Science
AI model ProFormer classifies cells and blood samples directly from protein measurements
AI-generated illustration
A deep-learning tool called ProFormer sorts samples, such as cell types or patient blood plasma, straight from protein measurements without slow data interpretation. The abstract reports it beat several other methods in laboratory datasets.
14model designs compared, with ProFormer outperforming the traditional machine-learning and image-based networks tested
The facts
- 1In the reported tests, ProFormer classified single cells by cell type, cell cycle stage or differentiation path (how a cell changes into a specialised type). It also classified patient disease status from blood plasma.
- 2It is a transformer, a type of deep-learning model. It reads table-style features of peptide ions taken from the first stage of mass spectrometry data, skipping time-consuming data interpretation.
- 3The authors compared 14 model designs. ProFormer beat traditional machine learning and image-based neural networks at classifying samples and at generalising to new data.
- 4Current instruments can measure thousands of proteins, even in single cells, but low throughput limits classifying large biological or clinical groups. The authors also show which individual peptides drove each decision.
- 5The authors point to patient stratification, early detection and single-cell analysis as possible uses. This abstract does not establish clinical use or real-world performance.
Why it matters
It shows a way to speed up sorting of protein data from large sample groups, and the peptide-level explanations let researchers question why the model made each call.
Sources
- Adapted from Nature Communications: ProFormer: generalizable classification of single-cell and plasma proteomes using deep learning · Karl K. Krull, Arlene Kühn, Julia Höhn, Syed Azmal Ali, Martin H. Otterbein, Titus J. Brinker, Jeroen Krijgsveld · Source license
