Hypothesis

Longitudinal, spatially resolved multi-omics data, when fed into a dynamic graph attention network that reweights genomics, proteomics, and metabolomics layers at each time point, will predict the emergence of lactate-driven immunosuppressive niches in circulating tumor cells before radiographic progression in patients receiving anti-PD-1 therapy.

Mechanistic Rationale

It's known that tumor hypoxia stabilizes HIF-1α, which upregulates LDHA and increases lactate export. Extracellular lactate binds GPR81 on tumor-associated macrophages, shifting them toward an M2-like phenotype that suppresses CD8+ T-cell activity and promotes PD-L1 up-regulation on tumor cells. Simultaneously, lactate-mediated histone lactylation alters transcription of genes involved in angiogenesis and immune evasion. These metabolic-immune adaptations generate spatially restricted niches that are invisible to bulk genomics but detectable by paired spatial transcriptomics (showing HIF-1α and GPR81 signatures), proteomics (LDHA, GPR81, PD-L1), and metabolomics (lactate, pyruvate ratios). Static fusion methods don't capture the temporal shift in layer importance; a graph attention network can learn to up-weight metabolomics and proteomics when lactate rises, thereby capturing the niche formation ahead of clinical failure.

Testable Predictions

1. In a prospective cohort, the dynamic multi-omics model will achieve a higher AUC for predicting progression-free survival at 6 months than a static multi-omics model (baseline AUC≈0.82 from prior work) – expected increase ≥0.07.
2. Patients classified as high-risk by the model will show, on paired serial biopsies, a significant rise in spatial lactate-HIF-1α-GPR81 co-localization (≥2-fold increase) compared with low-risk patients.
3. Experimental blockade of GPR81 in patient-derived xenograft models will attenuate the model-predicted risk score and delay tumor growth, falsifying the hypothesis if no effect is observed.

Experimental Design

We're enrolling 120 patients with metastatic melanoma or NSCLC starting anti-PD-1 therapy. We're collecting baseline and every 6-week peripheral blood for circulating tumor cell (CTC) clusters; performing spatial transcriptomics (10x Visium) on CTC clusters, targeted proteomics (Olink) for LDHA, GPR81, PD-L1, and untargeted metabolomics (LC-MS) for lactate/pyruvate. We're constructing a heterogeneous graph where nodes represent omics features and edges represent known biochemical links; applying a graph attention network that learns time-dependent attention weights. Primary endpoint: prediction of progression-free survival at 6 months; secondary: correlation with spatial lactate-HIF-1α-GPR81 signals in matched tumor biopsies (optional). Statistical analysis: compare AUCs using DeLong test; test risk-group differences with logistic regression; assess GPR81 blockade effect in PDX models using tumor growth inhibition curves.

If the model fails to outperform static multi-omics or if lactate-HIF-1α-GPR81 co-localization does not rise in high-risk patients, the hypothesis is falsified. Conversely, confirming the predictions would support a mechanistic link between lactate signaling, immune remodeling, and dynamic multi-omic weighting as a leading indicator of therapeutic resistance.