Full Research Blueprint: Tabular Foundation Models as Surrogate Epigenetic Clocks

The Problem, Precisely Stated

Gold standard: Epigenetic clocks (GrimAge, DunedinPACE, Horvath, PhenoAge-DNAm) require Illumina EPIC 850K methylation arrays. $300–500/sample. 3–6 week lab processing. Specialized bioinformatics pipeline.

Current blood-based alternative: PhenoAge (Levine et al., 2018, Nature Medicine), linear regression on 9 blood markers trained to predict chronological age, not methylation clock outputs. Linear. 2018. Never updated. Not actually trained to approximate GrimAge or DunedinPACE.

The gap we fill: Train a deep tabular model directly on the 30–60 blood chemistry markers available in a standard clinical panel, with methylation clock scores as the prediction target. Multi-task across 4 clocks simultaneously. Calibrated uncertainty. Graceful missingness. Deployable for $30/sample.

Our unique edge: Antiaging Labs is operationally running PhenoAge on real clients. We understand what 50-marker panels look like in practice, what's missable, and what "clinical deployment" requires. No competing researcher has that grounding.

Section 1, Dataset Selection

Primary Dataset: UK Biobank (UKB)

This is the only dataset that gives you scale, breadth, and both methylation + blood chemistry simultaneously.

What it has:

• ~500K participants total
• ~45,000 participants with EPIC 850K methylation arrays
• Blood biochemistry: 34 markers, Albumin, ALP, ALT, ApoA, ApoB, AST, Calcium, Cholesterol, CRP, Creatinine, Cystatin C, Direct bilirubin, eGFR, GGT, Glucose, HbA1c, HDL, IGF-1, LDL, Lipoprotein-A, Phosphate, SHBG, Testosterone, Total bilirubin, Total protein, Triglycerides, Urate, Urea, Vitamin D
• Full CBC: WBC, RBC, Haemoglobin, MCV, Platelets, Neutrophils, Lymphocytes, Monocytes, Eosinophils, Basophils
• ~42 blood markers total in the methylation subset

How to calculate your targets from this data:

• Horvath clock, from methylation CpGs using the methylclock R package
• GrimAge, from methylation using grimage R package (Lu et al., 2019)
• DunedinPACE, from methylation using DunedinPACE R package (Belsky et al., 2020)
• PhenoAge-DNAm, the methylation-based version (distinct from the blood-based PhenoAge)

Training data becomes: [blood chemistry vector] → [4 clock output scores]

How to apply:

1. Register at ukbiobank.ac.uk, the process takes 2–4 months. Start today.
2. Complete mandatory GDPR and data handling training
3. Submit a project application with your research question (below)
4. Data delivered to a Tier 2 secure compute environment

Application pitch:

"We aim to develop deep learning surrogate models for epigenetic biological age clocks using standard clinical blood biochemistry. Using paired methylation array and blood biochemistry data, we will train tabular transformer architectures to predict established clock outputs (GrimAge, DunedinPACE, Horvath's DNAmAge, PhenoAge-DNAm) from routine blood markers alone. This work aims to enable scalable, low-cost population-level biological age monitoring without requiring DNA methylation arrays."

Start this application this week. It is the long pole in the tent.

Secondary Datasets (external validation, critical for Nature-tier papers)

The Indian cohort angle (our unique contribution):
Antiaging Labs clients, even a small cohort of 30–50, give us something no competing lab has: Indian urban professional blood chemistry paired with biological age calculation. After the main model, we validate on this cohort. Cross-ancestry generalizability writes itself as a finding. Most epigenetic clock validation is 95% European-ancestry cohorts.

Section 2, Architecture Design

Baselines (we must beat all of these)

Our novel architecture: BioAge-Transformer

Core design principles:

1. Feature tokenization (each biomarker becomes a vector, not a raw float)
2. Biological pathway-aware attention (metabolic markers attend to each other, etc.)
3. Learned missing value tokens (not imputation, critical for clinical use)
4. Multi-task heads (one per clock, shared encoder, separate decoders)
5. Calibrated uncertainty (Gaussian NLL outputs mean + variance per clock)

Architecture in full:

class BioAgeTransformer(nn.Module):
    """
    Multi-task tabular transformer for surrogate epigenetic clock prediction.
    Predicts GrimAge, DunedinPACE, Horvath, and PhenoAge-DNAm
    from standard clinical blood chemistry panels.
    """
    def __init__(
        self,
        n_features: int,          # ~42 blood markers
        d_model: int = 192,
        n_heads: int = 8,
        n_layers: int = 6,
        n_clocks: int = 4,
        n_pathways: int = 6,      # metabolic, lipid, inflammation,
                                   # hormonal, CBC, nutrients
        dropout: float = 0.1,
    ):
        super().__init__()

        # ── Feature tokenization ──────────────────────────────────────
        self.feature_embedding = nn.Embedding(n_features, d_model)
        self.value_projector   = nn.Linear(1, d_model)

        # Learnable missing-value token (one per feature, not global)
        self.missing_tokens = nn.Parameter(torch.randn(n_features, d_model))

        # ── Biological pathway embeddings ────────────────────────────
        self.pathway_embedding = nn.Embedding(n_pathways, d_model)

        # ── [CLS] token ───────────────────────────────────────────────
        self.cls_token = nn.Parameter(torch.randn(1, 1, d_model))

        # ── Transformer encoder ───────────────────────────────────────
        encoder_layer = nn.TransformerEncoderLayer(
            d_model=d_model, nhead=n_heads,
            dim_feedforward=4 * d_model, dropout=dropout,
            activation='gelu', batch_first=True, norm_first=True,
        )
        self.transformer = nn.TransformerEncoder(encoder_layer, n_layers)

        # ── Shared projection head ─────────────────────────────────────
        self.shared_head = nn.Sequential(
            nn.LayerNorm(d_model), nn.Linear(d_model, 128),
            nn.GELU(), nn.Dropout(dropout),
        )

        # ── Per-clock output heads (mean + log_variance) ──────────────
        self.clock_heads = nn.ModuleList([
            nn.Sequential(nn.Linear(128, 64), nn.GELU(), nn.Linear(64, 2))
            for _ in range(n_clocks)
        ])

        # ── Auxiliary head: chronological age ────────────────────────
        self.age_head = nn.Linear(128, 1)

    def forward(self, x: Tensor, mask: Tensor = None):
        B = x.shape[0]
        feat_ids = torch.arange(x.shape[1], device=x.device)
        feat_emb = self.feature_embedding(feat_ids)
        val_emb  = self.value_projector(x.unsqueeze(-1))
        tokens   = feat_emb.unsqueeze(0) + val_emb

        if mask is not None:
            missing_emb = self.missing_tokens.unsqueeze(0)
            tokens = torch.where(mask.unsqueeze(-1).expand_as(tokens),
                                 missing_emb.expand(B, -1, -1), tokens)

        pathway_emb = self.pathway_embedding(self.feature_to_pathway)
        tokens = tokens + pathway_emb.unsqueeze(0)

        cls    = self.cls_token.expand(B, -1, -1)
        tokens = torch.cat([cls, tokens], dim=1)
        out    = self.transformer(tokens)
        shared = self.shared_head(out[:, 0])

        clock_preds = [head(shared) for head in self.clock_heads]
        age_pred    = self.age_head(shared).squeeze(-1)
        return clock_preds, age_pred

Training loss:

def gaussian_nll(pred, target):
    mean, log_var = pred[:, 0], pred[:, 1].clamp(-6, 6)
    return (0.5 * (log_var + (target - mean)**2 / log_var.exp())).mean()

def total_loss(clock_preds, clock_targets, age_pred, age_target,
               clock_weights=(1.0, 1.0, 0.8, 0.8), lambda_age=0.1):
    clock_loss = sum(w * gaussian_nll(p, t)
                     for w, p, t in zip(clock_weights, clock_preds, clock_targets))
    return clock_loss + lambda_age * F.mse_loss(age_pred, age_target)

Section 3, The Full Experiment Suite

Experiment 1: Main comparison table

Run all models on same train/val/test split (80/10/10, stratified by age decade and sex). Metrics: MAE, R², Pearson r, Spearman ρ per model × clock. Target for GrimAge: r > 0.80 (prior blood-based approaches achieve ~0.65–0.75).

Experiment 2: Feature minimization curve

Start with all 42 features. Remove in reverse order of SHAP importance. Plot MAE vs. panel size. Key finding to demonstrate: "A 20-marker panel achieves 95% of full-panel performance. The 9-marker PhenoAge panel achieves only 78%." This is the clinical deployment argument.

Experiment 3: Missing data robustness

Introduce missingness at 5%, 10%, 20%, 30%, 50%, 70% under MCAR, MAR, and MNAR mechanisms. Compare learned token approach vs. mean imputation, iterative imputation, KNN, and GAIN. Expected result: learned token outperforms all imputation strategies at all missingness levels.

Experiment 4: Multi-task vs. single-task

Train 4 single-task models vs. BioAge-Transformer. Expected result: multi-task model outperforms all single-task models, especially on DunedinPACE (hardest clock to predict from blood alone).

Experiment 5: Calibration analysis

Reliability diagram: for each 10th-percentile prediction interval, what fraction of true values fall inside? Report Expected Calibration Error (ECE) per model. Clinical framing: our model's 90% intervals should contain the true value 90% of the time.

Experiment 6: External validation

Train on UK Biobank (~40K). Validate on Generation Scotland (~3K), MESA (~1.2K), Framingham (~2K). Expect 5–15% MAE increase on external cohorts. Show it's not catastrophic.

Experiment 7: Longitudinal tracking

Apply model to participants with repeat blood draws (Framingham, LBC). Does predicted biological age increase at the expected rate? Does it rise faster in people who develop disease?

Experiment 8: Interpretability

Attention weight heatmap per head. SHAP values per feature per clock. Expected biological story: GrimAge dominated by albumin, inflammation, ApoA; DunedinPACE dominated by metabolic markers. This section should read like biology, not ML.

Experiment 9: Subgroup analysis

All metrics stratified by sex, age decade, BMI, smoking status, ancestry. Check for heterogeneous performance. Investigate and fix, or disclose clearly.

Experiment 10: Comparison to existing DL papers

Directly compare against Lee et al. 2023 (Aging), Bae et al. 2021 (npj Digital Medicine), Putin et al. 2016 (Aging). Run their models on our test set. Show superiority directly.

Section 4, Figures Strategy

Figure 1, Architecture: Clean schematic of the full BioAge-Transformer. Pathway-color-coded. Should be readable by a non-ML clinician.

Figure 2, Main results: Grouped bar charts, MAE per model per clock, 95% bootstrap CIs.

Figure 3, Feature minimization curve: x = panel size, y = normalized MAE. Mark the "20-marker elbow." Add cost axis on right.

Figure 4, Calibration: Reliability diagram. Our model follows the diagonal; others don't.

Figure 5, Interpretability: (A) Attention heatmap by pathway. (B) SHAP dot plot, top 15 features per clock.

Figure 6, External validation: Performance on UKB test set vs. each external cohort, with error bars.

Section 5, Week-by-Week Timeline

Weeks 1–2: Apply for UK Biobank access (critical, 2–4 month wait). Apply MESA via dbGaP. Download NHANES (free, immediate). Set up GPU compute.

Weeks 3–6: Build full data pipeline on NHANES + GEO data while waiting for UKB. Implement all baseline models.

Weeks 7–12: UK Biobank access arrives. Calculate clock targets from methylation data in R. Extract and join blood chemistry features. Run baselines.

Weeks 13–18: Implement BioAge-Transformer in PyTorch. Optuna hyperparameter sweep (200 trials). Run all 10 experiments. Log everything in W&B.

Weeks 19–24: External validation. Interpretability analysis. Statistical analysis (bootstrap CIs). Supplementary material.

Weeks 25–30: Paper writing in LaTeX. Figure production (publication-quality matplotlib/seaborn). Submit to npj Digital Medicine.

Compute requirements

GCP a2-highgpu-1g (1× A100) = ~$3.67/hr. Full project: budget ~$500–800 compute.

Section 6, The 5 Novelty Claims

1. First deep tabular model trained to directly approximate methylation-derived epigenetic clocks from blood chemistry, PhenoAge approximates mortality risk, not GrimAge/DunedinPACE/Horvath. These are different targets.
2. Multi-task learning across 4 biological age clocks, joint training exploits shared variance and improves performance on harder clocks like DunedinPACE.
3. Principled missingness handling via learned tokens, all prior work uses imputation. Our model handles arbitrary missingness patterns without separate imputation.
4. Calibrated uncertainty quantification, clinical deployment requires confidence intervals. We provide per-sample prediction intervals that are demonstrably well-calibrated.
5. Feature minimization analysis establishing minimum effective panels, what is the minimum blood panel for each level of clock prediction performance?

Section 7, Journal Selection

Primary target: npj Digital Medicine (Nature Portfolio), IF ~15, scope exactly this, open access, 4–8 weeks to first decision.

Secondary: Nature Communications, broader scope, slightly harder.

Tertiary: Aging (Albany NY), home journal of most clock papers, IF ~4–6, more likely if external validation is limited.

Do not target Nature Medicine, Cell, or Lancet Digital Health on first submission.

Section 8, The Antiaging Labs Data Angle

"To demonstrate real-world clinical utility, we apply BioAge-Transformer to a prospectively collected cohort of n=X urban Indian professionals enrolled in the Antiaging Labs longevity program (Hyderabad, India). This cohort differs from UK Biobank in ancestry, diet, and metabolic risk profile. Despite these differences, BioAge-Transformer [maintains performance / shows X% performance drop], suggesting [generalizability beyond European-ancestry populations / the need for fine-tuning in South Asian populations]. Notably, n=Y clients had repeat blood draws at 90-day intervals following a structured lifestyle intervention, the first demonstration of surrogate clock tracking over a short-term intervention period."

Even n=20 paired before/after samples is a publishable finding if the surrogate clock moves with the intervention. Apply to an Indian IRB before collecting data for research purposes.

The 8 Papers to Read This Week

1. Levine et al., 2018, PhenoAge, Nature Medicine. Your benchmark.
2. Lu et al., 2019, GrimAge, Aging. Primary prediction target.
3. Belsky et al., 2022, DunedinPACE, eLife. Hardest prediction target.
4. Horvath, 2013, Original epigenetic clock, Genome Biology.
5. Gorishniy et al., 2021, FT-Transformer, NeurIPS. Primary architecture baseline.
6. Hollmann et al., 2022, TabPFN, NeurIPS. Second architecture baseline.
7. Lee et al., 2023, Deep learning for biological age from blood. Closest existing work.
8. Bae et al., 2021, DL biological age from CBC, npj Digital Medicine. Primary "we improve upon" citation.

Immediate Action Items

The UK Biobank application is the hardest blocker. Start today.