Skip to content

Anshul Kundaje — sequence-to-accessibility models (ChromBPNet) on fibroblast reprogramming

Speaker: Anshul Kundaje — Associate Professor of Computer Science & Genetics, Stanford University. Graduate students Surag Nair and Mohamed Ameen ("Moe") — co-first authors of the reprogramming paper. Collaboration with Ionis; "Kevin"/Helen. Talk title (SCGD26): "Deep learning the regulatory code of fetal development and cellular reprogramming from single cell chromatin accessibility" Slot: SCGD26, Jun 12, 2026 Method: ChromBPNet — base-resolution sequence→ATAC-seq deep-learning models + interpretation (DeepLIFT/DeepSHAP, TF-MoDISco) Recording: SCGD26 livestream — paper, data, and interactive browser at the reprogramming browser Status: Captured from session transcript

Thesis

How is one fixed genome read differently across cell states? Using Yamanaka (OSKM) reprogramming of human dermal fibroblasts → iPSCs (Sendai delivery, 14-point time course, scRNA + scATAC), sequence-to-accessibility models give mechanistic insight — not just predictions — about how DNA sequence, TF concentration, and motif affinity drive chromatin dynamics.

Setup & models

  • ~15 cell states from fibroblast → "high-OSK" → pre-iPSC → iPSC, with off-target trajectories (keratinocyte-like, partially reprogrammed). Endogenous vs. exogenous OSKM separable by isoform structure → the high-OSK state has supraphysiological OSKM; iPSC has physiological levels.
  • A separate ChromBPNet (small bias-corrected seq→profile model) trained per pseudobulk cell state (Pearson ~0.75–0.875). Interpreted with DeepLIFT/DeepSHAP contribution scores + TF-MoDISco to track per-nucleotide, per-state motif activity in individual enhancers (e.g. the NanoG locus, with CTCF insulators bookmarking dynamic enhancers).

Key findings

  • ~50–60k transient peaks open then close, are out-of-distribution (absent from ENCODE/Roadmap and beginning/end states), and map to genes with no obvious enrichment — invisible without dense time-sampling.
  • Affinity × concentration trade-off. At supraphysiological OSKM, OCT4/SOX2 bind low-affinity, non-canonical motifs (e.g. POU "CAT→AAT") that physiological levels can't. Transient peaks with the weakest OCT-SOX motifs close first as TF concentration drops — explaining the temporal decay. ChromBPNet, trained only on sequence + accessibility, reproduces footprint changes as a function of both concentration and affinity despite never being told either (synthetic-motif insertion experiments for KLF, SOX2).
  • "Repression by theft" model. Transient low-affinity OCT-SOX peaks also carry AP-1 sites. The flood of new peaks sponges AP-1 away from fibroblast enhancers → rapid (~5×, within 2 days) shutdown of fibroblast genes. Confirmed by single-cell multiome (day 1–2): per-cell accessibility at newly opened AP-1 peaks anti-correlates with fibroblast-program expression.

Q&A highlights (with Rahul Satija)

  • Low-affinity motifs are pervasive, not a reprogramming artifact — standard motif-calling/binarization (ChromVAR-style thresholds) hides them; collectively many weak sites rival a single strong one (cites Julia/related work; Mars/Barkai-style low-affinity work).
  • One model for promoters + enhancers is fine — they're ends of a continuous spectrum; fine-tune sampling/learning rates for subtleties (vs. a separate-models view).
  • Gene-expression (RNA) prediction is a data problem, not architecture — only ~20k genes, each enhancer seen once in native context; needs perturbation data (enhancer hopping, deletions, inversions) to learn enhancer→gene syntax.
  • What a "virtual cell" should be: not just accurate prediction but mechanistically anchored internals — interpret the model and show predictions arise through real causal biology.

Connections to the corpus

Sources