Increasing Clinical Trial Success Rates
Better animal models, organoids, keeping organs alive after death, causal inference, and AI-driven human data mining
A recent study showed that 54% of phase 3 clinical trials fail. For neurodegenerative disorders, a lot more clinical trials fail than average. Most trials fail because of lack of efficacy and we can do definitely do better. Imagine a world in which 90% of clinical trials succeeded? We might push a lot harder to get these expensive ($B) clinical trials run because each one would all but promise a new treatment or even a new cure. The question is how do we do better?
I would argue that we need better evidence that a drug is going to work before a clinical trial is initiated. I think we can design smarter approaches that hit a more predictive fail/no-fail inflection for a smaller price tag pre-clinically. This will allow more shots on goal and ultimately cost less because clinical trials will be more de-risked before initiation.
Let’s take a look at what has gone wrong that has caused the high failure rate for clinical trials. Let’s take the example of Alzheimer’s disease clinical trials. Why have these trials either failed or been only moderately successful? A big problem is that we still do not know whether the proteins that have been targeted with drugs- mainly amyloid plaques or other species of amyloid protein- actually cause Alzheimer’s disease. You can’t treat or cure a disease if you are eliminating the wrong protein. Some experts believe that we just haven’t found the right species of amyloid to target and others, including myself, believe that amyloid isn’t causal at all.
Another problem is that we don’t know the timing that the intervention must start to be successful. Can we reverse Alzheimer’s with the drug we are creating or only prevent it from getting worse? Do we need to start with people when they first start to have memory problems or is late stage disease treatable? So how did we spend so much money, >$42B, on clinical trials for drugs that target proteins that might not even cause the disease in time frames/patient segments that we aren’t clear about?
Part of this is certainly political and there is also a story about data fraud in key papers. Here I will focus on how the science has contributed to high clinical trial failure rates and how we can do better. To understand how to improve clinical trial success rates, we first need to understand how pre-clinical drug validation works. We rely on animal models, usually mice, for validation because by being genetically identical and identically housed, they eliminate sources of noise. We take animals with a disease and give one group the drug we are testing (treatment group) and we give another identical group, placebo (control group). Then, if we have seen significant disease improvement or cure in the treatment group compared to the control group, the drug is validated. Seems reasonable right? Yes, except that mice are very different from people and often don’t get the same diseases. This is particularly problematic for neurodegenerative diseases, which mice don’t normally experience. They don’t get dementia. They don’t get plaque build up in their brains. The changes that happen with age within neurons in mice are often in the opposite direction of humans. So how do you establish causation in mice if they don’t get Alzheimer’s disease?
Scientists have artificially created diseases in mice in various ways. In the case of Alzheimer’s disease, they have engineered mice that produce large amounts of the human amyloid protein in their brains. Those mice, unsurprisingly, have memory problems. They then gave those mice a drug that removes that human protein and voilà! The mice were cured of their memory problems. Pretty neat right? Sounds like causation? I would argue, not. If you flood a mouse brain with a large amount of any human protein they would likely have memory problems. Reversing that artificial problem created by scientists seems quite likely to improve that mouse’s memory. That doesn’t mean that amyloid is causing Alzheimer’s in humans or that the anti-amyloid drug would work to treat the human disease.
So how do we test drugs better before clinical trials? I think there are a few ways including 1. using better animal models 2. using human systems like organoids 3. using AI and causal inference in existing human datasets.
BETTER ANIMAL MODELS
Mice have been used historically as universal models for human disease because they are relatively cheap and we can breed and raise them at Universities. So the tools to manipulate animals- genetic engineering etc.- and the resources to breed and obtain tissue have all been built for mice. In a research setting, this is ok for many diseases, especially cancer, which mice do get. This might be part of the reason why we have seen more success with cancer clinical trials than neuro clinical trials. But for many diseases mice are really bad models.
I propose that we need to match the animal model to the disease that we are modeling and invest more money into developing tools and programs for species other than mice.
DOGS
Note: no dogs would be harmed in this scenario (I’m a dog lover), these would be clinical trials like with people, not animal testing.
For neurodegenerative disorders, dogs are actually much more similar to people than mice. They get dementia, plaques, and tangles like people do. Companion dogs also share our environment- they drink the same water, breath the same air, and are exposed to the same things that we are. A dog clinical trial costs ~$5M. A human clinical trial costs ~$1B. What if 90% of drugs that were successful in dogs, were successful in humans? That would be a significant cost savings for human clinical trials. And we would have drugs for our furry friends in the meantime. This is the approach that Loyal and the Dog Aging Project are taking and I think will not just apply to dog longevity in the future but will become standard in many companies.
LOWER ORDER ANIMALS WITH HUMAN SIMILARITY FOR AN INDICATION
In general, mice are more similar genetically to humans than say, fruit flies. But what if some lower order animals were more similar to humans than mice for a particular indication? This would be great because they would be better models of a disease while being cheaper to study. How do we know which lower animal model to choose? We need to design a framework for this. I would suggest that this framework should include proof that the biology of interest- be it a disease or hallmark of aging- naturally exists in the animal we are interested in. For NeuroAge’s purposes this would start by demonstrating that whole genome transcriptome changes with age in the organism’s brain match those in human to a high degree. Does the animal normally get neurodegeneration? If not, it’s not a good model. Do fruit flies fit this criteria? How about water fleas? This is something that I think is worth investigating.
NON-HUMAN PRIMATES
The closest relatives to humans are non-human primates. We know that their brains age in very similar ways to people. While having a Chimpanzee colony might not be realistic for an early stage biotech startup, having a Tree Shrew colony might be. Tree Shrews are tiny (the size of a squirrel) and the cost to house them is more than mice but a lot less than dogs. Having non-human primate data would de-risk clinical trials a lot.
PRECLINICAL TESTING IN HUMAN TISSUES- ORGANOIDS & KEEPING ORGANS ALIVE
What could be better for de-risking clinical trials in humans than testing in human organs before the trial? We are very interested in this and technology in the last five years is showing promise. The FDA passed legislation in the last year that animal testing is no longer required for human clinical trials. I think we are still about 3-5 years out from this being a reality but imagine a world in which high quality data in human organs was brought to the table before a clinical trial was run? And no animals would be harmed in testing? Hold onto your hats because things are about to get very exciting.
ORGANOIDS
Organoids are laboratory grown human organs, where drug efficacy and toxicity can be tested. Imagine this, we grow a human brain on a chip that has Alzheimer’s pathology and then we test drugs on it to show that they can remove the pathology and make the brain young again. This is exactly what is being developed by companies like Emulate. There are some kinks to be worked out still with this technology. The structure and cell types of these lab grown brains and other organs are not perfectly like humans yet but we are close.
KEEPING ORGANS AND CELLS ALIVE AFTER DEATH
So what if instead of growing organs we kept them alive soon after death? This would bypass the issue of Organoids not yet being perfectly like human organs. There are several companies working on this in dogs and pigs. In the case of dogs, these furry friends have a terminal illnesses and their owners were putting them down. The owners then agree to donate their brains to science after they have passed. These companies have the intention of moving on to human brains after they succeed with dogs and pigs. So far these companies have kept brains alive for a few days. This is almost long enough to start testing drugs.
Another remarkable strategy is to culture human neurons directly from the recently deceased. Instead of creating immortalized neuron cells lines or deriving neurons from stem cells, we can instead just grow them in dishes after harvesting them from human brains of people who have died within the last 24 hrs.
USING HUMAN DATA- GENETIC CAUSAL INFERENCE AND MINING HUMAN HEALTH RECORDS
What could be better to establish causation and de-risk clinical trials than using human data itself? We can do just that in multiple different ways. Here I highlight two very promising avenues- genetic causal inference and patient data mining.
GENETIC CAUSAL INFERENCE
Traditionally scientists have determined what is genetically causing a disease by testing every genetic difference between people who have a disease vs. don’t have a disease using Genome Wide Association Studies (GWAS). Then a statistical p-value is used to determine what is noise and what might actually be causing disease. There are a few issues with this. The first issue is (as the name implies) that these are association studies and not necessarily causal. So we don’t know that these genetic differences actually cause the disease or if a nearby gene is actually the smoking gun. Mathematical models have been created to allow us to get a lot closer to what is causal without creating animal models. Welcome to Mendelian Randomization. I’m not going to get into this except to say that we can use the genetics of people not only to test association but also to test causation. Why is this not in every GWAS and a requirement for drug discovery? Good question. My guess is that outside of computer scientists and the cutting edge of bioinformatics, the wider world hasn’t caught onto this yet. The left hand (computer scientists) and the right hand (drug hunters and biologists) need to start talking.
MINING HUMAN HEALTH RECORDS
What if your pre-clinical evidence for your already approved drug comes directly from patient records? Let’s say that there is an approved drug for diabetes that you think might also work for leukemia. With enough big data and access to patient records, you could look for leukemia patients who also happen to be on the diabetes drug vs. those that are not to see if being on the diabetes drug is associating with improved outcomes in Leukemia. You would have to be very careful to control for confounders and this would be correlative and not causal. But it would provide evidence in addition to your animal data that the drug works before running a clinical trial. It also can generate new leads for drug repurposing if done in a high throughput manner for many drugs.
WHY THE FUTURE OF CLINICAL TRIALS IS BRIGHT
We can do so much better with clinical trial success rates particularly for neurodegenerative disorders. We can de-risk for cheaper, and with the added bonus of harming fewer animals. All of this is happening now or will be in the next 3-5 years. We need to invest in cross-pollination between startups that are moving new preclinical systems forward. We need to move beyond mouse models and into the AI-driven, furry friend by our side, organ on a chip, and back from the dead future.