As someone with zero background in this area, here is what I tried to do in ten minutes:
I looked up Machine learning in bioinformatics on Wikipedia. Searching the page for "2021", I came across a fairly recent paper titled Specialized metabolic functions of keystone taxa sustain soil microbiome stability. In the Discussion section, one of the main themes that emerges near the end is:
Interestingly, numerous studies have revealed that nitrogen cycling taxa are consistently identified as keystone taxa across diverse ecosystems. In contrast, the function of “phosphonate and phosphinate metabolism” is less well documented. However, the most abundant Gemmatimonas genus is one of the dominant groups identified in agricultural ecosystems. ... We speculate that the abundant bacterial genera that include Nitrospira, Rhizobacter, and Burkholderia referring to the keystone functions of “nitrogen metabolism” and the abundant genera that include Gemmatimonas and Brevundimonas referring to the keystone functions of “phosphonate and phosphinate metabolism” are pivotal taxa that may contribute to soil microbiome stability.
An unproven but strongly suggested speculation! That sounds like a topic ripe for research. I wonder if we can experimentally mess with Gemmatimonas and wreck soils somehow. Before I start excitedly writing up my research proposal, though, I note that this article has been cited 45 times (as of 8 June 2022). I will have to spend the afternoon looking through those citations to make sure none of them have already worked on my brilliant idea.
In summary, here were the steps I followed:
- Locate a recent, interesting article
- Skim to the end and scan the Discussion for interesting new thoughts.
- (To be done) Check articles that cited this article to see if those thoughts had been followed up. (If they had, start over from step 1 with this article.)
While I would love to leave you with this framework and wish you success, there are two features of your situation that seem very precarious to me.
The first feature is that "machine learning in bioinformatics" is a set of methods. Unless you are in a very specialized group, you run the risk of floating adrift in the wild seas of method development with very little to show for it at the end of your PhD. In conventional research you can often know that your methods will give you the real answer, even if it sadly isn't the answer you'd like to have. In method development, you do not even know if your method is correct, so you will have to work out (1) if your method gives you the real answer and (2) why it is so much better than other conventional methods out there.
Think of carpentry: when you learn carpentry, you start by making a piece of furniture, like a chair or a table or a cabinet. You don't start by making a saw, or a hammer, or nails. That is a different, and much more specialized and difficult, field.
The second feature is that you are asking random strangers on the Internet for advice that should be coming from your supervisor. I cannot stress enough how hazardous that is. You do not know me! You cannot know for certain that the advice I am giving you is good or sound. And if you follow my advice and things go wrong, there is every reason for your supervisor to say "well, why were you asking random strangers on the Internet for advice?" The whole point of a supervisor is that a supervisor is someone who, in theory, has gone where you're going, and has some stake in your success (in that they will get to co-author the papers you eventually write, after having helped you get going). They should be obligated to give you sound, useful advice in a way that a stranger on the Internet is not.
If you do not sense that your supervisor has invested themself in your success, and is just demanding results without telling you how to obtain them, you need to fix your situation ASAP.