LPDP: Variable-Length DNA Generation with Edit Flows
A novel technique known as Local Perturbation Discrete Programming (LPDP) facilitates reward management during inference for generating DNA sequences of varying lengths through Edit Flows. In contrast to fixed-length models, Edit Flows support biologically relevant insertions, deletions, and substitutions. LPDP functions without training, evaluating one-step root edits, maintaining a nearly optimal root band, and reordering each root by addressing a constrained local discrete program related to its child sequence. This strategy leverages the structured geometry of edit actions to concentrate on cohesive subgraphs and compiles local continuations.
Key facts
- LPDP is a training-free, intermediate-state and action-aware local re-solving operator
- It operates on variable-length DNA edit-action generators at inference time
- Edit Flows enable biologically plausible insertion, deletion, and substitution operations
- LPDP scores one-step root edits and retains a near-best root band
- It re-ranks each retained root by solving a bounded local discrete program
- The local program uses typed geometry of edit actions
- It focuses on coherent substitution, insertion, or deletion subgraphs
- The method aggregates local continuations
Entities
—