Hypothesis

A latent factor extracted by a variational autoencoder (VAE) from integrated tumor mitochondrial proteomics, plasma metabolomics, and germline whole‑genome sequencing predicts progression‑free survival (PFS) in patients receiving anti‑PD‑1/PD‑L1 immunotherapy, and this factor can be causally shifted by metformin treatment, thereby altering PFS.

Rationale

Multi‑omics integration consistently outperforms single‑layer approaches for patient stratification and drug target prioritization [1] [2]. Mitochondrial dysfunction links altered metabolism to immune evasion, yet most studies examine only one layer [3]. VAEs excel at learning shared representations while handling missing data and platform‑specific noise [4]. We therefore hypothesize that a VAE‑derived mitochondrial metabo‑proteo‑genomic factor captures a causal axis linking tumor bioenergetics, systemic metabolites, and host germline variation to immunotherapy efficacy.

Testable Predictions

1. In a prospective cohort of melanoma or NSCLC patients (n≈200) receiving checkpoint blockade, the baseline VAE latent factor score will stratify patients into high‑ and low‑risk PFS groups (hazard ratio >2.0, p<0.01).
2. The factor will remain significant after adjusting for established biomarkers (PD‑L1 TPS, TMB, CD8⁺ infiltration) in a multivariate Cox model.
3. Metformin administration (standard dose for 12 weeks prior to therapy) will shift the latent factor toward the "high‑response" direction in a randomized sub‑study (n=40), producing a measurable change in factor score (Δ>0.5 SD, p<0.05).
4. Patients whose factor shifts toward the high‑response zone will exhibit improved PFS compared with those whose factor does not shift, establishing a causal link.

Experimental Design

• Cohort assembly: Collect pretreatment tumor biopsies (for mitochondrial‑enriched proteomics via TMT‑labelled LC‑MS/MS), plasma (untargeted metabolomics), and blood (whole‑genome sequencing) from patients about to start anti‑PD‑1/PD‑L1 therapy.
• Data processing: Apply quality control, batch correction (ComBat), and feed the three matrices into a joint VAE with modality‑specific encoders and a shared latent space (dimension=10).
• Factor identification: Use literature‑guided loading inspection to select the latent dimension most enriched for mitochondrial proteins, TCA‑cycle metabolites, and nuclear‑encoded mitochondrial SNPs.
• Outcome modeling: Perform Kaplan‑Meier stratification and Cox regression; validate in an independent external cohort (n=100) from a public dataset.
• Intervention arm: Randomly assign 40 patients to receive metformin 1500 mg daily for 12 weeks before immunotherapy; repeat multi‑omics profiling and compute factor change.
• Statistical plan: Power analysis assumes HR=2.0, α=0.05, 80 % power → n≈180; metformin sub‑study powered to detect 0.5 SD shift (α=0.05, 80 % power → n≈34).

Potential Outcomes and Falsifiability

• If the VAE factor does not predict PFS beyond clinical covariates, or if metformin fails to shift the factor or improve PFS, the hypothesis is falsified.
• Conversely, replication of predictive power and metformin‑induced factor modulation would support a mechanistic link between tumor mitochondrial state, systemic metabolism, germline background, and immunotherapy response, suggesting a combinatorial biomarker‑intervention strategy.

Implications

A validated VAE‑derived mitochondrial metabo‑proteo‑genomic factor would provide a concrete, measurable target for early intervention (e.g., metformin, exercise, NAD⁺ boosters) to prime patients for checkpoint blockade, moving multi‑omics from descriptive stratification to actionable precision oncology.