In plain English
The framework uses three checks. PACE scores ten cell-type-specific immune-aging modules; CellQ compresses each single-cell transcriptome into eight discrete tokens and predicts perturbations; and an Analyzer–Planner–Auditor agent screens compounds and revises its scoring rule when experiments expose an imbalance. The authors report a 110-compound primary-PBMC screen and follow-up in an independent 13-compound T-cell assay.
How the study worked
A plain-language walk through the work behind the result.
Built PACE from ten directionally scored immune-aging modules selected for stability across four PBMC cohorts.
Built CellQ with a multimodal language model and residual vector quantization to represent each transcriptome with eight tokens.
Used an Analyzer–Planner–Auditor agent to screen 110 compounds and revise its own scoring objective before an independent assay.
What they found
- PACE outperformed five established aging clocks in an in-house cohort of 434 elderly donors, according to the authors.
- The agent changed from an equal-weight mean to a balance-constrained minimum, and the revised objective generalized to a 13-compound T-cell assay.
Why it matters
The work proposes a concrete way to make an AI discovery loop check biological phenotype, predicted perturbation, and its own optimization rule against experiments.
The catch
- The work is a preprint and has not completed peer review.
- Claims of self-auditing and immune rejuvenation are the authors' framing, not evidence of therapeutic efficacy.
- The external compound assay contains 13 compounds, and code and data availability are not stated in the abstract.