From DNA Design to DNA Slimming: Auditable Agentic Discovery of a Deletion-Only Designer

1 Introduction

• Compact payloads can improve vector properties such as transducibility, but natural flanking sequences may also carry desirable properties.The paper therefore frames slimming as a size/function tradeoff.Source: §1 · p. 1
• Model-based designers typically optimize oracle-predicted activity through substitutions, and prior quantitative comparisons did not include deletion-based designers.These approaches have been biologically validated.Source: §1 · p. 1
• GrAdaBeam combines gradient guidance with adaptive discrete search, but substitution gradients do not directly transfer to deletions because downstream coordinates shift and new junctions form.Ledidi and FastSeqProp gradients likewise cannot be interpreted directly as deletion gradients.Source: §1 · pp. 1–2
• The paper defines a machine-verifiable task: return an exact-length, order-preserving subsequence of an existing functional regulatory sequence.Prior computational enhancer design used fixed-length substitutions or mixed substitution, insertion, and deletion paths.Source: §1 · p. 2
• ERA searches over executable designer programs, while each proposed program searches over DNA subsequences.This makes the outer-loop output inspectable, versionable, and directly testable code.Source: §1 · p. 2
• The paper contributes a slimming evaluation framework, an ERA-discovered two-phase designer, and five held-out BPNet studies against random and greedy controls.The studies use a target-separated protocol.Source: §1 · p. 2

2 Defining sequence slimming

• A valid slimmer returns an exact-length candidate and increasing source indices, with each output base matching its indexed source base.Substitution, insertion, position reuse, reordering, and inversion are disallowed.Source: §2 · p. 2
• The objective minimizes oracle energy after embedding the candidate in a fixed-length background; the evaluation uses one available design background.The BPNet oracle expects length Lmin, which exceeds target length m in these evaluations.Source: §2 · p. 2
• The evaluator reconstructs the indexed subsequence and rejects outputs that fail sequence reconstruction or index checks.These checks detect substitutions, insertions, duplicated or invalid indices, and reordering or inversion.Source: §2 · p. 2
• Exactly m bases and fixed target lengths avoid confounding methods by different output sizes or incentivizing substitutions over deletions.The framework evaluates multiple fixed values of m to obtain a length–activity curve.Source: §2 · p. 2
• R = 1 retains the source’s predicted lift over background, while R > 1 improves it.Raw candidate energy was rejected as the primary measure because its scale and baseline vary across starts and contexts.Source: §2 · p. 3
• Scores average across five test backgrounds per sequence, then paired method differences are bootstrapped over starting sequences.The protocol uses a wall-clock limit to account for differing computational costs.Source: §2 · p. 3

3 Agentic discovery of a deletion-only designer

• ERA received the slimming prompt and GrAdaBeam, with hard checks rejecting invalid, rearranged, substituted, or overlength outputs.Accepted candidates were centered in a target-specific low-activity background for scoring, and instructions prioritized satisfying the length constraint before optimizing activity.Source: §3 · p. 3
• ERA evaluated proposed designers on one high-scoring ATAC start and one MYC start, averaging target-specific BPNet-lite energies as its objective.Failures, timeouts, and invalid subsequences received the worst objective value.Source: §3 · p. 3
• GRADASLIM removes roughly half the remaining length gap per step, then refines exact-length candidates with swaps and shifts guided by learned salience.It updates salience from better-than-average candidates, diffuses some credit to neighbors, and uses simulated annealing and an archive to preserve strong candidates.Source: §3 · p. 3

4 Five-target BPNet evaluation

• ERA has the highest mean retained effect in 9/10 settings; at 400 bp all five ERA−greedy intervals exceed zero, while 100-bp E2F3 favors greedy (−0.299, 95% CI [−0.516, −0.072]).Table 1 reports retained effect and paired ERA−greedy intervals by target and length.Source: Table 1 · p. 4, §4 · p. 4
• At 100 bp, ERA’s mean is numerically higher for ELF4, MAX, MECOM, and RAD21, but all four ERA−greedy intervals overlap zero.The 100-bp setting is substantially shorter than the 300-bp slimming focus during ERA program search.Source: §4 · p. 4
• ERA’s mean retained effect exceeds 1 in every setting, indicating improved predicted lift over background relative to full-length sources.Each study compares random exact deletion, batched greedy deletion with swap refinement, and GRADASLIM on five starts at 400 and 100 bp.Source: §4 · pp. 3–4

• The evaluation uses five starts and one optimizer seed per target, tests only two output lengths, and selects targets and high-activity starts using NucleoBench results.Source: §5 · p. 4
• No wet-lab validation or orthogonal-model test establishes biological preservation; frequent R > 1 values may reflect oracle exploitation or artificial deletion junctions.The 100-bp evaluation also tests more aggressive slimming than ERA's program search used during discovery.Source: §5 · p. 4
• Bootstrap intervals have limited resolution because they use only five starting sequences per target.Source: §5 · p. 4
• ERA shows a consistent advantage at 400 bp and mixed performance at 100 bp.Source: §5 · p. 4

• The evaluation implementation returns the archived best candidate and its source indices, with the algorithm's proposal and credit-assignment rules discovered by ERA.Source: A Complete discovered algorithm · pp. 5–6
• Phase 1 rapidly reaches the target length by proposing batches that delete a selected number of retained positions.The deletion count is at least one and is based on the gap between the current and target lengths.Source: A Complete discovered algorithm · p. 5
• At exact length, Phase 2 proposes salience-biased swaps, single-position shifts, and order-preserving contiguous-block shifts.The block length ranges from 1 to 11.Source: A Complete discovered algorithm · pp. 5–6
• Each scored batch updates retained bases using candidate energy relative to the batch mean, diffuses credit to neighboring coordinates, and accepts a best proposal or uses simulated annealing.An archive preserves the best exact-length solution found.Source: A Complete discovered algorithm · p. 6
• The salience vector is a persistent credit-assignment reservoir, not a model gradient; it guides deletions and restoration of omitted bases.Neighbor diffusion may preserve motifs but may also amplify correlated artifacts.Source: A Complete discovered algorithm · p. 6