Stanford's Kundaje fires nine-tweet rebuttal over TF cooperativity study
On November 19, Stanford computational genomics researcher Anshul Kundaje posted nine tweets systematically responding to a DNA transcription factor analysis by the BarkaiLab team, which was built on his own team's open-source models and data resources. He welcomed the rapid open reuse of those resources but explicitly disagreed with the study's conclusions, arguing the analysis was overly conservative — neither demonstrating a lack of transcription factor cooperativity nor actually testing binding selectivity.
Confirmed
- Kundaje noted that when using directional experiments like ChIP-seq to test whether a transcription factor's binding depends on other TFs, the study's methods were overly conservative
- He stressed that the analysis did not explicitly test selectivity (i.e., why certain motifs are bound while others are not); short-context models like BPNet trained on peaks and GC-matched backgrounds are poorly suited to answering such questions, whereas the discriminative performance of genome-wide long-context models is more appropriate
- He reviewed existing evidence that TF binding relies on multiple mechanisms: direct interactions (composites and hard syntax) plus mostly chromatin-mediated soft syntax; in addition, non-canonical sequence signals such as IDRs and non-DNA-binding domains (e.g., low-complexity sequences that may serve as IDR targets) also participate in targeting, and his models have already learned these predictable features
- He previewed a brief follow-up, in about a week, explaining why the analysis neither demonstrates a lack of cooperativity nor tests selectivity, and said his team would need several months and multiple methods to complete a comprehensive analysis of the resource
Why it matters
- The dispute involves Kundaje's own open-source models and data being used by a third party to draw broad conclusions he rejects — a public clash over the tension between open-science resource reuse and interpretation of results
- Kundaje's team encourages everyone to keep using these models and data and has committed to a fuller multi-method analysis later; the episode also offers lessons for the methodology of evaluating genomic foundation models (e.g., choosing model context length)
2026-11-19 ~ 2026-11-19 · 10 related posts
Primary sources
- [source] Stanford's Anshul Kundaje fires 9-tweet rebuttal at study claiming TF binding lacks selectivity — anshulkundaje · 2026-11-19
- Kundaje thread (2/9): preview promised on cooperativity and selectivity critique — anshulkundaje · 2026-11-19
- Kundaje thread (3/9): directional ChIP-seq approach overly conservative for TF dependence — anshulkundaje · 2026-11-19
- Kundaje thread (4/9): analysis never explicitly tests motif selectivity — anshulkundaje · 2026-11-19
- Kundaje thread (5/9): short-context models ill-suited for motif selectivity questions — anshulkundaje · 2026-11-19
- Kundaje thread (6/9): extensive evidence of TF binding dependence mechanisms — anshulkundaje · 2026-11-19
- Kundaje thread (7/9): IDRs and non-canonical sequence features as localization signals — anshulkundaje · 2026-11-19
- [source] Kundaje thread (9/9): restricted analyses can't support the study's broad conclusions — anshulkundaje · 2026-11-19
- Stanford genomics researcher defends model resource, previews motif-calling improvements — anshulkundaje · 2026-11-19
1 near-duplicate retellings: anshulkundaje