Now, when you get sick, every possible compound known to humankind can be administered to an infinite set of you’s who are living your reality and are equally as sick or unhealthy. With time travel we’ll have your clones instantly skip a couple years and see which compound resulted in the healthiest version of you. That compound is administered to the ‘real’ you and boom, disease cured.

It’s quite a functional system with no reliance on traditional biology. There are also no test tubes, pipettes, imaging machines or even scientists necessary. In this reality, those things might still be helpful in figuring out why a particular drug would cure a disease, but they are unnecessary to the process of finding a that drug. In fact, we can abstract away all the biological stuff — as long as a manipulation produces the desired output in the intended system (i.e. the sick individual), it’s a drug. With infinite copies of the intended system, we can assess the entire space of possible manipulations. And with time travel, we can instantaneously know which manipulation works best. The work of figuring out the reasons behind compound A curing person B is best left to academics who study biology for its own sake, not to industrial processes charged with eliminating disease in humans.

Of course, the real problem is that none of these three elements are even remotely possible under the current laws of physics, but the exercise is still a useful thought experiment. If this reality is considered an ideal system under which all disease is eliminated, then we must create technologies that bring us closer to this future. It follows that innovation in biotechnology must accomplish one of three things:

1. Manipulate Life: Grow the set of practical manipulations of the human body. This includes discovering new chemical compounds, but also editing tools, like CRISPR, and even prosthetics. (ultimate goal: infinity, the set of possibilities should keep growing rapidly)
2. Simulate Life: Build scalable representations of the human body that are increasingly representative of its biological complexity, environmental interactions, and interspecies variation. (ultimate goal: infinite perfect copies of every human being living their exact reality)
3. Fast-Forward Life: Reduce the time required to assess how a manipulation from (1) functionally alters the living system from (2). This includes predictive biomarkers and early-stage diagnostic tools. (ultimate goal: instantaneous assessment of future health, i.e. time travel? )

Generally, most biotech innovation today is happening along the first aim. Pharmaceutical startups are growing faster than ever, each investigating the applicability of a new set of chemical compounds to medicine. Companies like Neuralink are developing new methods to manipulate our brain circuitry, while technologies like CRISPR allow us to manipulate the genome. Others, like Recursion are applying innovative approaches to build bigger proprietary banks, faster.

What about aims 2 and 3? Well, we don’t actually have an infinite number of humans to expend. So instead, we construct biological mappings between observed disease and molecular mechanisms. We’ll study the brains of patients with mental trouble and reduce their troubles to a set of common molecular interactions, which we’ll call Alzheimer’s disease. We assume that if we can adjust those interactions in just the right way, we can reduce or eliminate the disease. Then, we’ll spend years finding manipulations that can displace disease-related interactions in test tubes, before conducting studies to observe how they affect increasingly complex systems from cells to mice.

If we find something that works in mice/rats (typically the most complex biological systems we have access to at scale) we’ll run a clinical trial. 9 times out of 10 that drug will fail to alleviate the disease in humans. All that time and money spent studying fundamental biology, building up a basic understanding of the drug’s mechanics, testing it increasingly complex systems, and then actually putting it in real people — goes wasted.

So in a way we do simulate life, but the empirical data shows us that we’ve just built really bad simulations. We use models that reduce the complexities of human life to simpler systems which are available to us at scale. It’s true that scale is crucial to approach the ‘infinite copies’ requirement of Aim 2; you can’t make 1000 clones of a human being, but you can have 1000 test tubes. What we ultimately need is a system that is closer to human life with regards to complexity that is available at the same, or even greater scale than test tubes.

That might mean building pseudo-humans in the lab: juxtaposing exact chemical copies of organs to build systems, and then juxtaposing those systems to simulate life. Indeed doing this would perhaps be the ultimate accomplishment for the field of synthetic biology — life built on the lab bench! I wonder what Mary Shelly would say about that. Frankenstein’s monster aside, the real issue here is just how far we are from ever being able to do this. We face the same fundamental problem as before — we have an extremely limited understanding of all of the various processes that make up life, and we can only reconstruct what we know — and we can’t even reconstruct most of that yet.

At Biosense, we take the opposite approach. Instead accounting for all these various mechanisms as building blocks to reconstruct the human, we like to be extremely reductive with biology. Every body is just another function with a number of inputs — perhaps including genome, environment, medical manipulations, current state, etc. — and a numerous outputs, in the form of the various biological signals and behaviors. To replicate a human life, you don’t have to reconstruct it from ground up, you just have to create a black box where, given all the same inputs, it would provide all the same outputs as the human you’re trying to replicate.

Constructing this black box is not a trivial thing. With enough manipulations for inputs, and measured outputs, it is possible to build a graph that algorithmically connects the two (i.e. the black box). But the space of inputs and outputs is immensely wide; and because we can’t apply inputs to human bodies and measure outputs at scale, we cannot even begin to traverse this space. Considering only chemical compounds as inputs, and digital health signals or self-reported symptoms as outputs are ways to narrow the both spaces with useful results. Even so, directly traversing these limited spaces is still quite impractical.

But everyday, scientists are manipulating and measuring systems just one level of complexity beneath the human body — animals. The measurements they make are already used to inform clinical trials, the only problem is that relationship between animal data and human data is extremely weak, as we see based on empirical success rates of clinical trials. We’ve spent the last year building and distributing a rapidly scalable digital health platform for mice. Now, we can generate the precise, granular data we need to systematically construct strong mathematical relationships between digital signals from mice and humans.

Once we build this mapping, we’ll be able to use real animal data that relates inputs to outputs to generate synthetic human data relating the same inputs to outputs in humans. We can use this synthetic data to train the world’s first digital organism — the ‘black box’. Eventually, we want to be able to create a personalized digital organism for every human being, that we adjust based on environmental data from their surroundings. ‘Infinite perfect copies of infinite humans living their exact reality’.

And what about Aim 3 — time travel? Yeah, someone else might have to get busy working on that. In any case, if we’re successful at turn all this biology into mathematical mappings, then processing speed will become the rate-limiting factor to getting results. That’s something we have a lot more control over than time itself.

All of this sounds a little crazy, and that’s because it is. We began by selling measurement tools that generate a lot of data from mice; the same data that your Apple watch might measure from you. What we realized is that if we could just connect the two datasets, we could take advantage of the scale and flexibility animals offer us, but interpret results as if they came from a human. There are obviously tons of limitations and complexities that we’re just beginning to discover and investigate. This is not an easy problem — remember, we are trying to simulate human biology. Right now all we can do is generate clean, structured, beautifully diverse data from living organisms. It’s chaotic and disorganized, but it’s still something, and as our good friend Mary Shelly reminds us “invention, it must be humbly admitted, does not consist in creating out of void, but out of chaos”. So we’re digging through the chaos, drawing lines and testing hypotheses, and someday very soon, we’ll be able digitally reconstruct life itself. How cool is that.

Interesting stuff? Continue the conversation! Get in touch with me at rohan@biosense.ai

Reimagining our world’s most crucial development pipeline

We all know how important user testing is in the development of a consumer product, but what happens when companies try to develop products that could hurt or kill humans if improperly designed? It’s a catch-22 situation thats plagued some of the world’s most crucial industries, including the one tasked with our lives — health care.

Fortunately, this industry has found a tried and tested solution: rodents. Particularly, researchers around the world — from the best universities to the most powerful pharmaceutical corporations — continue to use mice as initial beta-testers of novel therapeutics due to their availability, size, low cost, fast reproduction rate, and ease of handling. Mice share 84% of their genome with the human genome on average, making them prime animal models of inherited disease, and encouraging researchers make use of them in a whole host of studies. Investigators rely on mice to study addiction, disease prognosis, psychology, sleep — you name it — without putting human lives at risk. In fact, between 20 and 30 million mice are used in biological and medical research annually in the United States alone — that’s a lot of cheese.

These creatures allow for crucial, early-stage research findings to be made in vivo — Latin for “within a living organism” — as opposed to in vitro, or within a test tube. Countless modern medical findings, techniques, and treatments can be credited to having their roots in studies involving mice and other rodents.

Of particular interest are preclinical rodent studies, which involve the development of new therapeutics. Typically these studies begin with a fundamental biological observation; perhaps that — like my colleague and I observed when researching lung cancer treatment (find out more here) — a particular compound is likely to selectively target cancer cells. This observation is tested in vitro, first in test tube assays and then in cell line studies involving islolated human cells. If the results are promising, the treatment in question can be administered to rodents in order to understand the details behind its mechanism of action, toxicity, solubility, lifespan, and other critical metrics, before it enters clinical trials with real human beings. Research concerning the translation of fundamental breakthroughs to novel treatments— taking solutions from bench to bedside, so to speak — is published under the guise of translational research.

But translational research has a problem —it doesn’t really work; at least, it doesn’t work efficiently, effectively, and rigorously, like science is expected to. It’s a problem thats giving rise to the phrase “lost in translation” , and spurred the U.S. National Institutes of Health to begin a whole host of initiatives including a center dedicated to “advancing translational science” (NCATS) and the Clinical and Translational Science Awards (CTSA) Program. And although rodents certainly can’t take all the credit for this problem, their central role in any translational study means they certainly contribute to it.

The fundamental assumption lying at the core of the translational science is if a particular treatment works in preclinical studies (i.e. mouse models), then it should work in clinical studies (i.e. human models). In the vast majority — over 80% — of cases, this assumption doesn't hold; a treatment that works in preclinical trials fails to work in humans. All scientists expect some significant deviations when putting theory into practice — and this especially true when moving from one species to another — but an 80% false positive rate is not just significant, it’s overwhelming. Look at it this way — although about 500 treatments have been reported as effective in animal models of stroke, only 2 work in humans. All in all, these failures cost companies and universities conducting translational research upwards of 10 billion dollars annually — and that translates directly to higher treatment prices for all of us. But its worse than that because we don’t even know how many potential treatments are “false-negatives” — meaning they don’t work in rodents but they could work in humans — and if that number is anything close to 80%, then there are a lot of lives being lost by virtue of real treatments never coming to market.

In an article titled “Preclinical research: Make mouse studies work” (published by Nature in March 2014) Steve Perrin, the chief scientific officer at the ALS Therapy Development Institute in Cambridge, MA, offers some insight into the reasons for these statistics, which he collectively calls “one of the most uncomfortable truths in translational research”:

“Animal models of disease are frequently condemned as poor predictors of whether an experimental drug can become an effective treatment. Often, though, the real reason is that the preclinical experiments were not rigorously designed”

The article (available here) goes on to provide numerous examples of poor experimental design and establish a causal relationship between these examples and lackluster results. From these analyses he extrapolates a series of guidlines that are supported by “nearly a decade of validation work” and “should reduce the number of false positives in preclinical studies and so prevent unwarranted clinical trials”. These include:

1. “rigorously assessing animals’ physical and biochemical traits in terms of human disease”
2. “characterizing when disease symptoms and death occur and being alert to unexpected variation”
3. “creating a mathematical model to aid experimental design, including how many mice must be included in a study.”

The types of characterization experiments that complete these guidelines are, according to Perrin and others in the preclinical research community “expensive, time-consuming and will not, in themselves, lead to new treatments” Nonetheless, when he asserts that “without this upfront investment, financial resources for clinical trials are being wasted and lives are being lost” — he’s not wrong.

The reason these types of “models” — linking experimental practice and animal traits to human disease prognosis as called for by Perrin and other reviewers — are not established is multi-fold, but can be understood from two different angles. The first is that, like Perrin acknowledges, its not traditionally the job of a biologist other life scientist to take part in this kind of mathematical modeling. That’s not to say mathematics doesn’t play a crucial role in life-sciences research, but rather that most biological laboratories don’t exist with the end-goal of establishing mathematical relationships between animal traits and human disease. Instead, their funding and valuable time is spent — perhaps more wisely — on investigating fundamental biological pathways or pursuing disease treatment.

I say more wisely, because the fact is, life scientists are probably not the ideal group to pursue model-building related research anyways. Not to over-generalize, but its clearly the kind of project that’s best suited to computer scientists, data scientists, mathematicians, and statisticians, who attend conferences and pursue publications that seek to understand the most advanced methods to build exactly these kinds of mathematical relationships. In fact, I’d bet that a machine learning researcher would be much more interested in using cutting-edge data processing techniques to develop a link between animal behaviors and human disease than a cancer biologist, who’d prefer to understand why the disease occurs, and how we could stop it.

But it goes deeper than a barrier between computational solutions and biological problems; mathematical modeling requires understanding the relationships between data, that is quantitive data. Simply put, to do math, researchers need numbers. But the majority of preclinical methods don’t output numbers; instead, most conclusions are made from observations alone. That’s because the kinds of things researchers extrapolate conclusions from — behaviors such as running, resting, rearing, and seizures instances — need to be seen, not measured with an instrument. But who decides what constitutes a behavior? — the individual observer associated with a particular cage, leading to observational bias both between observers within a study, and across independent studies. There are efforts to standardize what constitutes a particular behavior using computer vision techniques, but applying these methods requires massive amounts of infrastructure, including multiple cameras per cage, and the associated cabling. Not only this, but they limit the kinds of experiments researchers can perform, making universal adoption highly unlikely, and moving the field nowhere closer to an uncontested standard.

And the little data that can be quantitive — metrics like temperature, respiration rate, and heart rate — is difficult to obtain. Collecting this kind of data requires invasive telemetry devices, which are expensive, require time-consuming surgeries before and after an experiment, and remain targets of criticism from ethics committees and animal rights advocates. Moreover, recent studies have shown they can alter thermoregulatory behaviors, introducing some level of bias into even numerical data. Even if researchers had access to numerical, un-biased data, extrapolating mathematical models from this data would require large amounts of it stored in a central location, all collected using the same standardized methods, something completely infeasible today.

So what will it take to create these kinds of models that are poised to eliminate “one of the most uncomfortable truths in translational research” , and spur a revolution that will create unprecedented opportunities in preclinical science? Nothing short of immense innovation in both technology and business practice. Researchers require a new kind of core technology, one that gives them automated access to all of the kinds of metrics they want, without limiting the kinds of studies they can perform. It must be noninvasive, nonintrusive, quantitive, and low-cost, and it must be adopted quickly and universally. Even a field fragmented by numerous such technologies will be incapable of following the best scientific practices, because different reporting formats and technical bias will make it impossible to independently repeat experiments. Moreover, not only is mathematical modeling best suited to data collected from a common instrument, but the models generated by data from one instrument will be best applicable to the same instrument, not some other device. For these reasons, the field must use this technology universally, and look towards it undisputed standard; this also means there needs to be little to no friction when it comes to adopting the technology, a requirement that even today’s most advanced systems fail to meet.

The data collected by the technology must be stored on a universally accessible platform — for all studies, even those that may not produce ideal results. This needs to happen automatically and seamlessly, at no additional trouble to the scientist. Further, a program must exist to incentivize and enable data scientists, computer scientists, and mathematicians — the true subject matter experts — to develop the kinds of links that researchers need from the data stored on this platform, so that biologists and life scientists are assisted in, not distracted from their core fields of research.

When it comes to these making these solutions a reality, it’s a long, dark, and rocky road forward — but that doesn’t mean its impossible. It’s exactly what my team and I are working on at FMB Technologies, where we’re developing and testing the kinds of technologies needed to fundamentally better preclinical science. Researchers, executives, patients — and maybe even rodents — everywhere are counting us to reimagine what is perhaps the most crucial development pipeline humanity has known, and we’re confident that very soon, we, along with the rest of the world, will be basking in the light at the end of this tunnel.

The Rodent Revolution was originally published in FMB Technologies on Medium, where people are continuing the conversation by highlighting and responding to this story.