CLIP Deep Learning - Usage Guide

Overview

Predict RBP binding from RNA sequence with deep neural networks. Modern models: RBPNet (2023, sequence-to-CL distribution at single-nt), RNAProt (RNN classifier; AUC 87-89%), GraphProt2 (GCN with structure), DeepCLIP, DeepRiPe. Use for variant-effect prediction, in silico binding-site discovery, and systematic RBP comparison. Training requires GPU; chromosome-split prevents data leakage; background must match foreground transcript region. Pretrained models exist for ~150 ENCODE RBPs.

Prerequisites

conda install -c pytorch pytorch torchvision
pip install rbpnet rnaprot graphprot2 deepclip biopython

Quick Start

Tell your AI agent: - "Predict per-base RBP binding from sequence with RBPNet" - "Train custom RNAProt on my eCLIP peaks with GC-matched 3' UTR background" - "Variant effect at heterozygous SNP using RBPNet log2 FC" - "Apply pretrained DeepRiPe to a list of GWAS variants" - "GraphProt2 with structure for RBPs that prefer specific structural contexts" - "Why is my model AUC 0.95 but predictions fail on novel sequences? Data leakage" - "Saliency map to recover motif - use TF-MoDISco for clean output"

Example Prompts

Variant-Effect Prediction

"RBPNet for TARDBP on heterozygous SNP rs12345 - log2 FC > 1 = strong effect"

"Genome-wide variant scoring for GWAS variants in eCLIP regions"

Custom Training

"RNAProt training on my new CLIP data - 50 epochs, 1:1 balanced sampling, chromosome-split"

"Use train chr1-20, test chr21-22 to prevent gene-neighbor leakage"

Production Inference

"DeepCLIP for fast inference on 1M sequences"

"RBPNet per-base score for a list of 3' UTR sequences"

Interpretation

"Saliency on trained DeepRiPe - extract motif with TF-MoDISco"

"GraphProt2 with RNAfold structure ensemble for structure-dependent RBP"

Transfer Learning

"RNA-FM foundation model fine-tuned on my eCLIP - emerging approach"

"Pretrained RBP not available - try transfer from closest paralog"

Diagnostics

"Variant effect different across windows - use model native 256 nt"

"GraphProt2 underperforms - structure ensemble not provided"

"Pretrained for related RBP works - paralog transfer valid"

What the Agent Will Do

1. Choose model: RBPNet (per-base), RNAProt (binary), GraphProt2 (structure), DeepCLIP (fast), DeepRiPe (multi-modal)
2. Verify pretrained model available; if not, train custom on user's CLIP peaks
3. For custom training: prepare GC-matched background; chromosome-split train/test
4. Apply to query sequences: per-base for RBPNet; per-sequence for binary models
5. For variant effect: predict reference + alt; compute log2 FC summed over 50 nt window
6. Interpret with saliency + TF-MoDISco for motif extraction
7. Cross-validate with eCLIP overlap, RBNS Kd, related RBP predictions
8. Flag pitfalls: data leakage, GPU memory, structure input, class imbalance

Tips

• RBPNet is the modern per-base model. Single-nt resolution.
• Chromosome-split prevents leakage. Random shuffle leaks gene-neighbor context.
• Background must match foreground region. 3' UTR peaks - 3' UTR background.
• Balanced sampling matters. RNAProt does this automatically.
• GPU required. CPU is 100x slower.
• Pretrained models are ~150 ENCODE RBPs. Not all proteins.
• Variant-effect window 256 nt for RBPNet. Native model size.
• GraphProt2 needs RNAfold pre-compute. Not automatic.
• Vanilla saliency is noisy. Use TF-MoDISco or DeepLIFT.
• Variant log2 FC > 1. Strong effect threshold.
• RNA foundation models emerging. RNA-FM, RNAErnie for transfer learning.

Related Skills

• clip-seq/clip-motif-analysis - Classical motif alternative
• clip-seq/crosslink-site-detection - CL sites for sequence-to-signal
• clip-seq/clip-peak-calling - Peak BEDs for training
• causal-genomics/mendelian-randomization - Variant-effect feeds MR
• causal-genomics/fine-mapping - Variant prioritization
• machine-learning/model-validation - Train/test methodology
• machine-learning/prediction-explanation - Saliency methods
• machine-learning/biomarker-discovery - ML framework