The Far-Reaching Ripple Effects of a Discredited Cancer Study

I’d look for a more complicated and nuanced explanation than that. Maybe, there are perverse incentives in scientific publishing (ever heard of publish or perish?), and insufficient checks and balances, and it’s hard to do the right thing and easy to do the wrong thing.

The more nuanced explanation gives you more traction for incremental improvement. One historically easy way to get people to do the right thing is to make the right thing easier.

People used to dump their used motor oil in the nearest stream. Now every auto shop is required to accept used motor oil and recycle it. Doing the right thing is so easy that you’d have to be crazy to do the wrong thing.

600 other papers cited it. Hundreds of other labs, professors, postdocs, and grad students had failed projects based on extending the research. Undoubtedly many of those people had their careers end in this hypercompetitive space due to this one paper.

The impact didn't destroy society or anything, it damaged the field. It made people think "wow that's interesting, tumors have unique microbiomes". the article says:

Meanwhile, research groups that, according to Salzberg, relied on the study’s data are dealing with the fallout, as are the journals that published the work. The publishers or journal editors of eight studies have begun reviewing the papers, according to spokespeople at Springer Nature, Frontiers, Wiley and the American Association for Cancer Research. The corresponding authors of these studies didn’t respond to requests for comment.

Looking at the analysis critique, yes, the original paper had an awful analysis. It smacks of a wet lab or clinical administrator who knows nothing about bioinformatics, having work done by say a grad student who themselves are not experienced in the field but clever enough to stitch together a nice sounding pipeline and write a Nature paper about it, and perhaps a bit cocky. This, I have seen before first hand. The long and short of it is, they used a big database of all bacteria genomes, and guess what, if a human grinds in a wet lab sequencing a bacteria genome, sometimes a bit of human DNA slips into the bacteria genome assembly, particularly if this was done some 50,000 times for 50,000 bacteria. So their analysis made no consideration to removing alignments to the human (or any other contaminant) genomes. So maybe their system says "aha! this bacteria is Staph Aureus!" but actually the Staph Aureus sequence has a fragment of human genome because the sequencing project therein had light contamination. It's absolutely self evident that if they were sequencing bacteria DNA off tumors, they should have controlled for human DNA contamination, given that the tumor the bacteria is sitting on, that tumor has a lot of human DNA.

Depends how we implement it. We could make bad science way faster. ChatGPT paper spam is already a thing, and their authors don't care about quality

What tasks being automated will make research more epistemologically sound?

That wold require automating reproductibility or requestin and analysing datasets that are not provided.