When an algorithm promises “I’m 90% reliable,” it’s easy to trust it. But in fields like drug discovery, that figure can be a trap. A new study shows that marginal conformal prediction — a popular uncertainty quantification technique meant to put an honest number on model reliability — achieves global coverage while almost entirely abandoning the rare classes, the very ones that often matter most.

The researchers tested the approach on four datasets. With a target coverage of 90%, marginal conformal prediction hits that mark overall. Yet when you drill down into minority classes, the numbers plummet: coverage falls to 64.8% on blood-brain-barrier penetration, and to a mere 4.2% on clinical-trial toxicity — effectively rendering the model blind to dangerous compounds. The flaw isn’t tied to any single model: it reappears whether you use a random forest, a graph neural network, or a frozen chemical language model (a specialized LLM).

The mechanism is familiar to statisticians but rarely discussed in deployment practice. Marginal conformal prediction guarantees an average success probability across all predictions, without per-class distinction. On imbalanced datasets — where a class like “toxic” appears in less than 5% of samples — the system can sacrifice the rare class to preserve accuracy on the majority, all while boasting excellent aggregate coverage.

The proposed fix is a class-conditional variant: instead of a single alpha threshold for all predictions, separate coverage constraints are enforced for each class. This lets the minority class regain the statistical protection it deserves. However, the correction isn’t free: it requires more granular calibration, and in high-throughput virtual screening it can translate into more conformity checks per compound.

This is where infrastructure enters the picture. Many pharmaceutical companies run such models on on-premise clusters to safeguard intellectual property and comply with regulations like GDPR. Adding per-class conditionality increases inference load and, at scale, affects the Total Cost of Ownership (TCO) of the hardware. Anyone evaluating on-premise deployment must now weigh a difficult trade-off: is the computational savings of marginal conformal prediction worth the risk of missing nearly 19 out of 20 toxic compounds?

The episode signals something deeper. Off-the-shelf statistical frameworks are increasingly adopted on the strength of global guarantees, without checking performance on critical subgroups. In a domain where a false negative on toxicity can have devastating consequences, the lesson is that statistical honesty must be tuned to domain priorities. Perhaps the real “honest number” isn’t the one that works on average, but the one that never lies about the class you can’t afford to ignore.