Rishav Sen
← All work
Life sciences & ecology · Complete

CRISPR

Guide RNA efficiency prediction with a convolutional network, attention and calibrated uncertainty

N
Headline result
The network's score against the Rule Set 2 baseline, and how often the confidence range held.

The question

Can a neural network predict how well a CRISPR-Cas9 guide RNA edits its target, and say how far to trust each prediction?

The method

I first rebuilt a Rule Set 2 style baseline with gradient-boosted trees and checked its score against published numbers before building anything harder. The network reads the 23-letter guide sequence through two convolution layers, then a self-attention step that shows which positions matter most. It combines that with sequence features such as melting temperature. Split conformal prediction then wraps each prediction in a range, and a calibration plot checks that the range holds as often as promised.

Tools and methods

Tools
PythonCNNAttentionConformal prediction
Methods
Convolutional neural networkSelf-attentionRule Set 2 baselineSplit conformal prediction

The evidence

Open On-target report (PDF)

Limits

It uses one dataset (Doench 2016) and predicts on-target efficiency only, not off-target edits. Scores come from one assay, so they may not carry over to other cell types.