In a lab where patient data resemble a scattering of gold dust, every test is a speck, and Alzheimer's disease is a river washing away the precious metal. Scientists set a task for the neural network: to tell whether a person received rapamycin, a drug that slows cellular aging, or not. But the sand is mixed with waste rock: noise in the tests, variation between people, the weak effect of therapy. To find the nuggets, you first have to discard the excess.
The TreatmentFormer model works like a multi-stage washing process. First, a "random forest" — an ensemble of decision trees — sifts through eight types of data: brain scans, inflammation markers, blood metabolites, the microbiome, and others. Out of hundreds of features, the 27 most valuable remain, three from each modality. Then separate encoders turn each data type into compact embedding vectors, and contrastive learning brings vectors from the same patient closer together while pushing others apart. Finally, a transformer assembles everything and gives an answer: treatment or no treatment.
The result was not perfect, but significant. Without filtering, the model guessed randomly — around 50%. With filtering, accuracy rose to 71.25% on average across ten runs, though the spread ranged from 50% to 87.5%. SHAP analysis showed that the main contribution comes from complete blood count parameters, especially RDW CV %, and the inflammatory marker MIP-1α. Blood metabolites provided 14 of the 27 features, but individually they were weaker. Demographics hardly mattered, while MRI and the microbiome added moderate benefit.
Before this work, it was unclear whether a small group of patients could reveal the drug's faint trace in the noise. Now it is clear: the key is not in the complexity of the model, but in cleaning the input data. The method was compared on the same data with and without filtering, and feature importance was checked in two independent ways. Already today, the approach can be used in clinical trials to select patients most likely to benefit from rapamycin and to search for response biomarkers.
The grains of sand glitter in the pan again, but now it is clear which of them are gold. The neural network does not just guess — it shows where to look.
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.