Researchers introduce STUART, a new machine‑learning system that aims to quickly identify signs of ionizing radiation exposure directly from raw sequencing reads. Unlike usual methods that first match reads to a reference genome (a slow step called alignment), STUART converts raw reads into short sequence fragments and uses those counts as numerical fingerprints. This mapping‑free approach is meant to cut computation and enable on‑site, real‑time triage with portable sequencers.

The pipeline works in two main stages. First, each RNA read is broken into overlapping 5‑letter pieces called k‑mers. Counting those k‑mers produces a “Bag of Words” style profile (the authors call this an NLP‑profile because it borrows ideas from natural language processing). A fast gating tool (noMapper) then filters reads into those “potentially related” to a chosen biomarker and those that are not. A set number of the potentially related reads are summed and scaled to make a single fingerprint. That fingerprint is fed to a machine‑learning predictor (the team tested Logistic Regression, Random Forest, and XGBoost) to decide whether a sample shows exposure.

The team tested STUART on Nanopore direct long‑read RNA data from two human cancer cell lines. They used fibrosarcoma cells (HT1080) with and without radiation (measured 24 hours after exposure) to train models. An independent liver cancer dataset (HepG2) served as external validation. Using feature‑selection methods, they shrank the initial 1024‑feature k‑mer space by over 98%, keeping as few as 17 sequence signatures. With those reduced signatures the models reached very high performance in their tests, reporting balanced accuracy and F1 scores in the 95–100% range. Their learning‑curve analysis also showed that stable predictions required aggregating only about 1,000 potentially related reads. External testing on the HepG2 data gave more than 99% specificity, which the authors cite as evidence the signatures are not tied to a single tissue type.