By Maeghal Jain
This study that tries to predict how drugs affect lifespan using gene-expression data.
Researchers combined 17,203 unlabeled transcriptional profiles from the LINCS L1000 database with lifespan measurements for just 127 compounds from DrugAge.
They trained a variational autoencoder (VAE) on the large unlabeled set to learn useful latent representations, then applied data augmentation and a simple neural network to predict lifespan change.
Current Results:
On held-out compounds the model scored R² = 0.173 and Spearman correlation = 0.164.
It performed better on compounds that extend lifespan (R² = 0.261) than on those that shorten it.
The model recovered a stronger signal on the training compounds (R² ≈ 0.72), showing the limits of the tiny labeled set.
Top predicted candidates (non-BRD compounds) include:
- glimepiride (~55 %)
- fluconazole (~37%)
- BAY-11-7082 (~32 %)
Several of these hit known aging pathways such as insulin/IGF-1 signaling, NF-κB inflammation, and HDAC activity.
Caveats
The labeled dataset is very small, and the transcriptomic data come from human cancer cell lines while the lifespan numbers mostly come from mammalian species such as Mus musculus, Rattus norvegicus, Mesocricetus auratus and non-mammalian species such as Saccharomyces cerevisiae, Acheta domesticus, Musca domestica, and Paramecium tetraurelia. Predictions remain untested in the lab.
The work shows that large-scale gene-expression signatures contain usable signal for longevity effects and supplies a shortlist of compounds worth experimental follow-up.
댓글 0