← Về thư mục
📄 / / profiles / r-and-d / raw_evidence / OpenAI4S / skills / bioskills / bio-alternative-splicing-single-cell-splicing / references / usage-guide.md

Single-Cell Splicing - Usage Guide

Overview

Analyze alternative splicing at single-cell resolution. The first decision is library chemistry - 10X 3' is fundamentally limited because RT primes from poly(A) and >70% of reads land in the 3' UTR. Plate-based full-length methods (Smart-seq3, FLASH-seq, VASA-seq, STORM-seq) and single-cell long-read (MAS-Iso-seq, scISOr-Seq2) are the chemistries that give per-cell isoform structure. Tools include MARVEL (R, Smart-seq), BRIE2 (Bayesian PSI with regulatory features), scQuint (junction-cluster, plate-based; not for 10X), SpliZ (annotation-free Z-score), Psix (graph-smoothness regulated AS), and Sierra (alternative polyadenylation, often confused with AS).

Prerequisites

# Python
pip install brie scanpy anndata scipy

# R - both from GitHub (Sierra not on Bioconductor; MARVEL archived from CRAN 2025-10)
# devtools::install_github('VCCRI/Sierra')
remotes::install_github('wenweixiong/MARVEL')

# scQuint, SpliZ, Psix from GitHub
pip install git+https://github.com/songlab-cal/scquint
pip install git+https://github.com/salzman-lab/SpliZ
pip install git+https://github.com/lareaulab/psix

Quick Start

Tell your AI agent what you want to do: - "Will my 10X 3' data support splicing analysis?" - "Analyze splicing in Smart-seq2 plate-based scRNA-seq with MARVEL" - "Find cell-type-specific splicing without an event database using SpliZ" - "Detect alternative polyadenylation in 10X data with Sierra" - "Pseudobulk single cells by cell type and run leafcutter for differential splicing"

Example Prompts

Chemistry Audit

"I have 10X Chromium 3' v3 data; can I do splicing analysis? What can I do instead?"

Plate-Based Workflow

"Run MARVEL on Smart-seq2 BAMs to quantify SE/A5SS/A3SS PSI, classify modality, and test differential splicing between cell types."

"Use BRIE2 with sequence-feature prior to estimate per-cell PSI for cassette exons in low-coverage Smart-seq3 data."

Discovery

"Run SpliZ to find genes with cell-state-associated splicing without using an event database."

Trajectory

"Use Psix to detect regulated AS along a developmental pseudotime trajectory, robust to dropout."

APA (10X 3')

"Use Sierra to peak-call 3' ends and detect alternative polyadenylation across cell types."

Pseudobulk

"Aggregate cells by cluster and run leafcutter or rMATS on pseudobulk junction counts for differential splicing between cell types."

Long-Read Single-Cell

"I have MAS-Iso-seq + 10X 5' data; demultiplex barcodes with skera/lima, then run FLAMES for full-length isoform quantification per cell."

What the Agent Will Do

  1. Audit the library chemistry and recommend tool subset based on what's recoverable
  2. For full-length plate data: quantify per-cell PSI; classify modality; test cell-state association
  3. For 10X 3' data: redirect to APA (Sierra). Note: scQuint authors recommend against use on 10X data (3' bias confounds junction-cluster modeling)
  4. For long-read single-cell: demultiplex, align with minimap2 splice:hq, quantify with FLAMES/Bambu/IsoQuant
  5. Recommend pseudobulk for between-cell-type comparisons; per-cell only for within-cluster heterogeneity
  6. Avoid PSI imputation - use graph-smoothness (Psix) or hierarchical Bayes (BRIE2) instead

Tips