Hypothesis

Integrating longitudinal metabolomic flux profiles with single-cell proteomic and epigenomic dynamics using a transformer‑based temporal graph neural network (TempGNN) will forecast the emergence of therapy‑resistant cancer clones at least 8 weeks before clinical detection, and this predictive signal will be mechanistically linked to glutamine‑dependent acetylation of ARID‑containing lysine demethylases (KDMs).

Rationale

Recent multi‑omics AI frameworks achieve AUCs of 0.81‑0.87 for early cancer detection by fusing static genomic, transcriptomic, proteomic, and metabolomic layers[1][2]. However, resistance evolution is a dynamic process driven by metabolic reprogramming and chromatin remodeling[3]. Emerging work shows glutamine fuels ARID‑mediated acetylation of KDM enzymes, altering histone marks that sustain drug‑tolerant states[0dbdae24]. By capturing temporal changes in glutamine‑derived metabolites (e.g., α‑KG, acetyl‑CoA) alongside ARID‑KDM acetylation levels and downstream proteomic shifts, TempGNN can learn the causal trajectory toward resistance.

Testable Predictions

1. Prediction Accuracy – In a prospective cohort of 200 patients receiving targeted therapy, TempGNN trained on serial plasma metabolomics, single‑cell proteomics, and ATAC‑seq from baseline and early‑treatment biopsies will achieve an AUC ≥0.90 for predicting radiographic progression at week 12, outperforming models using only baseline multi‑omics (AUC ≈0.80) and clinical covariates alone (AUC ≈0.65).
2. Mechanistic Link – Patients predicted to develop resistance will show a significant rise in glutamine‑derived acetyl‑CoA flux (≥1.5‑fold increase) coupled with increased ARID‑KDM acetylation (≥2‑fold) at week 4, whereas non‑progressors will not. Pharmacologic inhibition of MTHFD2 (a key glutamine‑linked one‑carbon enzyme) in patient‑derived organoids will attenuate both the metabolic flux shift and ARID‑KDM acetylation, delaying resistance emergence in vitro.
3. Falsifiability – If TempGNN’s predictions do not surpass baseline models by ≥0.10 AUC, or if glutamine flux and ARID‑KDM acetylation changes are absent in predicted progressors, the hypothesis is falsified.

Experimental Design

• Cohort – Enroll 200 metastatic cancer patients initiating EGFR or HER2 targeted therapy; collect blood and tumor biopsies at baseline, week 2, week 4, and week 8.
• Omics Layers – Untargeted metabolomics (LC‑MS), single‑cell proteomics (SCoPE2), single‑cell ATAC‑seq, and targeted ARID‑KDM acetyl‑lysine quantification (PRM).
• Model – TempGNN architecture: metabolite nodes linked to protein and epigenomic nodes via temporal edges; transformer encoder captures sequential dependencies; output predicts resistance probability.
• Validation – Compare predictions to independent endpoint (ctDNA‑detected resistance mutations or radiographic progression). Perform multivariate Cox regression to assess added value of TempGNN beyond clinical variables.
• Mechanistic Assay – Use CRISPRi to knock down ARID1A/ARID1B in organoids; measure KDM acetylation (Western blot) and sensitivity to drug after glutamine deprivation or MTHFD2 inhibition (AGI‑6780).

Impact

Confirming this hypothesis would establish a clinically actionable, mechanism‑grounded early‑warning system for resistance, guiding preemptive therapy switches. It would also highlight glutamine‑ARID‑KDM acetylation as a druggable node, bridging metabolic and epigenetic vulnerabilities in precision oncology.

References

[1] AI-driven multi-omics integration in precision oncology [2] The Future of Multi-Omics in Cancer Clinical Trials [3] Integrative multi-omics approaches identify molecular pathways and improve Alzheimer's Disease risk prediction [0dbdae24] Glutamine‑Fueled ARID Acetylation of KDM