LOCOGEN
A locality-guided method for regulatory DNA sequence design. Through invariant-motif mask learning and OOD damping, it applies localized guidance to a frozen diffusion model during sampling, balancing activity, novelty and plausibility.
Timeline (last 90 days)
The core of LOCOGEN is to apply localized guidance to a frozen diffusion backbone network during sampling, without retraining the backbone.
LOCOGEN contains two key components: invariant motif mask learning (identifying motif positions that are critical to function and protecting them from excessive guidance) and OOD damping (penalizing generations that deviate too far from the natural distribution at background positions, but acting only on background regions).
LOCOGEN was experimented on three cell lines, HepG2, K562 and SK-N-SH, and the results show that LOCOGEN can reach previously unexplored regions, that is, improving novelty and plausibility while maintaining high predicted activity.
LOCOGEN achieved a balance among performance, novelty and plausibility on all three cell lines, whereas the baseline methods could not satisfy all three simultaneously.
The LOCOGEN code has been open-sourced on GitHub (https://github.com/Lotfollahi-lab/LocoGen).
LOCOGEN's locality idea may offer inspiration for other sequence design tasks (such as protein engineering).
All Facts (6)
The core of LOCOGEN is to apply localized guidance to a frozen diffusion backbone network during sampling, without retraining the backbone.
50%UnverifiedLOCOGEN contains two key components: invariant motif mask learning (identifying motif positions that are critical to function and protecting them from excessive guidance) and OOD damping (penalizing generations that deviate too far from the natural distribution at background positions, but acting only on background regions).
50%UnverifiedLOCOGEN was experimented on three cell lines, HepG2, K562 and SK-N-SH, and the results show that LOCOGEN can reach previously unexplored regions, that is, improving novelty and plausibility while maintaining high predicted activity.
50%UnverifiedLOCOGEN achieved a balance among performance, novelty and plausibility on all three cell lines, whereas the baseline methods could not satisfy all three simultaneously.
50%UnverifiedThe LOCOGEN code has been open-sourced on GitHub (https://github.com/Lotfollahi-lab/LocoGen).
50%UnverifiedLOCOGEN's locality idea may offer inspiration for other sequence design tasks (such as protein engineering).
50%