TranscriptML

TranscriptML provides data preparation, model training, evaluation, and interpretation tools for RNA sequence-to-function models. It is designed for two common starting points:

• transcript-level measurements paired with annotated transcript sequences, modeled with Saluki; and

• MPRA measurements paired with one variable RNA insert per construct, modeled with LegNet.

Interpretation tools include single-nucleotide in silico mutagenesis (ISM), region/junction ablations, motif ablations, motif context scans, motif epistasis analyses, and Saluki-specific codon ISM. These analyses can expose learned regulatory sequence features as well as technical artifacts in the model or assay.

RBPNet/eCLIP preprocessing, descriptive scanning, region selection, materialized dataset construction, and structured profile/enrichment training are supported.

Warning

The entire RBPNet/eCLIP workflow is experimental, including preprocessing, scanning and selection, bundle construction, modeling, and evaluation. It has been minimally tested and has only been confirmed to preprocess data successfully and train reasonable models on PUM2 eCLIP data. It needs substantially more validation than other TranscriptML functionality.

Start here

See Installation for the smallest install that covers your use case. The Usage guide walks through Saluki and MPRA projects from input tables through cross-validation and interpretation. The Training Configuration guide describes every Saluki, LegNet, and shared training option. Use the API reference reference when calling TranscriptML from Python.

Contents

• Installation
• Usage
  • Saluki Workflow
  • MPRA-LegNet Workflow
  • Other Useful Commands
• RBPNet/eCLIP data workflow
  • 1. Canonical preprocessing
  • 2. Descriptive window scanning
  • 3. Region selection
  • 4. Materialized RBPNet bundle
  • 5. RBPNet model and training
• Training Configuration
  • Saluki Starter Configuration
  • MPRA-LegNet Starter Configuration
  • RBPNet Starter Configuration
  • Top-Level Training Settings
  • Saluki Model Parameters
  • MPRA-LegNet Model Parameters
  • Loss Configuration
  • Split Configuration
  • Sequence Controls
  • Using These Settings On Sherlock
• API reference
  • Data
  • RBPNet/eCLIP data
  • Models
  • Training and evaluation
  • Interpretation
  • Plotting
  • Run setup