Benchmark: pre-transformer models vs transformers for biological-age clocks
In this abstract, eight model classes predicted biological age across six public cohorts; pre-transformer models won on five cohorts and tied on the sixth, and the lowest-error model had the weakest mortality association. [1]
Automatic summary · not editorially reviewed
The key findings
Eight model classes were benchmarked for biological-age prediction across six public cohorts spanning five modalities, all tuned and evaluated under one nested cross-validation protocol on identical splits. [1]
Pre-transformer models won outright on five of six cohorts and essentially tied on the sixth; from-scratch transformers never won. [1]
The model with the lowest chronological-age error, the pretrained transformer, carried the weakest mortality association, and both transformers ranked at the bottom for risk stratification. [1]
The source at a glance
- Study type / format
- Benchmark study (abstract)
- Population / context
- Six public cohorts spanning five data modalities used for biological-age prediction.
- Summary based on
- Study abstract
- Publication status
- Preprint · peer-review status not confirmed
- Original publication
- July 22, 2026
Eight model classes across six cohorts
The abstract describes a benchmark of eight model classes for biological-age prediction in six public cohorts covering five modalities: blood DNA methylation (GSE40279), blood biochemistry (NHANES), whole-blood transcriptome (GTEx), gut microbiome (American Gut), wearable accelerometry (NHANES) and brain MRI (Cam-CAN). All models were tuned and evaluated under one nested cross-validation protocol on identical splits. Scoring covered accuracy, data efficiency, compute, interpretability, cross-cohort transfer and the biological validity of each model's age-acceleration residual against all-cause mortality.
Supporting source [1]
Pre-transformer wins; lowest error, weakest mortality link
Pre-transformer models won outright on five of six cohorts and essentially tied on the sixth. On tabular molecular omics, penalized regression and gradient boosting were best or tied-best (methylation mean absolute error 3.4 years, transcriptome 8.1 years); on structured modalities the modality-native pre-transformer deep net won (accelerometry 12.7 years, brain MRI 4.7 years). From-scratch transformers never won. A pretrained transformer edged ahead only on the single large cohort (n=34,000), by 0.10 years, and only past about 23,000 labelled samples, at three to four orders of magnitude more compute. The lowest-error model (the pretrained transformer) had the weakest mortality association, and both transformers ranked at the bottom for risk stratification (hazard ratio about 1.13 per standard deviation versus about 1.28 for elastic net).
Supporting source [1]
Transfer advantage of penalized regression
Penalized regression transferred across cohorts best of all, with a degradation of 1.2 years, while the from-scratch and non-pretrained deep models degraded more than twice as much (2.8 to 4.0 years). The abstract concludes that for the low-sample, high-dimensional, interpretability-sensitive and regulator-facing setting of ageing research, the pre-transformer toolkit is a rational default and transformers are a targeted addition, not a replacement.
Supporting source [1]
How to interpret this
This is an abstract; the full text was not reviewed. Cohort details, tuning choices and the mortality analyses are reported only in summary form, and the findings concern model performance in these public cohorts, not causal effects on ageing or lifespan.
Based on the study abstract; the full paper, tables and supplementary data were not reviewed.
Source & supporting passages
Feng J. · europepmc.org · July 22, 2026