What AI can see before peptides become a protein average

Three pieces of the same protein did not tell the same story. In the ProFormer study, peptides from complement C3 helped distinguish severe COVID-19 plasma profiles from comparison profiles, yet one changed in the opposite direction to the others. The authors suggest protein cleavage as an explanation; they did not establish that mechanism experimentally.[1]

That observation matters because a protein’s name can conceal several molecular forms. Proteins can be cut, modified or assembled into different structures. In a common proteomics workflow, researchers deliberately digest them into smaller peptides, separate those peptides by liquid chromatography and measure them by mass spectrometry. Software then maps the fragments back to their likely parent proteins. The mapping is useful, but digestion has already broken the physical connection between each fragment and the exact molecule it came from.[2]

An average can erase a difference

Take an explicitly hypothetical example. Two peptide signals associated with one protein both have an intensity of 100 in a reference sample. In a second sample, one rises to 150 while the other falls to 50. Their sum remains 200 and their average remains 100. Each signal has changed by 50%, but an equally weighted protein-level average reports no change.

Real protein-quantification methods need not use that simple average, and peptide intensities are not interchangeable molecule counts. The example isolates an information problem: collapsing several measurements into one number is a choice about which differences will remain visible. Keeping the separate values preserves the possibility of asking a different question later.[2]

Earlier work provides a concrete biological example. The COPF method grouped peptides whose quantities varied together across samples. In its 2021 analysis of proteasome assembly, two peptide groups associated with the PSMB7 protein behaved differently. Targeted re-extraction of a peptide characteristic of the processed form supported the distinction between the precursor and the form incorporated into the mature complex. A discrepancy among fragments could describe a maturation step rather than measurement noise.[2]

ProFormer feeds a transformer tables of peptide-ion mass, charge, retention time and intensity. A learned merger compresses variable-length tables into tokens. Peptide and protein names are withheld from the classifier, then restored to interpret influential measurements.[1]

There are two different information choices here. A chemical identity helps a scientist explain what a signal means. The measured pattern helps a classifier distinguish samples. Keeping those roles separate allows a model to learn from the pattern while leaving an investigator a route back to the underlying biology.

The comparison has to be biologically demanding

The plasma data came from the COMBAT blood atlas. Its original 2022 study compared COVID-19 of varying severity with healthy volunteers, influenza and sepsis. The main proteomics analysis reported 105 proteins across 257 individuals and 340 samples. It identified changes in acute-phase and complement proteins, among other processes. Those comparisons let researchers investigate what differed with disease severity and what was shared with other serious infections.[3]

That is a more revealing question than whether an acutely ill person differs from a healthy volunteer. A model can succeed at the latter by recognizing illness generally. Including other infections asks whether the pattern provides a more specific distinction. It changes the meaning of success before any algorithm is chosen.

ProFormer’s Figure 5 reports a test set of 74 profiles: 35 severe COVID-19 and 39 comparators. ROC AUC was 0.863. AUC measures discrimination across score thresholds; it does not mean 86.3% of patients would receive a correct diagnosis at a particular cutoff.[1]

This remains a proof of concept: identities were masked, but extracting MS1 features still used MS2-based peptide identification.[1]

The more interesting promise is therefore an analytical one. A protein label can organize a collection of fragments without explaining why some fragments disagree. Preserving those measurements gives researchers a chance to discover that the disagreement is the biological signal they were looking for.

Report an error or suggest a correction