Skip to content

SEMPLR

SNP Effect Matrix Pipeline in R — a Bioconductor package for predicting transcription-factor-binding effects of genetic variants using pre-computed SNP Effect Matrices (SEMs); ships 223 built-in matrices and supports custom matrices, with enrichment and variant-comparison utilities on top.

  • Maintainer: Grace E Kenney (Boyle / Phanstiel labs, University of Michigan) — kenney.grace6@gmail.com
  • Bioconductor: SEMPLR v1.0.0
  • Source: github.com/grkenney/SEMPLR · project site
  • Vignettes: SEMPLR intro
  • License: MIT
  • Status at GBCC2025: methods talk on TF-binding variant effects (first major release — Bioc 3.23, v1.0.0)

Talk at GBCC2025

  • Speakers: Grace E Kenney, Rintsen N Sherpa, Alan P Boyle, Douglas H Phanstiel (University of Michigan)
  • Day / session: Day 2 (Tue Jun 24, 2025) — Oral Session 2B, Grace Auditorium (chair: Vincent Carey)
  • Talk title: "SEMplR: An R package for predicting the effects of genetic variation on transcription factor binding"
  • Slides (Zenodo DOI): TBD — Zenodo deposits typically 2–4 weeks post-conference
  • Video: GBCC2025 playlist
  • Abstract / schedule: GBCC2025 program

What it does

SEMPLR scores genomic locations and SNPs against SNP Effect Matrices — a per-TF representation of binding affinity that captures position-specific tolerance for substitutions. With 223 pre-computed SEMs in the box, a user can take a list of variants (GWAS hits, somatic mutations, eQTL fine-mapping outputs) and ask which TFs lose or gain binding at each site, plus look at TF-specific enrichment. The companion data package SEMplR_data (github.com/rnsherpa/SEMplR_data) hosts the raw matrices.

How it works

Core idea. A SNP Effect Matrix is a position-by-nucleotide table of binding-affinity scores rather than the frequency-derived probabilities used in a PWM/PFM. Each cell encodes the relative binding energetics for a nucleotide at that position, so SEMPLR slides the SEM across both strands of a variant's flanking sequence and reports the best-scoring window per TF.

Inputs / outputs. Input is a set of genomic coordinates (e.g. a GRanges of variants) plus a reference genome (BSgenome.Hsapiens). Output for each variant × TF pair is a raw score, a baseline-normalized scoreNorm (the binding vs. non-binding signal), and the index of the best-matching subsequence. Enrichment outputs are TF-level p-values across a query set.

Key innovation. Energy-based affinity scoring rather than frequency-based probability — SEMs are derived from experimental binding measurements, so they distinguish "frequent" from "preferred" nucleotides, which matters when scoring single-base substitutions that PWMs treat as roughly equivalent. Shipping 223 pre-computed SEMs makes it usable out of the box.

Parameters worth knowing. - nFlank — nucleotides of flanking sequence around each query position (defaults to the minimum needed for the longest SEM). - seqId — metadata column for unique variant identification. - threshold — adjusted p-value cutoff (~0.05) for enrichment calls. - method — clustering metric (e.g. "WPCC") for SEM dendrogram visualizations.

Canonical example. The vignette scores chr12:94136009 by building gr <- GRanges(seqnames="chr12", ranges=94136009) and calling scoreBinding(x=gr, sem=SEMC, genome=Hsapiens), yielding 446 predictions (223 TFs × 2 orientations). A downstream enrichment example mixes 200 JUN-derived sequences with 800 random ones and recovers JUN as significantly enriched (p<1e-167) — a positive-control validation of the matrix-scoring approach.

Where it fits in the corpus

  • AACR 2026: axis bioinfo-tools — regulatory variant interpretation is an upstream step for many cancer-genomics dossiers
  • EuroBioC2025: ties into the TF-binding-prediction theme — Koen Van den Berge's transfactor talk on Day 1 PM is the deep-learning counterpart; SEMPLR is the classical-ML / matrix-scoring counterpart
  • ISMB 2026: scaffold; gene regulation / TF-binding overlap
  • RECOMB 2026: scaffold

Notes

Boyle and Phanstiel labs are well-known for chromatin and 3D-genome work — SEMPLR is the variant-effect prediction surface of that program, packaged as classical-ML matrix scoring rather than a deep-learning model. First Bioconductor release (v1.0.0) — the GBCC talk is effectively the launch presentation.