The rigour nobody asked for
Lately I’ve been preparing a small paper for submission, a rather naive piece of work from when I’d only just started my postgraduate studies. At first the process gave me a faint sense of achievement: something I’d put off for years was finally going to see the light of day. Yet the longer I spent looking at it, the more I began to doubt myself.
The methods it relies on aren’t really sound. I had my misgivings when I first built the pipeline, and I took them to my supervisor, only to be given a “lecture”: “Your biggest problem is that you lack confidence and you’re always judging yourself. If you don’t believe in your own work, how are you ever going to publish or present it?” An inaccurate model, though, can’t be wished away with a bit of self-reassurance, or even self-deception. My doubts didn’t come out of nowhere. Through repeated use, I had found that the model’s effects were unstable, or heavily swayed by certain references.
Since leaving the rigorous bioinformatics training of my undergraduate degree, I’ve gradually come to see that the way bioinformatics is actually used in research bears little resemblance to how I was taught to approach problems. The prevailing attitude is “if it works, it’ll do”, ideally with results that fit the trend you were hoping for. At first I was openly scornful of this. Over time, though, I realised that what looks “more accurate” from a data-analysis standpoint is still miles away from the real biological world. The relentless pursuit of accuracy in data, methods, or models may well be a futile exercise.
The only thing the scientific community will buy is what can be experimentally validated. It’s an entirely outcome-driven standard. Whatever method you used, and whatever principles lie behind it, counts for next to nothing in the real world. All anyone wants is for the one molecule you picked out of a vast sea of candidates to turn out to be real. For that one molecule, false positive and false negative rates are just background figures nobody pays attention to.
These contradictions leave me at a loss. If I abandon the rigour I was trained in, then as a bioinformatician I’m really just reducing myself to a tool. If I hold on to my original way of thinking and try to understand and apply every method as thoroughly as possible, I only end up tying myself in knots. But I’ve never been able to ignore these flaws in the methodology. They’re like thorns lodged under my skin, a constant ache I can’t get away from. Many times I’ve wanted to tear up the paper under submission and start again with a sounder methodology, and each time my supervisor has vetoed the idea. After all, the experiments are done and the conclusions are in. Rebuilding a supposedly better analysis pipeline would merely be dressing up a conclusion that already exists. At best it would give me some peace of mind; in any worldly sense it would be worthless.
Two or three years ago, when I first started using AI tools, I realised that bioinformatics would eventually revert to being a tool, and should no longer be pursued as a discipline in its own right. That doesn’t mean I should turn myself into a tool as well. Yet I have no idea what I actually am in scientific research. Within my own work, I can’t find where I belong.