During the 2 years of the MSC action, I worked on developing NUTRIclock. First, the available metagenomics and metabolomics datasets useful for its development were identified. Since it’s rare to find both types of datasets for the same volunteer group in human studies, the nutrients’ relative abundances were inferred from metagenomic data. The metagenomic data source used was the R package CuratedMetagenomicDataset, which contains metagenomic datasets processed using the Metaphlan3 pipeline to ensure consistency. Data was filtered, preprocessed and studied preliminary using different R packages (e.g. MMUPHin, MaAsLin2, MelonnPan). The final samples obtained belong to 3,664 healthy patients, aged 18 and over, assessing the batch effect among different studies and retaining bacterial data present in at least 10% of the samples and with a minimum relative abundance of 1e-5. This data revealed an increase in alpha diversity with age but no changes in beta diversity. Metabolites information was inferred using microbiome data and aging-related microbiome species along with metabolites were identified.
Next, different architectures and combinations of data were tested to obtain the best balance between age prediction and performance of the algorithm. The following architectures were used: I) Multilayer perceptron NN (MLPNN), a classical NN approach with hidden layers extracting information from input data. II) Convolutional NN (CNN), specific for images treatment. III) Variational autoencoders (VAE), architectures that extract information of the input data generating a latent representation that might be used to reconstruct the original data. These architectures were used on microbiome relative abundances, inferred metabolites relative abundances and/or metabolites-related pathway abundances. The architecture with the better prediction/performance balance was a CNN + VAE on microbiome data combined with MLPNN + VAE on inferred metabolites data.
Finally, the NUTRIclock algorithm was tested on samples from a diseased population and from a nutritional intervention study. I) The algorithm showed increased prediction of biological age in patients with age-related diseases like diabetes and cancer. This was the expected result, whereas they are considered “older”, and confirmed that the algorithm works. The effect was particularly increasing with the severity of the conditions. II) When applied to a nutritional intervention study in overweight/obese patients, it showed a discreet rise in biological age for non-responders and no changes for responders, aligning with previous results given the short duration of the intervention (3 weeks) and mild responder criteria (only 2kg weight loss).
