Alzheimer's disease affects about 50 million people worldwide, yet drugs that target its early mechanisms are almost nonexistent. Rapamycin, an FDA-approved drug, slows aging and protects neurons, but its effect is hard to assess because so many factors are in play: genetics, metabolism, inflammation, blood vessels. Traditional AI models in this field usually make a diagnosis rather than predict a response to treatment. This work stands apart from neighboring studies that focus on classifying Alzheimer's disease from MRI or genetic data, and it is the first to apply a multimodal approach to predicting rapamycin therapy specifically.
The authors collected data from 23 patients aged 40-60 in Columbia, Missouri, with two measurements each: before treatment and after. Eight data types included structural MRI, perfusion imaging, inflammation markers, blood metabolites, cerebral blood flow, microbiome, a complete blood count, and demographics. First, a random forest selected the 27 most important tabular features, keeping at least three from each modality. Then separate encoders converted each data type into embeddings, which were aligned through contrastive learning: features from the same patient moved closer together, those from different patients moved apart. Finally, a transformer integrated the embeddings and predicted treatment status. Evaluation used patient-level splits: 16 for training, 3 for validation, 4 for testing, repeated 10 times with different splits.
Without feature filtering, the model delivered about 50% accuracy, that is, chance level. After adding the random forest, mean accuracy rose to 71.25% across 10 runs. Results ranged from 50% to 87.5% depending on the split. Among the selected features, complete blood count measures (RDW CV %) and the inflammatory marker MIP-1α dominated. Blood metabolites supplied 14 of the 27 features, though on their own they were weaker. SHAP analysis confirmed that blood and inflammation contribute most to the predictions, while demographics barely matter. MRI and microbiome added moderate but useful information.
This work differs from neighboring studies in that instead of diagnosing Alzheimer's disease, it predicts the response to a specific treatment, combining eight heterogeneous data sources. Neighboring papers in the archive focus on classifying disease stages from brain scans or genetic markers, whereas this one is the first to test whether a therapy's effect can be detected in a small sample. The success of feature filtering shows that when patients are few, what matters most is not so much model complexity as cleaning the input data of noise. This opens a path toward finding biomarkers of rapamycin response and personalizing treatment.
The topic could move toward larger samples and more diverse populations, since the current patients are mostly of European ancestry. Adding longitudinal data would make it possible to model long-term effects of therapy, not just a short-term shift. Ablation studies are also possible, to compare different feature selection methods and architectures. Analyzing individual brain regions before and after treatment could reveal neuroanatomical correlates of response.
The method could be useful in clinical trials for stratifying patients and identifying those most likely to respond to therapy. It also applies to other neurodegenerative diseases and age-related disorders where systemic biomarkers matter.
The next step is validation on a larger and more diverse cohort, plus a check of whether the selected features reproduce. It would help to run an ablation to understand the contribution of each stage: filtering, contrastive learning, and the transformer.
Medical AI still struggles with small samples and high-dimensional data, where models tend to overfit. This work offers a concrete trick, filtering features with a random forest before deep learning, which could be generalized to other tasks with small patient sets.
Rapamycin was first isolated from a bacterium found in the soil of Easter Island (Rapa Nui), hence the name. The drug suppresses the mTOR signaling pathway, which regulates cell growth and aging, and extends the lifespan of mice by 10-15%.
The idea of predicting how a body will respond to a drug from a multitude of biological signals echoes Frank Herbert's novel Dune, where doctors use complex diagnostic systems to assess a person's condition. Unlike science fiction, though, this is about real blood and brain data.
Not published on arXiv. Here are its page, the PDF and the formalization: you can check it yourself, not only read our account.