Precious1GPT

An aging clock returns a number. A methylation array goes in, a predicted age comes out, and the gap between that prediction and the person's birthday is the signal everyone cares about. What it almost never returns is a gene worth drugging: the elastic-net regressions that dominate the field pick a few hundred CpG sites and stop there, and each clock is welded to the one assay it was fit on.

Precious1GPT, published by Insilico Medicine in Aging in June 2023, treats that as an architecture problem. A single tissue-agnostic regressor consumes either promoter-averaged methylation β-values or bulk expression TPMs, with the data modality and the tissue supplied as categorical inputs alongside the numeric omics vector, so both assays are learned into one representation. The interesting move comes after training: the authors froze every weight except the final layer and re-fit the model as a case-control classifier for idiopathic pulmonary fibrosis, COPD, Parkinson's disease and heart failure. Running SHAP over that classifier ranks genes by relevance to aging and to a specific disease at once, which is the output a target-discovery pipeline can actually consume.

Despite the name, nothing here is autoregressive or generatively pretrained — the model is a supervised tabular transformer, and "GPT" is branding carried across Insilico's PreciousGPT lineup. The lineage becomes a genuine language model only at its successors, Precious2GPT and Precious3GPT, architecturally distinct models rather than retrained versions of this one.

#Key Features

• One clock across two omics layers: methylation and expression profiles pass through the same regressor, with tissue and data type as categorical features, instead of one model per assay.
• Frozen-backbone transfer to a new task: all weights but the last are held fixed while the model is re-fit as a disease classifier, and for several of the diseases the resulting classifiers modestly but significantly beat the same architecture trained on the complete feature set.
• Interpretability as the deliverable: SHAP over the regressor ranks genes for aging relevance and over the classifiers for disease relevance, making attribution the product rather than a diagnostic.
• Tissue-agnostic training: one model spans all GTEx tissues rather than the per-tissue clocks that fragment the field.
• Validation on data it never saw: the trained regressor was applied without refitting to two GEO methylation datasets, recovering the expected trajectory in both.

#Technical Details

The backbone is a TabTransformer implemented through PyTorch Tabular, with hyperparameters chosen by Optuna: hidden layers of 128, 2048 and 128 units, ELU activations, dropout 0, learning rate 0.00023, AdamW with no weight decay, batch size 96, MSE loss. Training used 8,374 Illumina 450k methylation samples from EWAS Data Hub — β-values normalized with GMQN, then averaged across probes annotated to the TSS200 proximal-promoter region, leaving roughly 14,000 features — plus 12,453 GTEx RNA-seq samples as protein-coding TPMs, ComBat-corrected and quantile-normalized. On a 20% hold-out split stratified by tissue, the regressor reached MAE 4.227 and R² 0.934 on the 4,019 methylation samples, MAE 6.287 and R² 0.584 on the 2,730 expression samples, and MAE 5.622 and R² 0.807 across all 6,749. The authors are explicit that this is below the state of the art for clocks specialized to a single data type, and argue the model's value lies in target discovery instead. Applied without refitting to GSE54848 (iPSC reprogramming) and GSE76641 (fetal fibroblasts), it showed iPSCs growing younger through induction and fetal tissue older through gestation. The paper reports no total parameter count, and neither code nor weights have been released.

#Applications

The intended workflow is target triage in aging biology: fit the clock once, freeze it, adapt it per disease, and read out ranked gene lists that carry aging context. For the four diseases studied, the top-200 expression-classifier genes fed into Insilico's PandaOmics TargetID platform, which surfaced APLNR as a top-20 target across all four and IL23R across three. The case-control cohorts and the approved-drug/target annotation both come from proprietary Insilico resources, so the downstream half of the pipeline is not reproducible from public data; the clock itself trains entirely on public corpora.

#Impact

Precious1GPT is the first entry in a lineup Insilico has continued to build on, and its contribution is less a benchmark than a design pattern: treat the aging clock as a transferable representation rather than a scalar predictor, then mine the frozen backbone for biology. That a layer trained only to predict chronological age retained enough signal to separate cases from controls in four unrelated diseases is the load-bearing evidence. The practical limits are real too: accuracy trails single-modality clocks, the input space is fixed to one array platform's promoter aggregates or GTEx-style expression vectors, the highlighted targets were nominated computationally rather than validated experimentally here, and with no released checkpoint or code the model cannot be run or audited outside Insilico.