Machine learning predicts lifespan-extending compounds in C. elegans
In the abstract, supervised machine learning using compound-protein interaction features from the STITCH database predicted lifespan-extending compounds in the nematode C. elegans. How the features were represented changed both accuracy and which proteins the models ranked as important. [1]
Automatic summary · not editorially reviewed
The key findings
The study tested supervised machine learning to predict lifespan-extending compounds in C. elegans using compound-protein interaction features from STITCH. [1]
Three feature representations were compared, and Random Forest and XGBoost were evaluated with stratified 10-fold cross-validation. [1]
Feature representation influenced both predictive performance and model interpretation, so metrics and explainability should be evaluated together. [1]
The source at a glance
- Study type / format
- Computational modelling study using supervised machine learning with stratified 10-fold cross-validation (abstract).
- Population / context
- Compounds and proteins represented in the STITCH database, evaluated for predicted lifespan extension in the nematode Caenorhabditis elegans; no human population.
- Summary based on
- Study abstract
- Publication status
- Preprint · peer-review status not confirmed
- Original publication
- October 19, 2026
Predicting lifespan compounds in C. elegans with machine learning
Finding compounds that extend lifespan is a key goal in ageing research, but experimental screening is costly and slow. This abstract describes a study testing supervised machine learning to predict lifespan-extending compounds in the nematode Caenorhabditis elegans. The models used compound-protein interaction features taken from the STITCH database.
Supporting source [1]
Three feature types compared with two classifiers
The researchers compared three ways of representing the interaction features: continuous interaction probabilities, binary features made with a fixed threshold of 0.5, and binary features made with a threshold tuned inside cross-validation. They tested Random Forest and XGBoost classifiers using stratified 10-fold cross-validation. Performance was measured by accuracy, area under the receiver operating characteristic curve (AUROC) and area under the precision recall curve (AUPRC). They also examined which proteins the models ranked as important, using several interpretation methods, and compared the top-ranked proteins between methods.
Supporting source [1]
Feature choice shifts accuracy and protein rankings
Random Forest reached its highest accuracy (78.84%), AUROC (0.7391) and AUPRC (0.5202) with the fixed-threshold binary features. XGBoost reached its highest accuracy (80.36%) with the optimised-threshold features, but its highest AUROC (0.7605) and AUPRC (0.5332) with continuous features. Two interpretation methods, SHAP and TreeSHAP, agreed most strongly, giving identical top-10 protein sets for all three XGBoost representations. The authors conclude that feature representation affects both prediction and interpretation, so classification metrics and explainability should be assessed together when prioritising candidate compounds. These are model results in a worm, not evidence of human lifespan benefit.
Supporting source [1]
How to interpret this
Abstract only; the full text was not reviewed. The abstract reports model performance in C. elegans, not experimental lifespan results or human outcomes. Details such as dataset size, compound counts and external validation are not reported here.
Based on the study abstract; the full paper, tables and supplementary data were not reviewed.
Source & supporting passages
Acharya DD, Freitas A. · europepmc.org · October 19, 2026