Why faster discovery still faces slow clinical development

A computational model can propose a molecule in hours. The development programme still needs to establish that the molecule reaches the intended tissue, changes the intended biology and produces enough benefit to justify its harms.

These tasks overlap, but their timescales differ. A drug that acts on a laboratory target can fail because the target is poorly chosen, because the exposure is inadequate or because the molecule affects other biology. Increasing the supply of candidates can expose those problems more often.

Retrospective research helps identify recurring failure patterns. A 2015 analysis of candidates from four pharmaceutical companies examined relationships between molecule properties and attrition. A later study analysed genetic factors associated with why trials stopped. Neither provides a simple failure probability for a new AI-designed candidate.

The target problem

Biological datasets contain many associations. A protein can be elevated in disease because it contributes to damage, because damaged tissue releases it or because the body is attempting repair. A model trained on those associations needs further evidence to establish which intervention would help.

Human genetics can supply evidence about causal involvement. A 2024 analysis found greater clinical success for mechanisms with genetic support, with the relationship varying across therapeutic areas and development stages. This is an observational result. It supports better target selection while leaving each intervention's behaviour to be tested.

The direction of action also needs attention. A variant that reduces a protein's activity throughout life does not reproduce every consequence of giving an inhibitor after disease has developed. Timing, tissue exposure and the degree of inhibition differ.

AI can help combine evidence and propose tests. Its useful contribution can include discarding an attractive but weakly supported target before a large clinical investment.

What the trial calendar contains

The FDA describes clinical development as a sequence of studies with different purposes. Early studies establish tolerability and how the body handles a candidate. Later studies examine effects in relevant participants and build evidence sufficient to assess benefit and risk.

A trial cannot observe a year of outcomes in a month. Researchers can choose earlier measurements, but must establish what those measurements predict. Recruitment and follow-up take time even if the underlying molecule was designed rapidly.

Ageing research adds a difficult choice of outcome. Preventing several diseases over many years requires a different study from changing an ageing biomarker over twelve weeks. The shorter experiment is useful when described accurately; it cannot supply observations it never collected.

Small early studies also leave uncertainty about less common harms. If an adverse effect is uncommon, a small group can pass through a study without anyone experiencing it. Larger populations and continued observation increase the opportunity to detect it.

For some gene therapies, FDA guidance recommends risk-based long-term monitoring because delayed adverse events are possible. This follow-up should not be confused with a universal rule requiring every therapy to wait a fixed number of years for approval.

Manufacturing is another scientific problem. A programme has to make material consistently and establish that changes in production have not changed the product's relevant properties. For complex biological products, the process can affect what participants receive.

A credible acceleration claim specifies the step shortened and the comparison used. It also reports failed candidates, not only the fastest survivor. Calendar speed, financial cost and probability of a useful result are different measures.