What would be the most efficient learning path from here to being able to analyze real public datasets (TCGA, GEO, MetaboLights)?

That entirely depends on you. There is no strict curriculum. When I started about 10 years ago, before AI existed, I just googled for basics tutorials, and did them top to bottom, until I felt confident. A foundation in python and R is beneficial. Tutorials on RNA-seq and OMICS are everywhere. Just pick one, do, repeat.

Are there specific tools/skills I should prioritize for cancer metabolomics specifically vs. general bioinformatics?

I doubt so. It all comes down to more or less the same type of analysis in most fields, and its the knowledge of literature and the system that is required to excel. Blindly applying bioinformatics creates a mess. I see this daily on Reddit with these "Independent Researchers" presenting their AI slop "research", and even here at my university I sometimes see talks where you clearly see that the analyst should have better read some textbooks on the subject first rather than just applying tool XYZ on the data at hand.

Arguably, since a lot of the differential analysis tools live in R you should get familiar with limma, since for me its the generic jac-of-all-trades when it comes to that sort of analysis. But other tools of course exist. Important is attractive data visualization, so become fluent in ggplot2, seaborn, or similar packages.

Any advice on finding a first small project to build a portfolio, given I don't have wet-lab funding?

Find a lab/supervisor to professionally support it. These self-cooked projects have limited scientific value. Not that youÄre not smart enough or anything like that, but without experience you quickly end up doing nonsense research that the field doesn't care about, nor will PIs to later give you a position.

Mostly agreeing with ATpoint, but I'd gently push back on one point: metabolomics genuinely does differ from the rest of omics in ways that catch people out, and since you asked specifically, they're worth knowing before you start.

The big one is that compound identification is itself an unsolved problem. In transcriptomics a gene has a stable ID; in untargeted metabolomics you have an m/z and a retention time, and mapping that to an actual molecule is probabilistic. Get familiar with the MSI confidence levels early, because a lot of published "metabolite X is elevated" is really a level-2 or level-3 annotation being reported as though it were certain.

Other things that don't transfer from RNA-seq:

• Missing values aren't random. A blank can mean below detection limit or genuinely absent, and those need different handling. Naively mean-imputing will bury real biology.
• Normalisation is different territory -- PQN, TIC, or QC-sample-based correction, not TMM or median-of-ratios.
• Instrument drift within a run is a real batch effect, which is why acquisition sequences interleave QC samples. If a public dataset has no QC injections, be sceptical of anything subtle you find in it.

So general skills transfer (R, tidyverse, ggplot2, and limma works perfectly well on log-transformed metabolite intensities), but the preprocessing is its own discipline. MetaboAnalyst is a reasonable way to get oriented conceptually, then xcms or MS-DIAL if you go deeper into raw data.

On the portfolio question, there's a middle path between "self-cooked project" and "wait for a supervisor": take a published HCC study with public data and reproduce its main figures. You learn the actual pipeline, you discover how underspecified most methods sections are, and it's defensible in a conversation with a PI in a way an invented analysis isn't -- you're demonstrating competence, not claiming a finding. TCGA-LIHC is free and well-trodden, and both MetaboLights and Metabolomics Workbench have HCC studies.

Last thing: being in Egypt is a genuine advantage here rather than a limitation. HCV-associated HCC is regionally distinctive, and framing your interest as "public multi-omics applied to a question that matters locally" is a lot more compelling to a prospective supervisor than another pan-cancer TCGA reanalysis.