AI Transformer Reads Raw Protein Fingerprints to Sort Cells and Spot Severe COVID-19

Proteins are the workhorses of every cell, and the complete set of proteins a cell or tissue produces—its proteome—offers one of the closest snapshots of biology in action. For decades, scientists have used mass spectrometry to catalogue these proteins, but a stubborn bottleneck has remained: turning the enormous, noisy raw data files produced by the machine into a clear answer about what kind of sample is being measured. A team at the German Cancer Research Center and Heidelberg University now reports in Nature Communications a deep-learning system called ProFormer that does exactly that, classifying proteomic samples directly from raw mass-spectrometry signals with a speed and robustness that traditional pipelines cannot match.

The core insight behind ProFormer is a shortcut. In conventional proteomics, the instrument first measures intact peptide ions—their mass-to-charge ratio, retention time and signal intensity (the so-called MS1 level)—and then painstakingly fragments selected ions to read their amino-acid sequences (MS2). That fragmentation step is essential when the goal is to identify every protein, but it is slow, and the downstream interpretation rests on a chain of assumptions. The Heidelberg team reasoned that for many practical questions—Is this cell a T cell or a cancer cell? Does this patient have severe disease?—the sample’s identity is already written into its MS1 fingerprint, the pattern of peptide ion signals, without ever needing to name the proteins involved.

Earlier attempts to exploit this idea converted MS1 data into image-like heatmaps and fed them to convolutional neural networks, the same architecture that revolutionized photo recognition. But the authors argue this approach has a fundamental flaw: unlike pixels in a photograph, neighboring peptide signals in a mass spectrum have no meaningful spatial relationship. Binning them into a rigid grid can split or merge unrelated signals, and the team’s own benchmarks suggest the resulting CNNs are prone to overfitting. When the researchers evaluated sixteen image-based CNN configurations on a test set measured six months after the training data, accuracy dropped by 20 to 57 percentage points compared with validation results—a warning sign that previously reported CNN successes may partly have been artifacts of non-independent testing.

ProFormer takes a different route. Instead of images, it consumes tabular lists of peptide ion features—mass, retention time, intensity and, where available, ion mobility—and uses a learnable component called a patch merger to compress hundreds of thousands of signals into a manageable set of tokens. The patch merger weighs each peptide ion against learned reference points and aggregates similar signals dynamically, without fixed grid borders. These tokens then flow through a lightweight transformer, the self-attention architecture that underpins modern language models, which can compare dependencies across the entire dataset rather than only within local neighborhoods. A final classification token is read out to produce a probability score for the sample’s identity.

The team stress-tested this design with a deliberately difficult experiment: distinguishing HeLa cells treated with interferon-gamma, a cytokine that produces only subtle proteomic changes in these cells, from untreated controls. Training data comprised 142 injections, and the test set—71 injections of the same samples—was acquired six months later on the same instrument, deliberately introducing a technical domain shift. The best ProFormer model reached 84.0 percent area under the curve, while the strongest CNN, a ResNet_v2_50, managed 79.9 percent and showed a large gap between validation and test performance. Traditional machine-learning approaches such as logistic regression, random forest and XGBoost topped out at 64.7 percent AUC, underscoring why deep learning was needed at all.

Generalization was tested further in two harder scenarios. In one, the model faced biologically distinct samples cultured by a different operator in 96-well plates; the 4D-ProFormer achieved roughly 97 percent AUC, the best of the four top architectures tested. In another, cells were treated with seven interferon-gamma concentrations spanning 0.001 to 50 nanograms per milliliter—doses absent from training. Only ProFormer showed the expected monotonic dose-response relationship, with a Spearman correlation of +0.72 and a sigmoidal fit of R² = 0.77, detecting the cytokine’s effect down to 9.44 nanograms per milliliter, roughly an order of magnitude below the training dose. Competing CNNs either lost sensitivity, lost specificity, or even reversed their predictions.

The system also proved strikingly frugal with data. Trained on just ten samples, ProFormer already reached about 85 percent accuracy on the bulk dataset and 75 percent on a single-cell task, with performance improving gradually as more cells were added. Filtering out low-abundance peptide ions, which suffer from missing values and inconsistent quantification, boosted single-cell accuracy by around five percentage points. When the researchers injected Gaussian noise into test features, ProFormer held its accuracy up to 10 percent noise, whereas the ResNet CNN degraded to random guessing at that level—a stark demonstration of the transformer’s resilience to the variability that plagues real-world proteomic measurements.

Applied to published single-cell proteomics datasets, ProFormer distinguished cell types and cell-cycle stages with AUCs between 94.2 and 100 percent, even when a dataset contained fewer than 800 peptide ions per cell. More subtle tasks—separating migrating from stationary cells (82.1 percent AUC) or differentiating pluripotency states of mouse embryonic stem cells (80.8 percent)—remained challenging but tractable, with accuracy tracking the biological similarity of the classes being compared. This positions the tool as a candidate solution to one of single-cell proteomics’ most pressing needs: annotating thousands of individual cells rapidly and reproducibly before downstream analysis.

The most clinically striking result came from blood plasma. Using data from the COMBAT consortium, which profiled 490 participants including severe COVID-19 patients, mild cases, severe influenza, sepsis and healthy volunteers, the team trained ProFormer to flag severe COVID-19 from plasma MS1 features alone. Splitting the mass-to-charge feature into separate mass and charge components lifted validation AUC from 72.7 to 86.3 percent, and the final model achieved 86.3 percent AUC on the test set with an F1 score of 76.7 percent. Tuned for maximum safety, it identified every severe COVID-19 case with 100 percent sensitivity while correctly ruling out the disease in 46 percent of non-cases—an encouraging profile for a triage or screening tool.

Crucially, the model’s decisions can be interrogated. Using SHAP values, a game-theoretic framework for attributing predictions to individual features, the researchers traced which peptide ions drove the COVID-19 classification. The most influential signals mapped to proteins of the complement system, including C3 and C4A, and to coagulation factors such as antithrombin III and the three fibrinogen chains—findings that mirror the original COMBAT study’s conventional differential analysis. One particularly informative peptide from C3 was inversely regulated, hinting at proteolytic cleavage during complement activation that protein-level analysis would have missed. Feeding just the top-ranked peptide ions into a simple random forest reproduced the deep model’s accuracy, while a model trained on all ions achieved only about 55 percent AUC—proof that ProFormer finds genuinely discriminative signals rather than exploiting technical artifacts. The authors envision a future version that operates entirely without MS2 sequencing, potentially accelerating acquisition further. For now, ProFormer offers a template for turning raw proteomic data into fast, explainable decisions, from annotating single cells to stratifying patients at the bedside.

Subject of Research: Deep-learning classification of single-cell and plasma proteomes from raw mass-spectrometry data

Article Title: ProFormer: generalizable classification of single-cell and plasma proteomes using deep learning

Article References: Krull, K. K., Kühn, A., Höhn, J., Ali, S. A., Otterbein, M. H., Brinker, T. J., & Krijgsveld, J. (2026). ProFormer: generalizable classification of single-cell and plasma proteomes using deep learning. Nature Communications, 17(1), Article 10493. https://doi.org/10.1038/s41467-026-78143-2

Image Credits: AI Generated

DOI: 10.1038/s41467-026-78143-2

Keywords: proteomics, deep learning, transformer, mass spectrometry, single-cell proteomics, plasma biomarkers, COVID-19, explainable AI, SHAP values, domain generalization, machine learning, diagnostics