Skip to content

consICA

Consensus Independent Component Analysis — data-driven deconvolution method using consensus ICA to decompose heterogeneous omics data and extract features for patient diagnostics and prognostics.

Talk at EuroBioC2025

  • Speaker: Maryna Chepeleva (presented; Nazarov is the maintainer / senior author)
  • Day / session: Day 2 (Thu Sep 18, 2025) — late-morning short-talk session, before the Ferruz keynote
  • Talk title: "consICA for functional signal decomposition"
  • Slides (Zenodo DOI): to verify after publish
  • Video: not recorded
  • Abstract / schedule: eurobioc2025 schedule

What it does

consICA runs Independent Component Analysis many times with different initializations, then aggregates the results into stable consensus components — the "consensus" step is what turns ICA from a stochastic exploratory method into a reproducible deconvolution tool. The output is a set of statistically independent signal components extracted from heterogeneous omics data, with downstream utilities for linking components to patient survival, clinical covariates, and gene-set enrichment.

How it works

Core idea. Run fastICA ntry times with different random initializations on the same expression matrix, then aggregate the per-run components into stable "consensus" components — a standard remedy for the well-known non-determinism of single-run ICA. The aggregation step (matching components across runs and averaging) is what turns ICA from an exploratory tool into a reproducible deconvolution method.

Inputs / outputs. Input: a SummarizedExperiment carrying an expression matrix (genes × samples), sample metadata, and optionally survival time/event columns. Output: an S matrix of consensus metagenes (genes × components), an M matrix of component weights per sample (components × samples), and per-component stability scores; downstream utilities link components to survival, clinical covariates, and gene-set enrichment.

Key innovation. The consensus criterion — components from independent runs that match consistently across initializations are kept; unstable components are flagged or discarded — is the methodological signature. Components are then explicitly framed as deconvolution signals to be tested against clinical metadata and pathway databases rather than treated as black-box features.

Parameters worth knowing. ncomp — the number of independent components to extract (analyst-chosen, the primary modeling decision). ntry — how many ICA restarts to aggregate (more restarts → more stable consensus, more compute). Sample metadata columns to test against (Var) and survival columns drive the downstream component-interpretation tooling. Specific aggregation algorithm not detailed in the vignette excerpt accessed.

Canonical example. The vignette uses TCGA skin cutaneous melanoma (SKCM) — 472 samples × the 2,454 most variable genes, plus 103 metadata columns and survival data — pre-loaded as a SummarizedExperiment via data("samples_data"). Consensus components are then linked to melanoma subtype and survival as the worked example.

Where it fits in the corpus

  • AACR 2026: no current dossier; relevant axis is bioinfo / AI methods — consICA complements PCA/NMF/omicsGMF dimensionality reduction with an ICA-based alternative
  • omicsGMF (entry) — adjacent matrix-factorization framing, different statistical assumptions
  • ISMB 2026: scaffold
  • GBCC 2025: queued
  • Nextflow Summit 2026: not mentioned

Notes

ICA-based deconvolution — explicitly oriented toward extracting clinically interpretable signal components, with the consensus step addressing the standard ICA reproducibility concern.