SPACER

AI Activity Prediction

Optional machine learning models that predict guide cleavage activity from sequence features.

Overview

AI activity prediction is an optional scoring component that uses pre-trained neural network models to estimate the likelihood of efficient target cleavage based on spacer sequence features. When enabled, the predicted activity score is incorporated into the composite score. SPACER automatically selects the correct model for the chosen enzyme family.

Info
AI activity prediction is available for both Cas12 (via the EasyDesign model) and Cas13 (via the ADAPT model). SPACER automatically selects the appropriate model based on the enzyme.

Supported Models

SPACER integrates two published prediction models, each targeting a different enzyme family:

ModelEnzymeSourceArchitecture
EasyDesignCas12Wei et al., 2024Two-stage CNN cascade (classification + regression) on one-hot encoded sequences
ADAPTCas13Metsky et al., 2022Two-stage CNN cascade (classification + regression) on one-hot encoded sequences with flanking context

EasyDesign Model (Cas12)

The EasyDesign model predicts Cas12 guide activity using a two-stage convolutional neural network trained on LbCas12a activity data. The input is:

  • 21 nt spacer sequence + 4 nt PAM (prepended internally)
  • 10 nt flanking context on each side of the target
  • One-hot encoding with 10 channels per position (5 target + 5 guide, with gap support)

The first stage classifies guides as active or inactive; the second stage performs regression on the active class to produce a continuous activity score.

ADAPT Model (Cas13)

The ADAPT model uses a two-stage convolutional neural network that takes the spacer sequence plus flanking context as input. It was trained on Cas13a (LwaCas13a) activity data and uses:

  • 28 nt guide sequence (no PAM required)
  • 10 nt flanking context on each side of the target
  • One-hot encoding with 8 channels per position (4 target + 4 guide)

Score Normalization

Both models produce raw predictions on their own internal scales. SPACER normalizes these to a comparable range so that the activity score contributes proportionally to the composite score regardless of which model is used.

When to Enable AI Prediction

AI activity prediction adds computational cost but can significantly improve guide ranking accuracy. Consider enabling it when:

  • You are designing guides for a diagnostic assay where sensitivity is critical
  • You have many candidate guides and need to differentiate between otherwise similar ones
  • Your target is a viral genome where activity data is available in the training set
Warning
When AI activity prediction is enabled, SPACER may constrain the spacer length to match the model's training input size (21 nt for EasyDesign/Cas12, 28 nt for ADAPT/Cas13) to ensure prediction accuracy.