Integrating machine learning-derived biomarkers from Multiple Sclerosis transcriptomic data into evolutionary trade-off models will identify gene expression patterns that predict the emergence of drug-resistant immune cell phenotypes.

In treatment-refractory ("smoldering") multiple sclerosis, CD8+ T cells at chronic active rim (CA-RIM) lesion margins express a reproducible transcriptomic signature — elevated DNMT1 (log2FC +1.59), ZNF740/BRD3 axis (log2FC +1.15), and CTSS (log2FC +1.16) — that (a) is detectable in single-cell or sorted CD8+ material with FDR<0.1 in ≥2 independent cohorts, (b) is causally linked via network proximity to STAT3/IFNG/CSF1R-PTPRC hub pathways rather than being a passive bystander marker, and (c) predicts, prospectively, which patients on disease-modifying therapy will progress to a drug-resistant/PIRA (progression independent of relapse activity) phenotype within 24 months, with area-under-ROC ≥0.70 in a held-out validation cohort. The hypothesis is falsifiable at each of these three sub-claims independently.

Minimum viable test (MVT): 3-arm tiered validation. Arm A (in silico replication, 4 weeks): re-run Phase 1–4 pipeline on 2 additional independent scRNA-seq MS cohorts (target: GSE180759 or equivalent public CD8+ sorted datasets) to confirm DNMT1/ZNF740/CTSS signal independent of GSE193770/GSE138614. Arm B (functional validation, 10–12 weeks): sorted CD8+ T cells from n=20 smoldering MS patients + n=20 matched RRMS/controls; scRNA-seq + flow cytometry validation of DNMT1/ZNF740/BRD3/CTSS protein-level expression; ex vivo BET/CTSS/DNMT inhibitor dose-response assays measuring IFN-γ, granzyme B, MBP-cleavage readouts. Arm C (predictive/clinical, retrospective, 8 weeks): apply composite biomarker score to an existing longitudinal MS biobank cohort (n≥150, ≥24-month follow-up, known PIRA/non-PIRA outcome) using banked PBMC/CSF samples; compute AUROC for drug-resistance/progression prediction.

# Arm A: replication pipeline
for cohort in [GSE180759, other_independent_CD8_dataset]:
    adata = load_and_QC(cohort)
    adata_int = scANVI.transfer(reference=atlas_32239cells, query=adata)
    cd8_clusters = leiden_subset(adata_int, marker=["CD8A","CD8B"])
    pseudobulk = aggregate_by_sample(cd8_clusters)
    deg_results = DESeq2(pseudobulk, design=~CA_RIM_status)
    replication_check(deg_results, targets=["DNMT1","ZNF740","CTSS"],
                       prior_effects={"DNMT1":1.59,"ZNF740":1.15,"CTSS":1.16})

# Network proximity permutation
observed_prox = string_proximity(targets, seeds=["STAT3","IFNG","CSF1R","PTPRC"])
null_dist = [string_proximity(random_geneset(50), seeds) for _ in range(1000)]
p_value = mean(null_dist <= observed_prox)

# Arm B: functional assay
for donor in cohort(smoldering_MS=20, control=20):
    cd8 = FACS_sort(donor.PBMC, marker="CD8+")
    scRNA = sequence(cd8)
    protein = western_flow(cd8, targets=["DNMT1","BRD3","CTSS"])
    for drug, doses in {"JQ1":[0.01,0.1,1,5,10],
                        "RO5459072":[0.01,0.1,1,5,10],
                        "decitabine":[10,30,50,100]}.items():
        for d in doses:
            treated = culture(cd8, drug=drug, dose=d)
            readouts = measure(treated, ["IFNG","GZMB","MBP_cleavage","CD74"])
            log(donor, drug, d, readouts)

# Arm C: predictive model
X = compute_composite_score(biobank_samples, formula=phase3_formula)
y = outcome_labels(biobank_samples, endpoint="PIRA_24mo")
model = XGBoostCox().fit(X_train, y_train)  # nested 5-fold CV
auroc, ci = bootstrap_auroc(model, X_holdout, y_holdout, n=1000)
delong_test(model_auroc=auroc, baseline_model=clinical_baseline(X_train,y_train))