Finding. A deep learner can shove 771 cancer-cell-line proteomes toward 550 real tumors even when the tables are full of missing values. After that shove, the lines look more immune, more ECM, more talkative, and they look less like transcription-and-RNA machines. They also look poorer in mitochondrial gene-expression proteins. If you pick a line because it “matches” a tumor, you may be picking a line whose mitochondrial-expression proteome has been talked down.
Why this paper matters
Everyone knows lines are not tumors. Transcriptomic integrators already exist. Proteomics has holes, so the equivalent tool was missing. ProtInt is that tool. The mitochondrial clause is a side result that matters for this site: the adaptation axis includes mitochondrial gene expression.
Whether that axis is biology or batch is not settled. Report it as a systematic shift.
What they actually measured
Integration performance versus other methods; recurrent up/down protein classes after adapting lines to tumors.
How to read the score
Low sixties. Methods win, mitochondrial hint. Score 64.
What to do with it
If you match lines to tumors on proteomics, try ProtInt and then check the mitochondrial gene-expression module by hand. Do not conclude tumors have weaker mitochondrial expression from the warped residual alone. The directional implication is that forcing cell-line proteomes to look like tumors systematically lowers mitochondrial gene-expression proteins.