Why do we continue to discover so late that certain drugs do not work, or present unacceptable risks?
Every year, billions of euros are invested in the development of drugs that will fail… often too late.
One figure alone sums up the scale of the problem: nearly 93% of drug candidates never reach the market. Even more striking, a significant proportion of these failures occur after years of development, sometimes at the most advanced stages of clinical trials. At that point, the costs were colossal, but above all, the warning signs already existed.
Based on this observation, why do we continue to discover so late that certain drugs do not work, or present unacceptable risks?
A system that validates too late, and learns too slowly
The problem is not a lack of data. It’s how we use them.
Today, the assessment of the benefit–risk ratio still relies largely on fragmented approaches: clinical trials carried out on limited populations, static analyses, late reviews. Even after marketing authorization, monitoring remains partial, with well-documented underreporting of adverse effects.
In this context, decisions are made based on successive “photographs”, often disconnected from each other. Result: weak signals go unnoticed, and major uncertainties persist until late stages.
This operation is no longer adapted to the current complexity of drug development.
The real blind spot: the absence of a dynamic vision of benefit–risk
What the accumulation of late failures reveals is not only a scientific problem, it is a problem of method.
We continue to assess benefit–risk as a fixed state, when in reality it is an evolving balance, which should be continually reassessed as new data emerges.
Between clinical trials, real-life data, safety histories and patient registers, we now have an unprecedented volume of information. However, this data remains largely under-exploited, because it is neither integrated nor analyzed on a continuous basis.
It is precisely this blind spot that fuels some of the avoidable failures.
Moving from one-off assessment to real-time monitoring
The tools now exist to change the paradigm.
Approaches based on artificial intelligence make it possible to integrate heterogeneous data sources, detect early signals and simulate different benefit–risk evolution scenarios. Where traditional methods expect late results, these models allow anticipation.
Concretely, this means that it becomes possible to:
- identify security risks earlier
- detect potential inefficiencies
- prioritize candidates with the best chance of success
- In other words, to make better decisions… sooner.
Regulators are ready, but waiting for solid evidence
Contrary to popular belief, regulatory authorities are not a brake on this development. They have already started the movement.
Initiatives at European level demonstrate a clear desire to integrate real-life data and modernize evaluation frameworks. Regulators are open to more dynamic approaches, provided they are rigorous, transparent and auditable.
The real issue is therefore less acceptance than the ability of innovators to offer tools that meet these requirements.
Reducing avoidable failures: a shared responsibility
Reducing the proportion of late failures is not an isolated technological innovation, but a collective change in practice.
Three conditions will be determining:
Reliable and explainable models, capable of gaining the trust of regulators and clinicians
Earlier dialogue, to align evidentiary expectations from the outset
Continuous evaluation frameworks, integrating the entire life cycle of the medicine
No longer accept the inevitable
Failures are part of drug development. But not all of them are inevitable.
At a time when costs are exploding, delays are lengthening and patient expectations are ever higher, continuing to discover problems late that can be detected earlier is no longer sustainable.
The question is no longer whether we can improve the benefit–risk assessment.
The question is how much longer we can afford not to do it.
