Yellow fever vaccine transcriptomic time-course analysis

What this research found

Systems vaccinology holds that the gene-expression response in blood during the first days after a vaccination encodes how strong the eventual immune response will be. K-Dense re-analysed the canonical yellow-fever cohort of 25 vaccinated adults with a machine-learning pipeline built to eliminate data leakage, predicting antibody and T-cell outcomes from expression measured by Day 7. The predictive signal held up, reaching an area under the ROC curve of 0.87 for neutralizing antibody and 0.76 for CD8 T-cell activation against a 0.52 no-information baseline, but the genes carrying that prediction barely overlap the interferon, stress-response, and plasma-cell programs known to dominate this vaccine's biology.

• Neutralizing-antibody class was predicted with an area under the ROC curve of 0.87 and 80% accuracy, 20 of 25 held-out subjects correct, using the within-subject change in expression from Day 0 to Day 7. CD8 T-cell activation was best predicted from raw Day 7 expression at 0.76 and 72% accuracy.
• Which representation works is endpoint-specific and the alternatives sat near chance: antibody prediction from raw Day 7 expression scored 0.54, and T-cell prediction from the fold-change scored 0.64.
• Of 21 genes selected in at least 80% of cross-validation folds across the two best models, only IL5 mapped to any established immune module, at an adjusted p-value of 0.041. None touched the interferon-stimulated genes, the ER-stress and unfolded-protein response, or the plasma-cell program that the published yellow-fever literature reports.
• The selected features are only moderately reproducible. Mean pairwise overlap between the top-20 gene sets chosen in different folds was 0.48 to 0.55 by Jaccard index, meaning roughly half the features are interchangeable between folds even while the model predicts well.
• Pathway enrichment of the predictive genes returned ion-transport and neuronal terms rather than immune ones, alongside a more plausible mTORC1 signaling hit at an adjusted p-value of 0.010 — a pattern the analysis attributes to selection variance rather than vaccine physiology.
• No independent validation cohort was available: a data-integrity check found that one of the two public accessions re-hosts the identical 87 arrays, a 100% overlap. Training on one trial and testing on the other gave a perfect 1.00 for the antibody model, but the median antibody split is partly aligned with trial membership, so that figure reflects batch structure rather than vaccine biology.

How it was done

The yellow-fever vaccination time course was retrieved from the NCBI Gene Expression Omnibus: 87 microarrays from 25 healthy adults sampled at Days 0, 1, 3, 7, and 21 across two trials, with per-subject neutralizing-antibody titer and a CD8 T-cell activation index. Probes were mapped to gene symbols and collapsed by highest mean expression, giving 8,433 unique genes. Two feature sets were built — the within-subject Day 7 minus Day 0 log2 fold-change, which removes each subject's baseline and the trial batch offset, and raw Day 7 expression — and each endpoint was split at its cohort median, yielding 13 high and 12 low subjects. Models were evaluated by leave-one-subject-out cross-validation across 25 folds in which standardization, filtering to the top 20 genes by point-biserial correlation, and an L1-penalized logistic classifier were all fit inside the training fold only. Feature-selection stability was measured across all 300 fold pairs, and the consensus genes were tested for enrichment against curated vaccine signatures using the 8,433 measured genes as the background universe.

Data sources

• NCBI Gene Expression Omnibus GSE13485 — 87 microarrays, 25 subjects, two trials, platform GPL7567
• Querec et al., Nature Immunology 10:116 (2009) — the original systems-biology analysis of this cohort
• Gaucher et al., Journal of Experimental Medicine 205:3119 (2008) — the canonical YF-17D response modules
• Li et al., Nature Immunology 15:195 (2014) — Blood Transcription Modules across five human vaccines
• Reactome, KEGG, WikiPathways, MSigDB Hallmark, and Gene Ontology via Enrichr
• Ambroise & McLachlan, PNAS 99:6562 (2002) and Varma & Simon, BMC Bioinformatics 7:91 (2006) — the cross-validation leakage cautions this design follows