e.g.
ROI_time_series_S25 — all 12 wells of one sample
Frames are matched by file name, A015 - 20260913_223113_roi_3.bmp = frame 15 of well 3. Only frames A015–A029 of wells roi_1–roi_12 are read; other files in the folder are ignored. Previews are brightened for display only.
Select the 15 frames (140×140 BMP) of one well at once, frames 15–29. They are sorted by file name. Tap a cell to replace one frame.
Each well is read from 15 fluorescence frames (frames 15–29, about 2:20–4:40 after the reaction starts, one frame every 10 s; 140×140 RGB ROI). Two separately trained networks, a CNN–LSTM and a CNN–Transformer, each read the frames under 4 flips; their outputs are averaged into 4 class probabilities (NC, 10 aM, 100 aM, ≥1 fM) and a log10 concentration.
The model was trained on laboratory dilution standards imaged on one instrument. Images from other instruments, other exposure settings or other frame timing are outside what it was validated on.
Authors: Minjie Pei, Shiyu Xie.
Model weights and code: huggingface.co/shiyuxie/crispr-kinetics-hpv16. The model runs inside your browser: images are not uploaded anywhere.