— We're living in a fascinating era where, despite all the advancements in medical science and the exponential growth in technologies for gathering and processing biomedical data, the cost of developing new drugs continues to soar, reaching hundreds of millions or even billions of dollars for each newly registered medication. The main reason for the high failure rate in expensive clinical trials is the lack of efficacy: new drugs often work well in animal models of human diseases, but fail to do so in humans.

The biotechnology and pharmaceutical industries are seeking ways to solve this problem, trying a variety of approaches. One of the most sensible hypotheses is that research on model organisms has limited value and that it would be better to have the ability to study disease mechanisms in humans.

As fantastical as it may sound, human medical research is becoming a reality. Digital technologies in medicine have led to the existence of hundreds of millions of electronic medical records. Tens of millions of people have been genotyped. This means we can "peek" at the results of the "clinical trials" that nature conducts on us and for us. The fact is, humans are not identical; we differ from each other genetically and, depending on this, when exposed to different medical situations, have significantly different chances of falling ill, healing, or even dying as a result of the development of various diseases. Some mutations in our DNA protect us, and therefore, can be mimicked by new drugs.

— There's probably an endless number of models, or even different types of models, that can be used to look for correlations in data. The advent of big data has given rise to so-called generative models—a new kind of algorithm capable of not just learning or predicting specific values (like blood parameters), but also determining the properties of the distributions of observed parameters. This is a much more powerful capability, as such algorithms can not only describe data but also, if necessary, generate new data that is indistinguishable from existing data. In this sense, the models become "digital avatars" of the individual.

If such capabilities can be achieved in models that utilize a small number of parameters, and if those parameters turn out to be interpretable—meaning they are understandable to experts and have medical or biological significance—then we can gain a better understanding of how various factors influence each other in determining human health. In this case, the models do more than just identify rare correlations in the data; they help us understand how the human body functions in different medical scenarios. This opens the door for conducting experiments that target specific physiological parameters, allowing us to predict the long-term outcomes of these interventions on chronic diseases.

— During the training phase, models can develop unexpected properties. Take ChatGPT as an example; a neural network initially designed to complete sentences turned out to be an engaging conversationalist capable of even passing some professional exams. In any case, models are tested on data they haven't seen during the training process.

— It's important to understand that very large datasets often contain numerous errors. Most of the time, the data you're dealing with was not specifically collected for your project but is combined with data from various other providers. Even setting aside data quality issues, a hundred million electronic medical records are a drop in the bucket compared to the total global population of around eight billion people and growing. This means that many hypotheses you might form based on even the largest available datasets could turn out to be unreliable when trying to generalize your findings to the entire human population.

There are ways to validate your conclusions within the same dataset you have. Additionally, in good research, there's always the opportunity to test the most intriguing assumptions with independent data (electronic medical records from other countries or regions within the same country, genetics from European and non-European cohorts).

In developed countries, there's a growing recognition of the need to collect high-quality medical data from significantly large groups of the population—hence the creation of national cohorts. Access to such data is the best method for model verification.

However, even the most elegant hypothesis remains just that—a hypothesis—until confirmed through animal experiments (pre-clinical trials) and human trials (clinical).

— Modern machine learning is a highly developed area of applied mathematics with its own culture, accomplishments, methodology, and experts. Applying these methods in adjacent sciences like drug development, biology, and medicine requires interdisciplinary skills. Accordingly, scientists from various fields need to work on these models together. A whole "cocktail" of methods is applied, including those from the physical and engineering sciences, to ensure the models are not just accurate in their predictions but also interpretable.

— At some point, senior colleagues explained to me that public attention (and therefore prestige, funding, etc.) towards physical sciences was related to defense. But in the early 21st century, other tasks came to the forefront. Developed countries underwent a demographic transition— a period of sharp decline in birth rates while life expectancy increased. This resulted in an exponential rise in the risks of chronic diseases. It became clear that in the developed world, there will soon be no one left to fight, but many will need medical treatment. To corroborate this, one can compare the arms market, which had shrunk to about half a trillion dollars, with the pharmaceutical industry market, which was at one and a half trillion dollars. It is not hard to see how biotech is becoming increasingly attractive. Colossal investments were made, and specialists from other fields followed. Speaking of physicists, we brought along skills in analyzing phenomena over time. The same principles that were once used for predicting eclipses or calculating the orbits of interplanetary spacecraft can be applied to understanding the mechanisms of aging and age-related diseases.

— We are interested in what are called longitudinal medical data, i.e., data collected multiple times over a person's life (like the results of their annual check-ups). The collection of such data on a large scale is impossible without digitization, so the very opportunity to analyze them appeared only after the digital transition in medicine and the emergence of electronic medical records. There's still not as much longitudinal data as we'd like.

Medical histories are often stored in hospital databases or national healthcare systems and cover millions of patients worldwide.There is also biobank data—these include electronic medical records (who got sick when, what medications were taken, blood tests, medical imaging, and molecular data like protein and metabolite concentrations in the blood, genetics).

— The straightforward answer to your question is: we work with what we have and compare it to biobanks, where the data is particularly well-organized. Metaphysically speaking, one of the marvels that modern machine learning shows us is that sometimes the volume of data solves the question of data quality. "Quantity is a quality all its own", the saying goes..

— Medical data is personal and therefore very sensitive. There are many legal (data protection laws) and technical (data protection system certification, depersonalization) restrictions. We don't collect the data ourselves; we gain or rent access to already collected and depersonalized data, either in recognition of scientific merits or for money. However, we are also responsible for data security on our end and must go through certification procedures and so on.

— To date, there's no approved drug specifically for aging. A recent review in Nature states that since human genome decoding began in 2003, only 40 genetic associations have been transformed into treatments for 40 diseases, mostly against the rare ones (36 out 40). Most human diseases are not caused by a dysfunction of a single gene. The human genome has tens of thousands of genes, and less than a thousand of them are targets for approved drugs. Moreover, fewer than ten drugs with new mechanisms of action are registered each year. Can you imagine how many years it will take to fully understand human molecular biology? Artificial intelligence offers a way to significantly accelerate research.

— Our statement is a bit more subtle: we assert that human longevity is limited if the sole focus of medical science is on fighting specific diseases rather than the primary cause of serious diseases—aging. This paper has greatly helped us draw attention to the issue, including interest from pharmaceutical companies and investors. We only discuss our results after they've been published in peer-reviewed journals, so keep an eye out for updates; they're coming soon.

— The advent of AI will lead to massive societal and technological changes, many of which we cannot yet envision. It's not wise to assume that future technologies will solely address the problems we currently deem important. New technologies will offer us new possibilities and present new challenges. For instance, the advancement of direct brain-computer interface technologies via AI could lead to the dissolution of human identity—we could upload our memories and experiences to computers, share thoughts and experiences with others. Personal boundaries will inevitably shift not just in space but also in time, allowing us to relive the best or most interesting moments of people from different continents or even eras. Our physical bodies may live much longer in the future. This does not mean, however, that notions of lifespan, self-identity would not be transformed in such a way that the whole concept of immortality may end up having the whole new meaning. Technology provides means and these questions for us to discover and for our children to find out.

— There are limitations on what we can disclose beyond what's in the press release. What can be said is that contrary to popular belief, Pfizer and other pharmaceutical companies are extremely interested in aging-related issues. The focus currently is more on the link between aging mechanisms and specific diseases. The reason is that age is the primary risk factor for most significant diseases. On the other hand, some age-related changes may turn out to be irreversible, and understanding this, the pharmaceutical industry is looking to differentiate between irreversible changes and those that can be potentially targeted by drugs. The more we understand about the mechanisms of aging and its relationship with diseases, the more effective treatments will become available on pharmacy shelves.

We hope our work brings together experts in gerontology, artificial intelligence, machine learning, complex systems science, and the world's best specialists in drug development, testing, and sales. All of this will be required to develop an anti-aging drug.