Why doesn't strong preclinical evidence guarantee a human clinical effect?

The translation from preclinical evidence to human clinical effect works reliably for some biology and less reliably for others. For receptor pharmacology at well-conserved receptors — GLP-1R, GIPR, glucagon receptor, melanocortin receptors — mouse and human are close enough that preclinical effect predicts clinical effect with reasonable reliability. This is why semaglutide, tirzepatide, retatrutide, and setmelanotide translated from preclinical to clinical successfully.

For other biology, translation is harder. Tissue-repair endpoints in rodents (BPC-157's tendon and gut work) differ from human clinical repair contexts: rodents heal faster in absolute terms, the specific injury models don't perfectly correspond to human injuries, and small sample sizes may produce effects that don't generalise. Aging endpoints (MOTS-c in aged mice) face the additional issue that mouse and human lifespan differ enough that 'aged mouse' and 'aged human' represent different biological states. Behavioural and cognitive endpoints (Selank, Semax in rat models) are the hardest because rat and human cognition differ enough that functional-clinical extrapolation is difficult.

The practical implication is that reading preclinical evidence requires knowing what kind of biology is being extrapolated. Receptor pharmacology at conserved receptors is more predictive; behavioural, aging, and tissue-repair extrapolations are less predictive. A rich preclinical evidence base with weak human evidence is not evidence that the compound doesn't work — but it is not sufficient evidence that it does, either. The honest reading is that we don't yet know at the level of confidence that a phase-3 trial would establish.