and the clock. We need everybody to be successful.
Kira Peikoff was the editor-in-chief of Leaps.org from 2017 to 2021. As a journalist, her work has appeared in The New York Times, Newsweek, Nautilus, Popular Mechanics, The New York Academy of Sciences, and other outlets. She is also the author of four suspense novels that explore controversial issues arising from scientific innovation: Living Proof, No Time to Die, Die Again Tomorrow, and Mother Knows Best. Peikoff holds a B.A. in Journalism from New York University and an M.S. in Bioethics from Columbia University. She lives in New Jersey with her husband and two young sons. Follow her on Twitter @KiraPeikoff.
Meet the Scientists on the Frontlines of Protecting Humanity from a Man-Made Pathogen
From left: Jean Peccoud, Randall Murch, and Neeraj Rao.
Jean Peccoud wasn't expecting an email from the FBI. He definitely wasn't expecting the agency to invite him to a meeting. "My reaction was, 'What did I do wrong to be on the FBI watch list?'" he recalls.
You use those blueprints for white-hat research—which is, indeed, why the open blueprints exist—or you can do the same for a black-hat attack.
He didn't know what the feds could possibly want from him. "I was mostly scared at this point," he says. "I was deeply disturbed by the whole thing."
But he decided to go anyway, and when he traveled to San Francisco for the 2008 gathering, the reason for the e-vite became clear: The FBI was reaching out to researchers like him—scientists interested in synthetic biology—in anticipation of the potential nefarious uses of this technology. "The whole purpose of the meeting was, 'Let's start talking to each other before we actually need to talk to each other,'" says Peccoud, now a professor of chemical and biological engineering at Colorado State University. "'And let's make sure next time you get an email from the FBI, you don't freak out."
Synthetic biology—which Peccoud defines as "the application of engineering methods to biological systems"—holds great power, and with that (as always) comes great responsibility. When you can synthesize genetic material in a lab, you can create new ways of diagnosing and treating people, and even new food ingredients. But you can also "print" the genetic sequence of a virus or virulent bacterium.
And while it's not easy, it's also not as hard as it could be, in part because dangerous sequences have publicly available blueprints. You use those blueprints for white-hat research—which is, indeed, why the open blueprints exist—or you can do the same for a black-hat attack. You could synthesize a dangerous pathogen's code on purpose, or you could unwittingly do so because someone tampered with your digital instructions. Ordering synthetic genes for viral sequences, says Peccoud, would likely be more difficult today than it was a decade ago.
"There is more awareness of the industry, and they are taking this more seriously," he says. "There is no specific regulation, though."
Trying to lock down the interconnected machines that enable synthetic biology, secure its lab processes, and keep dangerous pathogens out of the hands of bad actors is part of a relatively new field: cyberbiosecurity, whose name Peccoud and colleagues introduced in a 2018 paper.
Biological threats feel especially acute right now, during the ongoing pandemic. COVID-19 is a natural pathogen -- not one engineered in a lab. But future outbreaks could start from a bug nature didn't build, if the wrong people get ahold of the right genetic sequences, and put them in the right sequence. Securing the equipment and processes that make synthetic biology possible -- so that doesn't happen -- is part of why the field of cyberbiosecurity was born.
The Origin Story
It is perhaps no coincidence that the FBI pinged Peccoud when it did: soon after a journalist ordered a sequence of smallpox DNA and wrote, for The Guardian, about how easy it was. "That was not good press for anybody," says Peccoud. Previously, in 2002, the Pentagon had funded SUNY Stonybrook researchers to try something similar: They ordered bits of polio DNA piecemeal and, over the course of three years, strung them together.
Although many years have passed since those early gotchas, the current patchwork of regulations still wouldn't necessarily prevent someone from pulling similar tricks now, and the technological systems that synthetic biology runs on are more intertwined — and so perhaps more hackable — than ever. Researchers like Peccoud are working to bring awareness to those potential problems, to promote accountability, and to provide early-detection tools that would catch the whiff of a rotten act before it became one.
Peccoud notes that if someone wants to get access to a specific pathogen, it is probably easier to collect it from the environment or take it from a biodefense lab than to whip it up synthetically. "However, people could use genetic databases to design a system that combines different genes in a way that would make them dangerous together without each of the components being dangerous on its own," he says. "This would be much more difficult to detect."
After his meeting with the FBI, Peccoud grew more interested in these sorts of security questions. So he was paying attention when, in 2010, the Department of Health and Human Services — now helping manage the response to COVID-19 — created guidance for how to screen synthetic biology orders, to make sure suppliers didn't accidentally send bad actors the sequences that make up bad genomes.
Guidance is nice, Peccoud thought, but it's just words. He wanted to turn those words into action: into a computer program. "I didn't know if it was something you can run on a desktop or if you need a supercomputer to run it," he says. So, one summer, he tasked a team of student researchers with poring over the sentences and turning them into scripts. "I let the FBI know," he says, having both learned his lesson and wanting to get in on the game.
Peccoud later joined forces with Randall Murch, a former FBI agent and current Virginia Tech professor, and a team of colleagues from both Virginia Tech and the University of Nebraska-Lincoln, on a prototype project for the Department of Defense. They went into a lab at the University of Nebraska at Lincoln and assessed all its cyberbio-vulnerabilities. The lab develops and produces prototype vaccines, therapeutics, and prophylactic components — exactly the kind of place that you always, and especially right now, want to keep secure.
"We were creating wiki of all these nasty things."
The team found dozens of Achilles' heels, and put them in a private report. Not long after that project, the two and their colleagues wrote the paper that first used the term "cyberbiosecurity." A second paper, led by Murch, came out five months later and provided a proposed definition and more comprehensive perspective on cyberbiosecurity. But although it's now a buzzword, it's the definition, not the jargon, that matters. "Frankly, I don't really care if they call it cyberbiosecurity," says Murch. Call it what you want: Just pay attention to its tenets.
A Database of Scary Sequences
Peccoud and Murch, of course, aren't the only ones working to screen sequences and secure devices. At the nonprofit Battelle Memorial Institute in Columbus, Ohio, for instance, scientists are working on solutions that balance the openness inherent to science and the closure that can stop bad stuff. "There's a challenge there that you want to enable research but you want to make sure that what people are ordering is safe," says the organization's Neeraj Rao.
Rao can't talk about the work Battelle does for the spy agency IARPA, the Intelligence Advanced Research Projects Activity, on a project called Fun GCAT, which aims to use computational tools to deep-screen gene-sequence orders to see if they pose a threat. It can, though, talk about a twin-type internal project: ThreatSEQ (pronounced, of course, "threat seek").
The project started when "a government customer" (as usual, no one will say which) asked Battelle to curate a list of dangerous toxins and pathogens, and their genetic sequences. The researchers even started tagging sequences according to their function — like whether a particular sequence is involved in a germ's virulence or toxicity. That helps if someone is trying to use synthetic biology not to gin up a yawn-inducing old bug but to engineer a totally new one. "How do you essentially predict what the function of a novel sequence is?" says Rao. You look at what other, similar bits of code do.
"We were creating wiki of all these nasty things," says Rao. As they were working, they realized that DNA manufacturers could potentially scan in sequences that people ordered, run them against the database, and see if anything scary matched up. Kind of like that plagiarism software your college professors used.
Battelle began offering their screening capability, as ThreatSEQ. When customers -- like, currently, Twist Bioscience -- throw their sequences in, and get a report back, the manufacturers make the final decision about whether to fulfill a flagged order — whether, in the analogy, to give an F for plagiarism. After all, legitimate researchers do legitimately need to have DNA from legitimately bad organisms.
"Maybe it's the CDC," says Rao. "If things check out, oftentimes [the manufacturers] will fulfill the order." If it's your aggrieved uncle seeking the virulent pathogen, maybe not. But ultimately, no one is stopping the manufacturers from doing so.
Beyond that kind of tampering, though, cyberbiosecurity also includes keeping a lockdown on the machines that make the genetic sequences. "Somebody now doesn't need physical access to infrastructure to tamper with it," says Rao. So it needs the same cyber protections as other internet-connected devices.
Scientists are also now using DNA to store data — encoding information in its bases, rather than into a hard drive. To download the data, you sequence the DNA and read it back into a computer. But if you think like a bad guy, you'd realize that a bad guy could then, for instance, insert a computer virus into the genetic code, and when the researcher went to nab her data, her desktop would crash or infect the others on the network.
Something like that actually happened in 2017 at the USENIX security symposium, an annual programming conference: Researchers from the University of Washington encoded malware into DNA, and when the gene sequencer assembled the DNA, it corrupted the sequencer's software, then the computer that controlled it.
"This vulnerability could be just the opening an adversary needs to compromise an organization's systems," Inspirion Biosciences' J. Craig Reed and Nicolas Dunaway wrote in a paper for Frontiers in Bioengineering and Biotechnology, included in an e-book that Murch edited called Mapping the Cyberbiosecurity Enterprise.
Where We Go From Here
So what to do about all this? That's hard to say, in part because we don't know how big a current problem any of it poses. As noted in Mapping the Cyberbiosecurity Enterprise, "Information about private sector infrastructure vulnerabilities or data breaches is protected from public release by the Protected Critical Infrastructure Information (PCII) Program," if the privateers share the information with the government. "Government sector vulnerabilities or data breaches," meanwhile, "are rarely shared with the public."
"What I think is encouraging right now is the fact that we're even having this discussion."
The regulations that could rein in problems aren't as robust as many would like them to be, and much good behavior is technically voluntary — although guidelines and best practices do exist from organizations like the International Gene Synthesis Consortium and the National Institute of Standards and Technology.
Rao thinks it would be smart if grant-giving agencies like the National Institutes of Health and the National Science Foundation required any scientists who took their money to work with manufacturing companies that screen sequences. But he also still thinks we're on our way to being ahead of the curve, in terms of preventing print-your-own bioproblems: "What I think is encouraging right now is the fact that we're even having this discussion," says Rao.
Peccoud, for his part, has worked to keep such conversations going, including by doing training for the FBI and planning a workshop for students in which they imagine and work to guard against the malicious use of their research. But actually, Peccoud believes that human error, flawed lab processes, and mislabeled samples might be bigger threats than the outside ones. "Way too often, I think that people think of security as, 'Oh, there is a bad guy going after me,' and the main thing you should be worried about is yourself and errors," he says.
Murch thinks we're only at the beginning of understanding where our weak points are, and how many times they've been bruised. Decreasing those contusions, though, won't just take more secure systems. "The answer won't be technical only," he says. It'll be social, political, policy-related, and economic — a cultural revolution all its own.