1) Aside from significant increase in cost and experiment turn-around time, there's a more basic question: if you don't use a replicable and scalable model of the disease, how do you induce the cancer you're interested in studying? Wait around for one of ~5000 mice to develop pancreatic cancer?

2) There's already significant efforts underway to sequence many thousands of human tumors (e.g. you probably know about TCGA). Once the cost drops a bit more, then it might be feasible to sequence all tumors. However, a pathologist can still tell you quite a bit that's hard to reverse-engineer from sequencing (e.g. does the cell look like a melanocyte?). Classifying by tissue of origin (+ the ontology of cancer subtypes) does actually capture a lot of the variation between cancers. Sequencing adds a little on top of that, but surprisingly not as much as people hoped. For example, how often can you predict actionable drug sensitivity from RNAseq or WES? In my experience so far, there's only even the possibility of clinical benefit in a small number of special cases (e.g. BRAF V600E).

3) Newly diagnosed patients often have great treatment options! Do you really want to RCTs with placebo arms on stage I breast cancer patients?

4) "Go all out on the immune approach" there's been a huge boost of funding in that direction (by the NCI/NIH and private donors like Sean Parker).

"with an emphasis on developing treatments with minimal side-effects that can be given to healthy people as a preventative treatment." That's an interesting idea and I don't know who's working on it. Write it up as a grant proposal and send it to the CRI?