Notes
A tumor RNA-seq profile carries readings for roughly 18,000 genes. The obvious move is to feed all of them to the model and let it sort things out. We do not. Provotics reads a curated panel of 3,882 genes, and the input is smaller on purpose. More columns are not more information; past a point, they are more noise.
The signal that tells you where a cancer started lives in a minority of genes. The rest are a mix of housekeeping expression, biological noise, and measurement artefacts that drift with how a sample was handled and sequenced. Hand all of it to a model and it does not just learn the signal; it also learns the noise, including the parts of the noise that are specific to the labs in the training data. That is exactly what you do not want it memorising.
Gene expression is heavily correlated. Thousands of genes move together, so the full transcriptome contains far less independent information than its width suggests. Throw 18,000 correlated, partly noisy features at a model and you give it endless ways to find patterns that hold on your data and nowhere else. Curating down to a panel that actually separates the sites removes most of that temptation before training even begins.
The real test is not the data you trained on; it is a sample from a centre you have never seen, run on a pipeline you do not control. A focused panel of informative genes is more stable across those shifts than the full transcriptome, because the genes that genuinely mark a site of origin behave more consistently than the long tail of noisy ones. Fewer, better genes are part of why the model holds up off its home turf instead of quietly falling apart.
A good panel does not make every site easy. Some anatomical sites are rare, with far fewer training profiles than the common ones, and no panel can conjure signal that the data barely contains. So on those, the honest behaviour is restraint. When a profile sits at the edge of what the model has learned, Provotics abstains and flags it rather than guessing, the same conformal behaviour that governs any unfamiliar input.
The 3,882-gene panel is a deliberate trade. It gives up the illusion that more inputs mean a better model, and in return it reads cleaner, travels further, and is more candid about what it cannot do. That is the version of the system we would rather put in a researcher's hands.
Provotics is a research and educational project. It is not a medical device and is not intended for clinical diagnosis or treatment decisions.
← All notes