DR-GEM turns worst reconstruction errors into a search signal for rare cell types

bravo_abad · x · 2026-09-25

A paper thread makes a counterintuitive point: the worst model fit may be the discovery signal.

• Most ML pipelines minimize error, but in scientific data the largest errors can reveal structure the model failed to understand
• Across 8 real single-cell datasets, rare cell types, states, or perturbations were 1.15–9.18x more likely to appear among cells with the highest reconstruction error — rare populations contribute too little to average loss to shape the learned representation
• DR-GEM uses error as a search signal: it identifies poorly reconstructed samples, gives them more influence when rebuilding the representation, then verifies the resulting structure reproduces across balanced subsamples
• Patterns the original embedding ignored become visible, offering a new paradigm for discovering rare biological phenomena

Original post →

More from Research

• Puro-2B: an open recipe trains a Qwen2-1.5B-beating LLM on RTX 5090s for just $4.4K — IgorCarron · 2026-09-25
• Oxford-led NeurIPS paper: frontier LRMs mirror human rule discovery and brain activity — sreejan_kumar · 2026-09-25
• bold_lab_ai hiring: postdocs and RAs due Oct 9, strategic partnership PM due Oct 14 — josephdviviano · 2026-09-25
• AIRA₂ Research Agents Hit 81.5% on MLE-bench-30, Beating Prior SoTA of 72.7% — mariofilhoml · 2026-09-25
• Overfitting the Validation Set Is Rarely a Big Issue, AIRA₂ Ablations Show — mariofilhoml · 2026-09-25
• Nature Machine Intelligence: Cybernetics and interoception framework for embodied AI agents — AnnaCiaunica · 2026-09-25

Research channel →