Dinuc shuffling preserves the CA initiator dinuc, a pitfall for seq2func genomics models
anshulkundaje · x · 2026-11-07
Anshul Kundaje retweets his team's thread on seq2func models: dinucleotide shuffling is the standard way to build negatives, preserving base and dinuc content while destroying TF motifs — but the dominant feature for TSS positioning is the initiator, itself a dinucleotide (CA), which shuffling fails to destroy.
- Preserves important covariates (base/dinuc composition)
- Destroys motif-level features
- Leakage risk via the CA initiator
More from Research
- Regularization can amplify subliminal learning, surprising AI safety finding — AdtRaghunathan · 2026-11-07
- OpenVax revamps personalized cancer vaccine toolkit; OpenVax2 adds long-read support — iskander · 2026-11-07
- Google's Muir linear attention fits 10M tokens in a 25MB state, beats GDN-2 — CategoryNormal149 · 2026-11-07
- Ollama's prompt cache changes LLM outputs even at temperature 0 with fixed seed — Full-Travel857 · 2026-11-07
- OpenAI launches BioBasicsBench: 100 real wet-lab tasks to stress-test AI agents — OpenAI · 2026-11-07
- Scaling systolic array diagrams from 8x8 to 64x64: 4096 PE matrix multiply visualized — satnam6502 · 2026-11-07