Sort of long answer.

In, say, civil or aerospace engineering, science is understood well enough to allow your building or airplane to be modelled and tested using computer modelling, CAD software, FEM and CFD algorithms and so on. You can design a house or an aircraft without ever building a single physical model, and it will stand (or fly), 99 times out of 100. It is oversimplification to a degree, but sufficiently close approximation.

No such thing exists in biology, pharmaceutics, biotech and so on. The accuracy of computer models and simulation is not sufficient to produce results with single-digit percent accuracy for any metrics, hence long and complex Phase I-II-III trials. Maybe 1 out of 100 candidate drugs works.

Why? Because we do not have the same level of understanding for biological systems as we do for buildings or aircraft, or software. Amount of information is much larger, complexity is far greater, enzymes and cell signalling network make biochemistry extremely non-linear. This makes the problem space vast. It is practically untapped domain and it can eat any amount of computational power and biologists, data scientists and software devs (manpower-wise).

Any incremental improvements in simulation, modelling and interpretation of biological system behaviour will generate downstream improvements in medicine, pharma, biotech. But general-purpose LLM AIs are not that useful in biology, you need more specialized solutions to improve both accuracy and performance of large number of algorithms that have tremendous computational complexity: computational chemistry, molecular dynamics, genomics->proteomics->interactomics->metabolomics (all of that for just intra-cellular behaviour - tissues, organism and organisms are multiple orders of magnitude harder).

But fundamentally it is a problem of missing software to better model biological systems (AI or non-AI). Once created, such a solution will enable large amount of very big breakthroughs in almost every biology-connected discipline.