Plasma proteomics foundation model pretrained on healthy UK Biobank participants, then fine-tuned to predict 16 common chronic diseases.
No providers recorded yet. Browse all providers
A proteomic risk score for type 2 diabetes is usually fitted directly on diabetes cases and controls. It learns which proteins separate the two groups, but never learns what an ordinary, healthy plasma proteome looks like to begin with. It therefore has no baseline of normal protein co-regulation against which a pre-disease shift would stand out. ProLM starts from that baseline. A BERT-style encoder is first pretrained, without any clinical labels, to reconstruct masked protein abundances across the plasma proteomes of people who stayed free of common chronic diseases. Only then is it fine-tuned, disease by disease, to predict who will fall ill.
ProLM was developed at Harbin Institute of Technology with epidemiologists from Peking University and the China Kadoorie Biobank group at the University of Oxford, together with collaborators at Harbin Medical University and Capital Medical University. It was published in Nature Communications in July 2026. The pretraining cohort is 15,499 participants from the UK Biobank Pharma Proteomics Project (UKB-PPP), each measured for 2,873 plasma proteins. Every one of them was free of the studied diseases at baseline and throughout a median 15.5 years of follow-up.
ProLM belongs to a small group of population-proteomics encoders that includes the Plasma Protein-Token Transformer. What sets it apart is that it treats each protein as a token carrying biological identity as well as a measured value.
The released v0.1.0 backbone is a 24-layer transformer encoder with hidden size 768. Proteins are split into blocks of 512 tokens, each with a [CLS] token, and the ProteinBERT inputs have 15,599 dimensions. Pretraining stopped early after 99 epochs, reaching 0.85 reconstruction accuracy and R² = 0.53 on masked values. With no labels, the learned CLS embeddings separate male from female participants. For downstream tasks the authors unfroze the last eight encoder layers and trained for 40 epochs. The fine-tuning cohorts had no overlap with the pretraining cohort.
In incident-disease prediction, ProLM reached AUC above 0.8 for 14 of 16 diseases. Examples include type 2 diabetes (0.945), prostate cancer (0.949), dementia (0.928) and COPD (0.920). It had the best AUC among the compared methods in 10 of 16 tasks. It significantly outperformed an MLP in 15 diseases and FT-Transformer in 13. Against gradient-boosting baselines it was comparable: significantly better than LightGBM in 7 diseases and CatBoost in 6. The ProLM-derived proteomic risk score (ProRS) beat an age-and-sex model for all 16 diseases, the ASCVD risk equation for 14 and a 35-variable clinical panel for 11. Key findings were externally evaluated on Olink Explore data from a China Kadoorie Biobank case-control substudy covering six outcomes.
ProLM targets early risk stratification for chronic disease from one blood draw. ProRS divides individuals into low-, medium- and high-risk groups whose disease-free survival curves diverge over follow-up. Integrated Gradients attribution ranks the proteins driving each prediction. These include GDF15, whose abundance began changing more than 15 years before clinical diagnosis. The authors cross-referenced prioritized proteins against druggable-genome catalogs to point to candidate targets. The same backbone also classifies subtypes of dementia, stroke, inflammatory bowel disease and diabetes. Fine-tuning requires plasma proteomic data and case labels. The UK Biobank and China Kadoorie Biobank cohorts used to build and evaluate the model are available only through their access application processes.
ProLM shows that the pretrain-then-fine-tune recipe of sequence language models transfers to population-scale protein abundance. Here a healthy reference proteome plays the role of unlabeled text. The code is MIT-licensed, and the pretrained checkpoint is public as a GitHub release asset, so other groups can fine-tune on their own Olink-style panels. The released repository contains fine-tuning code only; the pretraining script is not included. The pretraining population is largely White and middle-aged, and the external evaluation covered six outcomes in a single case-control design. The authors state that prospective validation is needed before any clinical use.
Much of this page is generated or calculated automatically. Flag anything that looks off and we will re-run it.