Actually overall, in vivo animal testing on vaccines, antivirals, and antibiotics are typically quite predictive from animals to humans. These types of tests are not historically very predictive when moving from in vitro to in vivo animal studies, however. This is typically because the treatments and vaccines are structural and/or expression-based (exploit a bacterial surface protein and poke holes in the cell, use a virus to get into T-cells and inject genetic material to let them identify and kill polio, etc). (yes, oversimplified). But moving from cell culture to a living organism This doesnt always work out. If it does, then it's relatively easy to confirm the effect if it works in animals to humans (virus is gone/never infects, etc). So more antibiotics and vaccines fail from in vitro to in vivo than from animal to human data.

In treating diseases of inflammation, rheumatic disease, pain, etc., it is generally quite predictive when moving from in vitro to in vivo animal studies (e.g. is it down-regulating inflammatory transcription factor NF-kB, or not?). But it becomes much less predictive translating endpoints from animals to humans. That's, very generally, because it is difficult to assess disease measures of improvement in inflammation in animals and translate that to humans (pain, discomfort, edema, these processes may present much differently in rats versus humans, for example). Great animal data may not mean you'll get such a strong effect in human subjects, and thus the failure rate for these drugs in clinical studies is quite high compared to vaccine and antibiotic human trials, because those vaccines were shown to be ineffective earlier in development. This was a very general explanation, to be sure, and I'll dig through my archives for some papers on this and try to add links later.