ADAPT Model
Two-stage CNN from the Broad Institute for predicting Cas13 guide RNA activity.
Overview
ADAPT (Activity-informed Design with All-inclusive Patrolling of Targets) is a model developed at the Broad Institute for predicting Cas13 guide RNA activity. It was designed specifically for diagnostic guide design and trained on large-scale Cas13a (LwaCas13a) activity screens against diverse viral sequences.
Metsky HC, Welch NL, Pillai PP, Haradhvala NJ, Rumker L, Mantena S, Zhang YB, Yang DK, Ackerman CM, Weller J, Blainey PC, Myhrvold C, Mitzenmacher M, Sabeti PC (2022). “Designing sensitive viral diagnostics with machine learning” Nature Biotechnology.
SPACER automatically selects the ADAPT model when AI activity prediction is enabled for Cas13 analyses.
Model Architecture
Like EasyDesign, ADAPT uses a two-stage CNN cascade:
- Stage 1 (Classification): Predicts whether a guide is active or inactive
- Stage 2 (Regression): For guides classified as active, predicts a continuous activity score
Input Encoding
| Parameter | Value |
|---|---|
| Guide length | 28 nt (fixed for this model) |
| PAM | Not required (Cas13 is PAM-independent) |
| Flanking context | 10 nt upstream + 10 nt downstream |
| Total input length | 48 nt (10 context + 28 guide + 10 context) |
| Encoding | Paired one-hot: 8 channels per position (4 target + 4 guide, no gap channel) |
| Tensor shape | 48 × 8 (positions × channels) |
Scoring Integration
When AI activity prediction is enabled for a Cas13 analysis, SPACER runs ADAPT on each candidate guide. The predicted activity score is incorporated into the assay score via the ML adjustment component.
Performance
| Metric | Value |
|---|---|
| Applicable enzyme | Cas13 |
| Required spacer length | 28 nt |
| Output range | Continuous activity score |
| Inference | Batched via ONNX Runtime |