Causal circuit tracing reveals distinct computational architectures in single-cell foundation models: inhibitory dominance, biological coherence, and cross-model convergence

Clicks: 1
ID: 317597
2026
Free PDF
Article Quality & Performance Metrics
Overall Quality
Not rated
Combines reader engagement with the AI quality analysis. This article has not been analysed, so there is no overall score — reader engagement is measured and shown alongside.
AI Quality Assessment
Not analyzed
Readership in this journal

Ranked #792 of 829 articles by views in BMC Bioinformatics

Most read Least read

Bar heights use a square-root scale. Only the 120 most-read articles are drawn; the journal has 829 in total.

Mint this article as an NFT
Not yet minted

Create a permanent, verifiable on-chain record of this article on the Scimatic Network. The NFT is held in your Journament account, and you can withdraw it to your own wallet at any time.

5 SUSD one-off · no wallet required
Sign in to mint
Abstract
MOTIVATION: Sparse autoencoders (SAEs) decompose foundation-model activations into interpretable features, but the model-internal causal interactions between those features (i.e., what ablating one feature does to the others, as distinct from the biological causal structure of the underlying cells)-and how those model-internal relationships relate to biological structure-are uncharacterised in single-cell foundation models (scFMs). RESULTS: We introduce model-internal causal circuit tracing-zeroing one SAE feature at a source layer and measuring the resulting change in all downstream SAE features, for each of 120 source features-and apply it to Geneformer V2-316M and scGPT whole-human across four conditions (96,892 ablation-derived edges, 80,191 forward passes). On annotation-selected source features, edges share GO/KEGG/Reactome/STRING/TRRUST ontology terms at 50.9-68.5%, a 2.9-6.2× enrichment over a configuration-preserving permutation null (p<0.002); on 20 randomly sampled source features this attenuates to 21.5-26.3%-still 2.5-3.1× above null-quantifying the annotation-selection contribution. Inhibitory dominance (fraction of ablation edges with d<0, i.e. source activation supports downstream target) is 65.5-89.4%. scGPT produces larger raw per-edge effects (mean |d|=1.40 vs. 1.05); after feature-share normalisation, Geneformer is stronger (paired gene-pair ratio 0.64 on 33,301 shared pairs). Cross-model consensus yields 1,142 architecture-invariant domain pairs (ordered pairs of GO biological-process categories "A→B" each connected by at least one ablation edge in both models; 10.6× enrichment over permutation null; p<0.001). Circuit edge magnitude explains <1% of the variance in marginal driver-gene co-expression on the same cells (R2=0.010, n=31,176): the graph encodes structure beyond bivariate correlation. Against a matched-cell-type ENCODE ChIP-seq prior, circuit-predicted TF→target pairs are enriched 2.06× (Fisher OR 5.84), markedly higher than 1.12× against TRRUST; direct ChIP-seq-supported target pairs show 10-30× larger CRISPRi sign-bias-corrected excess than indirect pairs. Gene-level CRISPRi validation on Replogle K562 and the non-cancer RPE1 arm (and a true primary-T-cell control from Shifrut 2018) after sign-bias correction shows excess over baseline of +0.03 and +0.35 percentage points on K562 and RPE1 respectively (baseline already 52-56% from sign marginals); effect-magnitude Spearman correlations ρ≈0. Bootstrap and per-cell-type stability (N∈{50,100,200{; B cell, CD4+ T, macrophage) give Pearson r≥0.97 on shared edges with 100% sign agreement; edge Jaccard grows monotonically with sample size. The circuit graph is therefore highly reproducible as an effect-size map, cell-type-specific in edge identity, consistent with co-expression encoding, and weakly-but-detectably enriched for ChIP-seq-supported direct regulatory edges.
Reference Key
openalex_W7164915083 Use this key to autocite in the manuscript while using SciMatic Manuscript Manager or Thesis Manager
Authors Ihor Kendiukhov
Journal BMC Bioinformatics
Year 2026
DOI
10.1093/bioinformatics/btag379
URL
https://doi.org/10.1093/bioinformatics/btag379
Keywords Keywords not found

Citations

No citations found. To add a citation, contact the admin at info@scimatic.org

Comments

No comments yet. Be the first to comment on this article.