RiboSense

riboswitch detection · classification

loading models

genomic riboswitch observatory

Find riboswitch domains in your sequence

Load DNA (or RNA with U→T), then watch a detector scan for riboswitch-like windows and a classifier assign one of 31 ligand families — with probabilities on every hit.

step 01 · load strand

Then open the strand theater to see detected domains locked on the helix.

how it works

Detect riboswitches, then classify the family

Both networks were trained in TensorFlow (Python), then ported for in-browser inference. A binary detector scores sliding windows; positives go to a 31-class classifier. IoU NMS keeps multiple hits on one strand.

cascade architecture

trained in TensorFlow · inferred via TF.js port

1 · SLIDING WINDOWSwindow iwindow i+1window k2 · DETECTOR3 · CLASSIFIERDetectorTensorFlow · binaryIs this window a riboswitch?P(ribo)gatethreshyesnoreject windowstop · no annotationClassifierTensorFlow · 31 riboswitch familiesP(family)TPPFMNPurineSAM…4 · MULTI-HIT OUTPUT · IoU NMSTPP riboswitchnt 1–190 · 100%Purine riboswitchnt 376–485 · 100%FMN riboswitchnt 626–810 · 100%

TensorFlow origin

Detector and classifier were trained in TensorFlow (Python). The app loads the ported weights locally.

Cascade gate

Only windows that clear the detector threshold continue. Rejects stop here — no family call.

Multi-hit NMS

Overlapping windows are merged with IoU non-max suppression so one strand can carry several families.