AI is more than a computer program that is capable of human-like language comprehension and communication. It is its data assimilation and ability to recognize patterns within billions of bits of information that makes it a powerful tool for the advancement of science.
Imagine you are a scientist researching a new cure for malignant lymphomas. Conventional means of doing research involve a comprehensive search of all papers previously published on the subject. In addition, the scientist can look at the newest papers posted on preprint servers. These latter ones would have yet to be peer-reviewed. A third search might compile a list of eminent scientists in the field with contact information for you to access as needed. But now you can ask an LLM to do all of those queries and deliver the information in less time than it will take you to read this article.
Schmidt anticipates that much of science can be automated creating what he calls “self-driving labs.” These robot-driven research facilities will conduct experiments which at the beginning will be devised by humans, but later by the AIs to produce publishable research results.
Schmidt doesn’t believe this will sideline scientists, but rather will open opportunities for more to join in including citizen scientists who today are already contributing to advancements in many fields.
Even peer review could become an activity better done by an AI than a human. A dispassionate, impartial assessment of research can be difficult to obtain because often the reviewers are known to the author who may favour their opinions. But an AI wouldn’t be burdened by human connections in looking at new hypotheses.
Already we are seeing AIs excel at some scientific tasks. One is protein folding. Back in 2013, I posted an article about FoldIt, a game created by scientists at the University of Washington to help understand protein structures and their role in diseases. It invited humans to play the game and come up with new ways to fold proteins and by doing this help find potential cures for diseases. In 2018, an AI named DeepMind took on protein folding leading to the creation of a new tool called AlphaFold. Tasked in competition with humans it solved 25 of 43 protein-folding problems compared to the second-place human finisher with only three. Since AlphaFold’s introduction, researchers at the University of Washington have come up with even more powerful AI protein folders. The latest is RoseTTAFold which concurrently can solve multiple protein-folding problems.
Reproducibility is another challenge for scientists. In a 2016 paper published in Nature, it quotes a survey of 1,576 researchers with more than 70% admitting they failed to reproduce published results from their experiments. I get in my inbox periodic reports from a site called Retraction Watch, which exposes bad science and scientists. But it only catches a small number of papers reporting unreproducible experiments. The vast majority slip by. And even the papers Retraction Watch catches continue to get cited by scientists doing new research. An AI tool would catch the unreproducible with Schmidt arguing it would create “a greater trust in science.”
Schmidt states that “AI is such a powerful tool because it allows humans to accomplish more with less: less time, less education, less equipment. But these capabilities make it a dangerous weapon in the wrong hands.” That’s why he is calling for well-informed regulation of AI tools and their use for the benefit of science and humanity. Exploiting AI’s potential and recognizing the potential perils of using it, Schmidt states will allow “us to combine information in novel ways and bring creativity and progress in the sciences to new heights.”
]]>The way a protein folds determines its function. Proteins are chains of amino acids that fold into intricate shapes. They are encoded by DNA with the genes generating the recipes for how they fold. If they fold just right they execute the tasks for which they were designed. When they don’t there are negative consequences.
Proteins are large and complex molecules responsible for nearly every function in our bodies. Their unique structures have been a puzzle for scientists until recently. And their form and function are critical to how they achieve their roles. A good example of unique sets of folds are antibody proteins used to identify and tag viruses or bacterial invaders. When properly folded these proteins form Y-shaped hooks that latch on to invaders identifying them for the T-cells in their search and destroy missions.
The following illustrations depict how the protein-folding mechanism works.
When a virus like COVID-19 first infects a host it attacks cells and uses the protein-making machinery within them to make copies of its own proteins which then help in viral replication. This commandeering of the cell is the mechanism by which the virus spreads within every infected person.
Being able to predict how a protein folds can help determine how the virus functions within an infected cell and that knowledge can then be used to devise therapeutical countermeasures to impede the commandeering of the protein-making machinery. What AlphaFold does is predict the protein structure the virus creates in the absence of certainty by other research means. Right now AlphaFold is studying the virus responsible for SARS, a relative of COVID-19. A previous tool developed by DeepMind called the CASP13 system was able to accurately predict that virus’ protein folds and helped virologists and the scientific community to understand how to mount a counterattack.
DeepMind, the creators of AlphaFold, use a neural network AI. Neural networks simulate the interconnected cells of human brains to learn things. to detect patterns in data, and to make decisions similar to the ones we humans make. A neural network is a self-learning AI.
With AlphaFold the neural network has learned to predict a proposed protein structure by looking at millions of examples within the protein landscape. The neural network plays with chains of amino acids to create substructures of its own design rather than an entire protein. It then, over time, makes incremental improvements to the substructures it invents slowly building up a knowledge base of what larger assembled structures should look like. In the end it replicate how proteins fold in nature.
The virus when it enters a healthy cell carries its own encoded protein which interacts with the cell’s protein-making machinery to make more copies. The copying proteins are polymerases, long chains of RNA enzymes. The goal of researchers trying to find a cure for COVID-19 is to block the RNA polymerases so the virus cannot replicate. The challenge is developing a drug that specifically targets the RNA polymerase without harming the host cell.
In both previous coronaviral epidemics, SARS and MERS, attempts to create a safe drug yielded some positive results in laboratory animal studies. Remdesivir, a drug originally developed to fight the Ebola virus, although less effective on it, proved better in laboratory studies treating infected animals.
In February, the National Institutes of Health started a clinical COVID-19 trial using the drug at the University of Nebraska Medical Center in Omaha. Remdesvir has been undergoing clinical trials in China as well. The results from both efforts are pending.
]]>
Around 2012, a computer scientist-turned-biophysicist, Alex Zhavoronkov, noticed that AI was getting better at image, voice, and text recognition. He knew that all three tasks shared a critical commonality, massive datasets available to use to train the AI. By 2014, Zhavoronkov was using the technology for discovering new drugs.
Zhavoronkov had heard about a new AI technique known as generative adversarial networks (GANs). How does it work? By pitting two neural networks against each other, with minimal instructions the technology could produce novel outcomes. Before this researchers were using GANs to do things like design new objects or create one-of-a-kind, fake human faces. It was Zhavoronkov who saw pharmacological possibilities. He figured GANs would allow researchers to verbally describe drug attributes in terms such as “The compound should inhibit protein X at concentration Y with minimal side effects in humans,” and then let the AI construct an appropriate molecule from scratch.
Zhavoronkov set up Insilico Medicine on the campus of Johns Hopkins University in Baltimore, Maryland, and over three years using GANs was able to develop a drug discovery engine that sifts through millions of data samples and determines the biological signature and characteristics of specific diseases. It can then identify the most promising treatment targets and generate molecules perfectly suited for them. States Zhavoronkov, “The result is an explosion in potential drug targets and a much more efficient testing process … AI allows us to do with fifty people what a typical drug company does with five thousand.”
Right now, the GANS is being used to find new drugs for cancer, aging, fibrosis, Parkinson’s, Alzheimer’s, ALS, diabetes, and many others. The first drug coming to market is a treatment for hair loss slated to start Phase I trials in 2020.
Insilico is also in the early stages of using its tools to predict the outcomes of clinical trials in advance. If successful, the technique will enable researchers to shorten the time and money spent in the traditional testing process.
For those of you unfamiliar with the role proteins play in our bodies I refer you to an article posted on this site back in 2013 and a game called FoldIt, created by researchers at the University of Washington, which uses the collective effort of crowdsourcing to develop new protein folds. And you may not be aware that CRISPR/Cas9, the gene-editing tool taking genetic engineering mainstream, is in fact, a tool that uses a protein to excise bad genes and replace them with healthy ones. So proteins are a hot subject in the world of pharmacology today and scientists are using AI to identify new drug targets that bind with them.
Proteins consist of amino acids that fold in different ways to take on different functions. A protein containing a hundred amino acids, rather small by body standards, produces a googol-cubed-worth of potential shapes. (That’s a one followed by three hundred zeroes.)
Back in 1994, a biannual protein-folding competition was created with little success until 2018 when the creators of DeepMind turned neural networks loose on the problem. DeepMind took on the task of determining the most likely distance between a protein’s base pairs and the angles of their chemical bonds. They called their end-product, AlphaFold. In its first entry into the competition, it solved 25 of 43 protein-folding problems compared to the second-place finisher with only three.
How does AlphaFold work? It uses two neural networks, one focused on predicting distances between pairs of specific amino acids, and a second evaluating the possible angles of the chemical bonds between them. The two return a score estimating likely structures that match a compilation of known protein fragments, and the generation of completely new protein structures.
The potential implications of AI-accelerated protein folding for new pharma cannot be underestimated and what it could mean for treating patients with Huntington’s, ALS, myotonic dystrophy, and other diseases.
Another theatre of war for improved drugs is in the realm of drug delivery. Converging exponential technologies are paving the way for massive implications in both human health and industry shifts. The aforementioned CRISPR stands to revolutionize synthetic biology and the treatment of genetically-linked diseases. Researchers are now demonstrating how this gene-editing tool can be applied to create materials that shape-shift on command. Think of the potential of materials that can dissolve instantaneously when stimulated, releasing a specified drug in a highly targeted location.
Then there are nanobots, micro-machines capable of providing targeted drug delivery. In a recent review of medical micro- and nanorobotics, the lead authors from the University of Texas at Austin and the University of California, San Diego found numerous successful tests of in vivo operations using medical micro- and nanorobots.
Inefficient, slow-to-innovate, and risk-averse industries will be disrupted in the years ahead and Big Pharma is an area worth watching right now. Two areas, in particular, are getting headlines, advances in addressing longevity, and in disease prevention. The aforementioned Insilico is one of the companies leading the way.
The convergence of massive datasets, growing computational power, the rise of quantum computing, and innovations in AI, are bringing us closer to a world in which personalized drugs delivered directly to specified targets will become the standard of care. Rejuvenation biotechnology will be commercially available sooner than you think.
Zhavoronkov sees the timeline at “maybe 20 years … a reasonable horizon for tangible rejuvenational biotechnology.”
How might any of us use an extra 20 or more healthy years in our lives? What impact would that have on the world?
