Proteomics: Cross-Platform Blood Signature for Lung Cancer

Low-dose CT scans pick up lung nodules that mostly turn out to be benign, and the false alarms drive biopsies, anxiety, and cost. A blood test that flags real malignancy would help decide who needs follow-up and who can be reassured. This is where proteomics earns its place. A group from Oxford Cancer Analytics and academic partners measured plasma proteins in 490 lung cancer patients and 124 matched controls, then used explainable machine learning to pull out a protein signature that holds up across more than one measurement technology.

Inside the Proteomics Pipeline

The study ran two measurement chemistries side by side on the same plasma. Data-independent acquisition mass spectrometry, paired with Top 14 abundant-protein depletion, quantified roughly 2,519 proteins. The Olink Explore proximity extension assay, an antibody-based panel, measured about 1,748, with 1,136 overlapping the mass spectrometry set. Models were trained on each readout using an 80/20 train and holdout split, and explainable AI surfaced the proteins they leaned on most. A DNA-aptamer proteomic method then served as a third lens, sharpening the real question: which markers survive when you change how you measure them?

Key Findings

• Two platforms, one question: DIA mass spectrometry quantified about 2,519 plasma proteins and the Olink panel about 1,748, sharing 1,136; models built on each reached AUROC 0.91 and 0.97 on held-out samples.

• A signature that travels: a DNA-aptamer proteomic readout helped define a cross-platform protein set that scored 0.88 on both measurement types and still separated cases from controls in an external cohort.

• Feature selection stayed explainable: explainable AI ranked model-consistent proteins, so the panel reflects features the models agreed on rather than one algorithm's quirk.

• Biology pointed to the tumor's neighborhood: enriched proteins clustered around chemotaxis, cell adhesion, wound healing, and immune response, consistent with a host reaction to a growing tumor more than tumor-shed protein alone.

Figure 1. Study design and comparison of the two proteomics platforms. Panel A outlines the workflow: plasma from 124 healthy donors and 490 lung cancer patients, randomized, then profiled by DIA mass spectrometry after Top 14 depletion and by the Olink Explore proximity extension assay, feeding a machine learning and explainable AI pipeline. Panel B is a Venn diagram of proteins quantified by each platform, with 1,136 shared. Panels C and D relate plasma concentration to detection rank and to missingness. Panels E and F summarize measurement variability and protein-level agreement, while Panel G shows pathway over-representation across markers found by mass spectrometry, the affinity panel, or both. Adapted from Gushterov et al. (2026), Communications Medicine.

From Two Platforms to One Cross-Platform Signature

A model that scores well on the instrument it was trained on is easy to oversell. The harder test is whether the same proteins matter when the chemistry changes. By comparing mass spectrometry against an affinity panel and then an aptamer assay, the authors kept only markers that behaved consistently. Building on plasma proteome profiling described by Geyer and colleagues (2016), that cross-platform discipline is what gave the signature a fighting chance outside the discovery cohort, where it reached an AUROC near 0.88.

Biology Behind the Markers

The proteins carrying the signal were not random. They concentrated in chemotaxis, leukocyte adhesion, wound healing, and interferon-driven immune signaling, which reads as the body reacting to a tumor rather than the tumor spilling its contents into blood. One caveat is worth stating plainly: the design compares known cancer against controls, so the signature still has to prove it can catch early disease in people who look healthy. Prospective screening data will settle that.

In Practice: Trusting a Cross-Platform Signature

Cross-platform concordance is the quiet workhorse of any plasma biomarker program, and this paper treats it seriously by checking whether mass spectrometry, an antibody panel, and an aptamer assay agree on the same proteins. When a signature survives that comparison, it is far likelier to hold up in an outside cohort than one tuned to a single instrument. We see the same pattern in client work at Dalton, where proteins that replicate across measurement chemistries are the ones worth carrying into validation, and the ones that drop out usually trace back to platform-specific affinity quirks rather than biology.

Frequently Asked Questions

What is proteomics used for in cancer detection? Proteomics measures the proteins circulating in blood or present in tissue to find patterns that separate disease from health. In cancer, it is used to discover and validate blood-based markers that flag malignancy earlier or with fewer false positives than imaging alone.

How accurate was the proteomics blood signature for lung cancer? Models trained on each platform reached AUROC values of 0.91 for mass spectrometry and 0.97 for the antibody panel on held-out samples. The cross-platform signature scored about 0.88 on both and still separated lung cancer from controls in an external cohort. Those are strong discovery-stage numbers that now need prospective testing.

How many samples does a plasma proteomics biomarker study need? This study used 490 patients and 124 controls, sizeable for discovery but modest for proving screening value. The right number depends on effect size, how many candidate proteins you screen, and whether you plan a separate validation cohort to confirm the signal.

Conclusion

What is believable here is that a blood protein signature can separate lung cancer from controls across three measurement technologies, a higher bar than most single-platform claims clear. What is not yet proven is early detection in a screening setting, where cases are not known in advance. For teams planning a biomarker program, the practical lesson is to design cross-platform checks in from the start.

Citation

Gushterov, N., Hankey, L., Kaneva, I., Dupalliwar, M., Syed, J., Mi, E., Mi, E., Szulc, D. A., Zhan, L. J., Patel, D., Fischer, R., Kessler, B. M., Liu, G., Halner, A., & Liu, P. J. (2026). Multidimensional proteomics and explainable AI feature selection identify cross-platform lung cancer molecular signature in blood plasma. Communications Medicine. https://doi.org/10.1038/s43856-026-01701-8

Note

This blog post summarizes findings from the above-cited research. Figures are adapted from the original publication. For full details, please refer to the source article.

By Seungjun Yeo, CEO at Dalton Bioanalytics. Specializing in multi-omics mass spectrometry for drug discovery and biomarker research.