Human-captured. AI-enhanced. One composite image.
Thank you for reading I2I Newsletter!
If you think someone else might enjoy this newsletter, please pass it forward or they can sign up here.
Hello! Brian, Nick and Szczepan here.
In last issue, we argued that better tools matter only when they carry a clear warranty on the decision they are meant to inform. This month, that logic meets one of the hardest realities in drug development: most programs do not fail because a single category turns red. They fail because benefit, burden, and variability drift toward each other faster than teams admit, while time and cost keep compounding in the background. That is where portfolio math, trial design, translational data, and governance stop being separate conversations and become the same one. Brian quantifies the shrinking margin between therapeutic gain and tolerated harm. Nick maps the tangled causes of trial failure and asks what kinds of data actually earn their place in the evidence chain. I push the argument one step further: if we cannot specify the overlap, the trigger, and the owner, we are not managing risk, we are narrating it. Together, these three essays turn de-risking from a slogan into an operating discipline.
INSIDE THIS ISSUE
Brian Berridge | Drug Development is a Math Problem
Why the margin between benefit and harm keeps shrinking, and why better prediction has to start earlier.
Nick Kelley | 90% of clinical drug development still fails. Will AI help de-risk trials?
Why AI will matter only when it is paired with fit-for-purpose data and tied to a decision that someone can actually make.
Szczepan Baran | Your Pipeline Dies in the Overlap
A practical operating model for deciding when benefit, burden, and variability have converged enough to change what you fund next.
BYTES
Brian Berridge | Drug Development is a Math Problem
Brian starts where every program eventually gets judged: the moving boundary between therapeutic benefit and tolerated harm.
Brian recasts drug development as a quantitative balancing act, not a sequence of hopeful checkpoints. The core variable is margin of safety, the distance between the dose that delivers meaningful benefit and the dose, duration, or exposure pattern that produces intolerable harm. That margin often narrows as candidates advance, which is why late stage attrition looks surprising in real time but obvious in hindsight. His point is not that the industry needs more front-end activity for its own sake. It is that we need earlier assessments that can model how injury progresses over time and identify the point where adaptive biology becomes maladaptive. That shifts the center of gravity toward human relevant in vitro systems, mechanistic readouts, and computational approaches that can integrate complex biological trajectories into usable predictions. The implication is practical: stop waiting for apical failure states to confirm what the system has been signaling for months. Better development decisions begin when we treat toxicity as a dynamic process and build preclinical evidence to predict when the math stops working.
“The margin that matters is the one between hoped-for benefit and tolerated harm.”
Nick Kelley | 90% of clinical drug development still fails. Will AI help de-risk trials?
Nick widens the frame and shows why trial failure refuses to stay inside neat buckets.
Nick argues that efficacy, safety, commercial pressure, and operational friction rarely fail one at a time. They compound. A molecule can look strong in assay space and still collapse in the clinic because dose is constrained by toxicity, tissue exposure is uneven, target validation is incomplete, the population is poorly selected, or the trial simply cannot execute cleanly enough to reveal the truth. AI can help, but only if it is applied to the right problem. He points to real opportunity in EHR-driven planning, target validation in human populations, faster competitive intelligence, and operational risk detection at sites. At the same time, he is blunt about the frontier that remains weak: translation. In many cases, the real bottleneck is not data extraction but data generation. That is why he presses on MPS, organoids, imaging, and other fit-for-purpose sources that could expose relevant biology earlier. His bottom line is disciplined and timely: data are valuable only when they are captured, interpreted, and routed toward a concrete decision.
“AI does not de-risk a program by being impressive. It de-risks a program when the data it uses can change a real decision.”
Szczepan Baran | Your Pipeline Dies in the Overlap
From there, Szczepan turns the theme into a management rule that can survive a real gate review.
My argument is simple: most leadership teams still talk about program risk in labels, when they should be talking about overlap. A program breaks when the exposure required for benefit sits too close to the exposure, duration, or variability the system cannot tolerate. That is why I reframe failure modes through three moves: shift, squeeze, and spread. Some factors move the benefit or harm threshold. Some shrink the usable band through time, cost, or execution constraints. Others widen variability until averages stop being decision tools. Once that lens is in place, the question becomes operational. What data are we generating to narrow the overlap, and what decision changes if that data come back bad? The Three Curves Test translates that logic into practice by forcing teams to quantify benefit, burden, and variability on one page before Phase II turns uncertainty into sunk cost. The goal is not more monitoring theater. The goal is an explicit trigger, an owner, and an action.
“If you cannot name the curve, the trigger, and the owner, you are not managing uncertainty. You are financing it.”
THE TAKEAWAY
The thread running through all three essays is straightforward: de-risking improves only when evidence is generated early enough, and specifically enough, to change a real decision. Whether the issue is a narrowing safety margin, a data stream that does not earn its keep, or a Phase II plan built on vague monitoring language, the real work is the same: make uncertainty explicit before it becomes sunk cost. If these essays surface a failure mode, an early marker, or a translational blind spot you have seen firsthand, send it our way. The strongest conversations in this field begin when someone stops defending the process and starts describing what actually broke.
About I2I Newsletter
From idea to impact: candid reflections and sharp provocations on the real journey of innovation in drug development and how to separate theater from true translational change.
Read the Archives