PAIRNet Web Tool: User Guide & Documentation

1. Overview

The PAIRNet Cleavage Predictor is a deep-learning tool designed to predict the cleavage efficiency of piRNA (guide) and target RNA pairs.

Unlike traditional thermodynamics-based predictors, this tool uses an ensemble of 10 neural networks trained on high-throughput experimental cleavage data. It provides a Relative Cleavage Score, comparing your specific target variant against a theoretical "Perfect Match."

2. Input Requirements

To ensure accurate predictions, please strictly follow these formatting rules. The model is sensitive to sequence length and directionality.

• Guide RNA Sequence (piRNA)
  • Length: Must be exactly 26 nucleotides.
  • Direction: 5'3' (Left to Right).
• Target RNA Sequence
  • Length: Must be exactly 26 nucleotides.
  • Direction: 3'5' (Left to Right).
  • Note: This is the reverse of standard gene listings. Ensure your target is oriented to pair antiparallel to the guide.
• Pairing Mode String

  This string describes the specific mismatch or pairing architecture. It must follow the strict format:

  g[Start]:g[End]:[Count]mm:[Position]:[Mutation]

  Breakdown of the format:

  • g2:g26: Indicates the region of the guide involved in pairing (usually bases 2 through 26).
  • Xmm: The number of mismatches (e.g., 1mm for one mismatch, 0mm for perfect match).
  • Position: The position of the mutation relative to the 5' end of the guide (e.g., 10).
  • Mutation: The specific base change (e.g., AU means an A in the wild type was mutated to U).

  Example: g2:g26:1mm:10:AU

  Interpretation: Pairing occurs from g2 to g26; there is 1 mismatch located at position 10; the mutation represents an A U change.

3. Demo: Try It Yourself

If you are new to the tool, try these inputs to see how the prediction works.

Scenario: A Single Point Mutation at Position 10

• Guide RNA (5'->3'): TGAGGTAGTAGGTTGTATAGTATCCA
• Target RNA (5'->3'): TGGATACTATACAACCGACTACCTCA
• Pairing Mode: g2:g26:1mm:10:AG

Step-by-Step:

• Copy the Guide and the Target (Variant) into the boxes.
• Enter the pairing mode string.
• Click PREDICT.

Other examples to try:

• Mutation at g10

  TGAGGTAGT A GGTTGTATAGTATCCA 5->3
  TGGATACTATACAACC G ACTACCTCA 5->3
  g2:g26:1mm:10:AG
  ==================================
  ENSEMBLE RESULTS (Median of 10 models)
  RELATIVE K RATIO:    0.1641
  ==================================================

• Mutation at g3

  TG A GGTAGTAGGTTGTATAGTATCCA 5->3
  TGGATACTATACAACCTACTACC A CA 5->3
  g2:g26:1mm:3:AA
  ==================================
  ENSEMBLE RESULTS (Median of 10 models)
  RELATIVE K RATIO:   0.5697

• Mutation at Position 20 (3' Region)

  TGAGGTAGTAGGTTGTATAG T ATCCA 5->3
  TGGAT G CTATACAACCTACTACCTCA 5->3
  g2:g26:1mm:21:UG
  ==================================
  ENSEMBLE RESULTS (Median of 10 models)
  RELATIVE K RATIO:   0.9675

4. How to Interpret the Results

The tool outputs two main values: the Relative Cleavage Rate and the Confidence Interval.

• Relative Cleavage Rate (The Ratio)

  Because absolute cleavage rates ($k_{obs}$) can vary between experiments, we report a Ratio relative to a perfect match.

  $$Score = \frac{Predicted\ \text{ }k\ \text{ }of\ \text{ }Your\ \text{ }Target}{Predicted\ \text{ }k\ \text{ }of\ \text{ }Perfect\ \text{ }Match}$$
  • Score≈1.0: High Efficiency. Your target is cleaved as effectively as a perfect Wild Type match.
  • Score≈0.5: Medium Efficiency. The mismatch has reduced cleavage activity by roughly 50%.
  • Score<0.1: No/Low Cleavage. The mismatch effectively abolishes cleavage.
  • Score>1.0: Enhanced Cleavage. In rare cases, certain geometries may cleave slightly faster than the canonical perfect match.
• Confidence Interval (Quantiles)

  We use an Ensemble of 10 Models to make every prediction.

  • Narrow Range (e.g., 0.45 - 0.55): All 10 models agree. The prediction is highly reliable.
  • Wide Range (e.g., 0.20 - 0.80): The models disagree. This usually happens if the input sequence is very unusual or falls outside the training data distribution (e.g., highly G-rich sequences or clustered mismatches).