I’ve often thought of drug development as a fierce struggle between good and evil. Most of us who work in drug development are driven by the potential for “good” which are the effects of drugs that improve the lives of patients. There is a bit of self-service in that intent since most of us will eventually be patients if we’re not already. Also and as I wrote in an earlier post, there is a “good” that is the opportunity for the developer and its stockholders to benefit economically from a successfully marketed drug.
The “evils” are the undesirable or toxic effects of drugs that undermine their opportunity for market approval or a patient’s ability to tolerate them enough to gain their benefit. Generally, drugs that cause severe and irreversible harm to patients don’t make it to market. There are a lot of caveats to that statement and I won’t go into them now but toxicity is a bad thing and significantly contributes to development attrition. Relatedly, not all toxicity is severe and irreversible but can still be a challenge for patients. I’m thinking of the all-too-common CNS effects of nausea, dizziness, anxiety, and sleep disruption that plague so many marketed drugs.
The fundamental aim of drug development is to design and develop a medicine that has a balance between its benefits and its potential harms where the benefits or “good” outweigh the harms or “evil”. As most recognize, getting that balance is indeed a struggle and one we lose most of the time.
By the numbers
Interestingly, that struggle is quantifiable in many ways. From the perspective of “good”, there is the dose needed to see benefit. There is the proportion of patients that might actually experience that benefit. That number is usually defined during clinical testing but often changes if a drug receives market approval and gets used in a much broader patient population. Additionally, the benefit itself can be quantifiable as increased life expectancy, time to return to health, delay in progression-free survival, decrease in healthcare costs, decrease in lost days worked, etc.
Countering quantifiable benefit is the dose where intolerable side effects emerge. Sometimes those side effects take time to emerge so there’s also a number associated with that. As with beneficial effects, there is the proportion of patients who will suffer those effects severely enough that they can’t tolerate the drug. As with the beneficial effects, that number and even the effects themselves can change with wider use of the drug.
There's the amount of time in development it takes before you're confident that an acceptable balance of good and evil is achievable. In a best-case scenario, our early preclinical studies tell us whether we have the right balance or not informing the decision to keep going or not. Even then, you’re looking at years of development before we get to those studies. In a worst-case scenario, we don’t find out that the balance isn’t right until Phase III clinical trials right before we hoped to submit a request for market approval to regulators. That’s a 10-12 year, $2+ billion worst-case scenario.
The cost of that time can be substantial. Like everything else in life, drug development continues to increase in cost. That cost increases exponentially as you progress in development. Preclinical costs are significantly greater than discovery costs. Clinical development costs are enormous. The longer you put resources into a candidate that isn’t going to have the right balance of good and evil, the greater the costs- including the lost opportunity to work on a candidate that might have had a better chance of success. This particular number is one that drives the “fast fail” culture I’ve mentioned in a previous post.
Eroding margins of safety
A typical scenario in drug development is a candidate drug with a target dose for efficacy progresses to a portfolio of preclinical safety assessment studies of increasing breadth and duration. The most definitive safety assessment studies are conducted in animals. Maximum tolerated doses are quickly identified as thresholds of acute survivability. Target organ toxicities are identified and characterized by their severity, adversity, and reversibility. Adverse effects define No Observed Adverse Effect Levels (NOAEL). “Margins of safety” are defined as the difference between the dose at which efficacy is expected and the dose at which adverse effects are identified. It’s not uncommon for those margins to decrease as candidates progress through safety assessment. A sufficient margin (determined by the toxicity of concern) and a sensitive monitoring strategy may support progression into clinical trials in human patients where the real answers are revealed. Often, some of those real answers are that it takes a higher dose to get a beneficial effect than was expected. Unfortunately, as our preclinical models can be poor models for clinical trial patients, clinical trial patients can be poor models for the broader patient population accessed after market approval.
The good and evil struggle of drug development and the mathematical conundrum it represents might be represented by Figure 1.
Figure 1. The dose required to get efficacious benefit often increases as development progresses. Contrastingly, the dose at which harmful effects are seen often decreases as development progresses. The consequence of these usual eventualities is loss of a margin of safety and development termination.
Cracking the math problem
There are a number of ways to mitigate this conundrum.
Increasing the potency of the candidate drug for its pharmacological target might lower the dose needed to get a beneficial effect. Unfortunately, that can also increase the likelihood of an adverse response if that adverse response is a consequence of hitting the intended or a closely related target.
Prolonging exposure might also allow a lower dose by providing more time for the drug to circulate and engage the pharmacological target. Relatedly, that prolonged exposure might also exacerbate or increase the likelihood of an adverse effect.
The “evil” effects of the drug might be mitigated by decreasing the potency of the drug for the mechanistic target of toxicity. Of course, that requires us to know what that target is which often isn’t the case. You would also want to preserve the “good” which can be tricky when you start modifying a drug’s design. Similarly, decreasing the level of exposure or dose has the potential trade-off of compromising benefit.
A common clinical approach to managing undesirable side effects of drugs that do get market approval is co-administering a mitigating therapy (e.g. giving an antiemetic to a cancer patient being treated with a nausea-inducing chemotherapeutic). That said, administering a drug to counteract the negative effects of another drug seems like a vicious circle but it’s a vicious circle that is not uncommon.
Another approach to mitigating the consequences of having a development-limiting imbalance in good vs. evil is to decrease the time and cost of coming to that conclusion. Unfortunately for the developer and for patients, those imbalances are often not discovered until Phase II and III clinical trials at great fiscal cost to the developer and potential risk to the patients in those trials. This particular approach is popular with developers and a common use for new technologies. Novel in vitro and in silico modeling systems are often used as screens before initiating more expensive animal studies. Unfortunately, this practice often contributes to the “additive innovation” that I wrote about in my last post- i.e. increasing the effort on the front end to avoid wasted effort on the back. That is a very tenuous value proposition as these “screens” may aid in lead selection but more often produce “hypotheses” that we test in the animal studies we routinely do anyway.
Focus on predictivity
Since our drug discovery and development culture is one where we already maximize pharmacological potency and spend little time trying to understand the mechanisms of toxicity, our best bet might be to improve the quantitative predictivity of our early assessments. This approach might also be more tractable as we’ve concentrated efforts to use more mechanistic non-animal modeling approaches early in development.
Improving the predictivity of our early preclinical assessments will require us to get better at modeling two key biological events before we progress to animal studies- progression of a host’s response to a drug over time and identifying the point (defined by dose and/or time) at which an adaptive or not harmful response by that host transitions to a maladaptive or harmful one (i.e. toxicity). For the former, we need to be able to model long term trajectories from short term exposures in contrast to usual practice where we model long term outcomes in long duration studies in animals. For the latter, we need to define toxicity in terms other than morphologic tissue-level or organ-level functional outcomes that are the primary focus of our usual animal safety assessment studies (arguably, this is also what we do in human clinical trials).
I think this is all doable with our fundamental understanding of the pathogenesis of injury, our vast experience modeling drug-induced toxicity (80+ years since the Food, Drug, and Cosmetics Act of 1938), and the power of our current computational capabilities. The opportunity is further supported by recent progress in the development of in vivo-relevant modeling systems and the promise of predictive AI.
Relevant tools
Physiologically-relevant in vitro modeling systems populated with human cells can increase the translational relevance of our pre-animal assessments. These modeling systems are not more translationally-relevant because they are populated with human cells but because they better mimic in vivo tissue architecture and physiology, are substrate for mechanistic endpoints, and are often more temporally stable than traditional 2D culture systems. Accordingly, we have the opportunity to model in vivo-relevant human biology more mechanistically and continuously over spans of time that can provide insights into temporal progression without having to model the ultimate target organ outcome. That’s a mouthful so let me say it in the context of an example. I can model a drug candidate’s temporally progressive disruption of mitochondrial function in a human cardiomyocyte in vitro that is functioning like it does in vivo. From that, I can predict an eventual decrease in the contractile capacity of that cardiomyocyte, the eventual death of that cardiomyocyte, and heart failure without having to wait until the cell dies or the heart fails. I could say the same thing about lots of other cellular targets of toxicity. Essentially, that’s operationalizing the Adverse Outcome Pathway concept (PMID: 29682628) that has been promoted as a framework for developing non-animal approaches.
Now, I’ve just represented those biological events as linear and, to some extent, binomial (e.g. mitochondria functioning or not, cell contracting or not, cell alive or dead, etc.) which isn’t the way biology really happens. It is much more complex, variable, and dynamic than that. That’s where the power of our computational capabilities will come to bare. There are lots of quantitative relationships between mitochondrial function represented as ATP production, calcium transients in and out of the cell as a primary effector of cardiomyocyte contraction, change in contractile function, induction of mitochondrial injury and repair pathways, the oxidative stress that would result from mitochondrial insufficiency, and even protein degradation pathways as a consequence of that oxidative stress. A mere mortal can’t quantitatively integrate all of those processes and relate them to known outcomes like an AI algorithm might. It’s the predictive AI that will translate my in vitro numbers into an in vivo prediction.
I’m getting into the gory details of how to mitigate our drug development math problem and have probably lost you doing it. The point of this really is that we need to think more about how to leverage the “math” of our biology of interest and the quantitative balance in good vs. evil we’re trying to achieve to become more predictive in our earliest assessments. We have the tools and know what we’re looking for. We just need to design, build, and test a novel paradigm focusing less on the morphology of apical endpoints and more on the quantitative progression of events that we know will end in a bad place.