Hypothesis

We hypothesize that integrating spatially resolved metabolomics, proteomics, and transcriptomics with AI‑driven network inference reveals metabolite‑protein hubs that drive epigenetic reprogramming preceding drug resistance in solid tumors.

Mechanistic Rationale

Multi‑omics integration already improves disease classification and causal target identification [1][2]. Methods like sparse PLS‑DA and variational autoencoders now handle high dimensionality and missing data, allowing true integration rather than simple concatenation [4]. It's becoming clear that metabolite‑enzyme contacts can act as rapid signaling nodes. Adding spatial information lets us see where metabolites accumulate relative to protein expression and gene activity within tumor niches. Recent work shows that combining proteogenomics with Mendelian randomization and AlphaFold3 pinpoints druggable proteins [5]. We extend this by proposing that specific oncometabolites (e.g., 2‑hydroxyglutarate, succinate) physically interact with chromatin‑modifying enzymes, altering their activity and thus reshaping the epigenetic landscape. These interactions are missed when omics layers are averaged across bulk tissue.

Testable Predictions

1. In pretreatment biopsies, spatial colocalization of high 2‑hydroxyglutarate levels with increased DNMT3A protein predicts subsequent resistance to EGFR inhibitors.
2. We don't expect these hubs to be evident in normal tissue.
3. Disrupting the metabolite‑enzyme interaction (using a small‑molecule blocker) will reduce histone methylation marks and restore drug sensitivity in patient‑derived organoids.
4. Multi‑omic spatial models will outperform bulk multi‑omic models in predicting progression‑free survival, with an expected increase in median MCC of at least 0.15.

Experimental Design

• Collect matched pretreatment and post‑progression samples from 60 patients with non‑small‑cell lung cancer receiving EGFR‑TKI therapy.
• We're performing imaging mass spectrometry for metabolites, multiplexed ion beam imaging for proteins, and spatial transcriptomics (10x Visium) on adjacent sections.
• Use a variational autoencoder framework to fuse the three modalities, generating a joint latent space that captures cross‑layer interactions.
• Apply causal network inference (e.g., ICM) to identify metabolite‑protein edges that correlate with changes in DNA‑methylation signatures measured by bisulfite sequencing.
• Validate top hubs by CRISPR‑knockout of the protein or chemical inhibition of the metabolite in organoid cultures, measuring drug response and epigenetic marks.
• Statistical analysis: compare prediction accuracy of spatial vs bulk models using DeLong test for AUC; require p<0.05 to reject null hypothesis of no improvement.

Potential Impact

If confirmed, this approach would provide a mechanistic link between microenvironmental metabolism and epigenetic drug resistance, enabling patient‑specific combinatorial therapies that target the metabolite‑protein axis before resistance emerges.