PanMETAI
Summary¶
PanMETAI is a diagnostic algorithm for pancreatic ductal adenocarcinoma (PDAC) built on TabPFN, a transformer-based tabular foundation model. It integrates serum ¹H NMR metabolomic and lipoprotein spectral features with age, CA19-9, and the protein biomarker Activin A, trained and validated on 902 participants (424 high-risk controls, 478 PDAC cases) across Taiwanese and Lithuanian cohorts. TabPFN-PanMETAI reached an AUC of 0.99 in Taiwanese validation and 0.93 in independent Lithuanian external validation, and — unlike CA19-9 alone or clinical/protein data alone — the addition of NMR metabolomics specifically boosted sensitivity for early-stage (I/II) disease [1].
Data and Modeling Approach¶
Serum samples underwent 600 MHz ¹H NMR profiling (high-throughput SampleJet system), yielding two spectral feature sets after normalization/trimming: NMR-M (small-molecule metabolites, 71,074 features) and NMR-M+L (metabolites plus lipoproteins, 42,500 features). Of eight candidate clinical/biochemical variables, silhouette-score analysis on UMAP clustering identified age and CA19-9 as the most discriminative, and Activin A was included given its reported association with early-stage PDAC. The final integrated input set was NMR-M, NMR-M+L, Activin A, age, and CA19-9.
Three modeling frameworks were compared:
- Multilayer SVM: per-modality feature selection (SelectKBest for NMR-M, recursive feature elimination for NMR-M+L), decision values from per-modality SVMs fused with clinical/protein features — designed to prevent any single high-dimensional modality from dominating the fused prediction.
- AutoGluon AutoML ensemble (14 models; best individual model was XGBoost).
- TabPFN: a transformer-based tabular foundation model using attention and in-context learning; because TabPFN is limited to <500 input features, a reduced feature set (215 NMR-M, 255 NMR-M+L features) was selected for it.
Performance¶
- Development-test set (blinded internal test): TabPFN outperformed multilayer SVM (AUC 0.96) and AutoGluon's best model, XGBoost (AUC 0.97), reaching AUC 0.98, sensitivity 0.98, specificity 0.90, accuracy 0.96 — with less training time and no hyperparameter tuning required.
- Blind test (Taiwanese cohort): TabPFN-PanMETAI achieved AUC 0.99 (95% CI 0.98–0.99), accuracy 0.94, sensitivity 0.93, specificity 0.94, outperforming the multilayer SVM (~0.90 accuracy) and XGBoost (~0.93 accuracy) comparators.
- External validation (Lithuanian cohort, n=322, geographically and behaviorally distinct from Taiwan): AUC 0.93 (0.90–0.95), sensitivity 0.90, specificity 0.83, with consistent performance across ancestrally and clinically different populations.
- Stage-stratified performance (Taiwanese cohort): late-stage (III/IV) PDAC — accuracy 0.95, sensitivity 0.95, specificity 0.95, AUC 0.99. Early-stage (I/II) PDAC — accuracy 0.92, sensitivity 0.78, specificity 0.95, AUC 0.95, with comparable results in the Lithuanian cohort.
- Modality ablation: clinical/biochemical parameters plus Activin A alone were sufficient for high accuracy on late-stage PDAC, but adding NMR metabolomic features specifically and substantially increased sensitivity for early-stage PDAC detection — the modality contribution that motivates the approach's clinical value, since CA19-9 alone is a poor early-detection marker.
Availability¶
TabPFN itself is open source at github.com/PriorLabs/TabPFN. Due to a pending patent (PCT/US25/51228), the TabPFN-PanMETAI implementation is distributed under managed academic access via Zenodo (PolyForm Noncommercial License 1.0.0, non-commercial academic use only). The AutoGluon comparator code is openly available at github.com/PanMETAI/AutoGluon_PDAC (Zenodo: 10.5281/zenodo.17701588).