acceptodds
Under review as a conference paper at ICLR 2027

Latent-to-Legal Gradient Planning over Dynamically Revealed CRISPR Edit Graphs

Abstract

Every intended genome edit changes the sequence that defines which subsequent edits are executable. On RegEdit, a frozen set of 50 regulatory tasks, the median Jaccard overlap between consecutive action sets is 0.10; in 80% of tasks the tra- jectory selects a later action that did not exist at the root—one that a root-defined program could not have named. We formulate this as planning with state-dependent executable action identities: a root-fixed program retains the original candidate list, while a dynamic program regenerates the executable set after every transition. In the native comparison the dynamic regime produces a higher model-predicted endpoint on all 50 tasks— yet under exposure-matched conditions the difference vanishes (bootstrap 95% CI crossing zero). The action space changes dramatically; the current search does not exploit them. The gap between reachable and reached defines the open problem: converting action-space turnover into utility. Within this regime, L2L-GP achieves the highest median gain at every tested native query scale.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
CRISPR, Gradient Planning

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_dwKvOpenReview ↗
Rating
2(+1)
Confidence
3
Informativeness
2
Presentation
2(+1)
Soundness
2(+1)
Read review
Paper summary

The paper formulates sequential regulatory genome editing as planning over state-dependent action sets, where each edit can change which subsequent edits satisfy computational targeting constraints. It introduces L2L-GP, which performs gradient ascent on a training-free proxy over engineered action features and projects the resulting vector onto legal candidate edits. Evaluation uses 50 RegEdit tasks with a frozen AlphaGenome predictor as the objective. Comparisons with a No-Ascent control show modest, task-dependent gains from the gradient proposal mechanism.

Critical strengths

The problem formulation captures a relevant constraint in sequential genome editing. Earlier edits can alter guide/PAM compatibility and subsequent targeting opportunities. Explicitly distinguishing fixed and dynamically regenerated candidate sets makes this dependency testable.

Critical weaknesses

First of all, the paper is poorly written, with many important details missing, including the core method, benchmark procedure, and biological rationale. The training-free proxy is not explicitly defined, and the complete 18-dimensional action encoding is not provided. It is therefore unclear how these features are constructed, where they come from, and what justifies their use. Furthermore, Section 5.2 states that “The exact per-task exposure-matching schedule will be specified with the formal receipts.” This is an unresolved placeholder for a procedure that directly supports the paper’s central conclusion, making it difficult to assess the validity and reproducibility of the comparison. Furthermore, the experiments do not sufficiently establish that the proposed method is useful and generalizable. The evaluation covers 50 expression-increase tasks, but their selection criteria and biological diversity are not adequately described. Dynamic regeneration shows no detectable benefit after exposure matching, while gradient ascent provides only a small median improvement over the No-Ascent control. The same AlphaGenome predictor is used for both optimization and evaluation, without independent biological validation or mechanistic analysis. Consequently, it remains unclear whether the reported improvements reflect meaningful biological effects or generalize across diverse sequence contexts.

Questions for authors
  1. What exactly is the training-free proxy, and how are the 18 action features constructed and justified?
  2. How were candidate exposure and query budgets matched, and why do the reported budgets differ between experiments?
  3. How were the 50 tasks selected, and how diverse are their biological contexts?
  4. Are the gains over No-Ascent statistically reliable, and is their biological relevance supported by independent evidence?
2 comments
AuthorsOpenReview ↗

Response to Reviewer dwKv: Method, Evaluation Protocol, and Reliability

Thank you for your detailed questions on the method, evaluation protocol, task construction, and reliability of the reported gains. The revision provides explicit proxy, feature, and search specifications and a new same-version ascent ablation.

Q1. Proxy and 18-dimensional action representation

L2L-GP uses an analytic, training-free proxy to guide continuous proposals, which are projected onto currently executable CRISPR edits. Section 4.1 gives the complete proxy and regularized waypoint objective (Eqs. 1–2); Appendix A.2 specifies the 18 action features, sources, normalization, ascent updates, projection, coverage, and tie-breaking.

The coefficients are fixed heuristics for targeting context, efficiency priors, and edit burden, not learned or biologically calibrated effect sizes. Only target missingness from the state vector explicitly enters the proxy; other state dependence arises mainly from the regenerated executable catalog and action features. Projection searches the full current catalog, enforcing targeting and coordinate-replay constraints after each committed edit.

Q2. Exposure matching and query budgets

For the Dynamic–FrozenRoot greedy comparison, the per-task, per-depth schedule is ki,t=min⁡(128,ni,tFR)k_{i,t}=\min(128,n^{\mathrm{FR}}_{i,t}), where ni,tFRn^{\mathrm{FR}}_{i,t} is native FrozenRoot's recorded evaluated count. It is determined without using paired terminal outcomes. Both arms use the same candidate-count schedule and selection rule, without padding short pools. This tests root-replay restrictions under FrozenRoot-derived exposure schedules; candidate identities and physical costs may differ.

The greedy comparison has a request ceiling of 384. The separate L2L-GP/No-Ascent comparison uses 1792 per arm, with projection/preview width 128, beam four, and one successor lookahead. Sections 5.2–5.3 and Appendix A.4 distinguish planner requests, predictor-ledger events, and physical calls, including the three 383/384 exported-count discrepancies. Historical comparisons use native caps and are not equal-compute rankings (Tables 1, 7).

Q3. Task selection and biological contexts

Development50 was selected using predefined technical eligibility, prior-exposure exclusion, positive oriented-effect filtering, deterministic locus/component representatives, and chromosome/orientation stratification. Neither AlphaGenome predictions nor effect-size magnitudes determine selection ranking. Thus, selection is not driven by favorable predicted outcomes (§5.1; Appendix A.3).

The frozen cohort covers 50 distinct genes/loci, eight chromosomes, and 26 tissue annotations, spanning multiple genomic and regulatory contexts. The supplement documents selected-task identities and annotations. All evaluated tasks concern expression increase; unseen-task generalization is not established, and complete upstream materialization/input-exposure records are unavailable.

Q4. Reliability of the No-Ascent comparison

Both arms were rerun with the same code, predictor, and search configuration, changing only ascent steps (32 versus zero). Each task uses a recomputed common unedited baseline. The endpoint is a reference-base-pair-weighted predicted RNA-seq signal over the full transcript, with deleted reference bases contributing zero and the original transcript length retained as denominator (§5.3; Appendix A.3).

Across 50 paired tasks, ascent yields 39 wins and 11 losses. The mean paired absolute-change difference is 0.001828 (95% CI [0.0000491, 0.004810]); the median is 0.0001512 ([0.0000425, 0.0003007]). Median paired relative-gain difference is 0.447 percentage points. Both confidence intervals exclude zero. We use 10,000 complete-pair bootstrap resamples (seed 20261117), with percentile endpoints; task-level data and analysis code are provided (§6.3; Appendix A.5).

The positive effect is not restricted to T07, although that task contributes 68.8% of the net paired sum. Excluding T07 as an influence diagnostic leaves 38/49 positive differences and a positive median interval, while the mean interval includes zero. All 50 tasks remain in the primary analysis. The earlier historical No-Ascent comparison is distinct; the same-version ablation is the primary component evidence. Stage-level candidate and branch traces were not retained, so the endpoints do not identify which intermediate changes carry the effect.

These findings support a predicted-utility contribution from ascent in this in silico planner, not independently verified biological efficacy or cross-predictor generalization (§7).

Correction of turnover interpretation: The original 40/50 count concerns tasks selecting root-absent exact IDs, not necessarily new operations. Of 68 selected root-absent records, 36 retain a root replay counterpart, while 32 lack a root replay key; those 32 occur in 24/50 tasks (§6.1; Appendix A.1).

Thank you for the detailed response and the revision. The added clarification and the clearer presentation make the paper much easier to follow and reproduce. In this sense, I have raised my score to 2. However, because the scientific contribution and rigor remains limited, I am not able to raise the score further.

Reviewer_AsmnOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

This paper proposes a novel formulation of multi-edit genomic sequence editing, framing it as a planning problem with state-dependent action identities. The authors first identify that after initial genomic sequence edits, the action set turnover for subsequent edits is large, and has a meaningful impact on utility. The authors find that dynamically regenerating action spaces does not provide a utility advantage over a frozen root, and propose a novel method called L2L-GP which actively searches and exploits the newly available edits within the dynamically regenerated action space through gradient ascent and projecting back to executable candidate actions. On the RegEdit benchmark, the authors show that L2L-GP achieves higher median scores compared to older baseline methods, although a version of their algorithm with no gradient ascent performed better on 20 out of 50 tasks.

Critical strengths
  1. Problem framing: The authors identify and address an important gap in multi-edit CRISPR design, introducing the idea of state-dependent action spaces into multi-edit CRISPR design. Other approaches typically frame the problem as a combinatorial optimization problem, and the authors identify that the action sets and subsequent edits are extremely dissimilar once initial edits are made to a genomic sequence. The proposed problem forming is therefore novel and unique, and addressing it would make a meaningful contribution to the field.
  2. Experimental design: The authors use strict attribution and ablation controls to isolate the utility gain for exposure-matched Dynamic regeneration versus FrozenRoot comparisons.
Critical weaknesses
  1. Evaluation: The evaluation of proposed edits relies on the AlphaGenome model-predicted endpoint, without external biological experiments or verifications.
  2. Writing style: The paper is difficult to follow, and needs more background writing and explanation of the domain and data modality in the introduction and abstract. Terms such as “root-fixed” and “model-predicted endpoint” are used without sufficient background explanation, which will suffice for people within this research area but is not well-written for a more general audience to understand easily.
Questions for authors
  1. Could the lack of utility difference under exposure matching be an artifact of the AlphaGenome evaluation model’s sensitivity rather than an intrinsic property of the biological landscape?
  2. Could the authors reorganize Figure 1 to allow each Figure subpanel to look larger? In its current form it is difficult to read, however the figure quality and illustration looks good.
2 comments
AuthorsOpenReview ↗

Response to Reviewer Asmn: Evaluation and Presentation Clarifications

Thank you for recognizing the significance of the sequential-editing formulation and the controlled experimental design. We have further clarified the matched-comparison evidence and improved the presentation.

Q1. Could AlphaGenome sensitivity explain the matched-comparison result?

Our trajectory audit clarifies the origin of the matching outcomes. In the shared-schedule Dynamic–FrozenRoot comparison, 35/50 pairs select identical action sequences and materialized sequence trajectories. All 32 exact endpoint ties occur within these shared trajectories; the remaining three shared-trajectory pairs have numerical differences below 10−910^{-9}.

The exact ties therefore arise from identical sequence inputs, rather than demonstrating that AlphaGenome is insensitive to distinct edited sequences.

Among the 15 pairs with different terminal sequences, seven favor Dynamic and eight FrozenRoot. Across all 50 tasks, the mean paired difference is −0.000178 (95% CI [−0.000530, 0.000098]). Evaluator sensitivity remains unresolved for the 15 pairs with different terminal sequences.

These findings support a specific conclusion: under the tested greedy controller, unrestricted catalog regeneration does not produce a detectable additional mean predicted-utility benefit. More importantly, the trajectory analysis establishes that changes in editing opportunities need not translate into changes in selected programs.

Our study is an in silico investigation of multi-step CRISPR planning, providing a computational framework for evaluating and prioritizing candidate editing programs. AlphaGenome guides both selection and evaluation; accordingly, our conclusions concern predicted regulatory outcomes and computational executability. We do not extrapolate these findings to experimentally established editing efficacy or the broader biological response landscape (§7).

Q2. Figure 1 and presentation.

We expanded the Introduction with a concrete deletion example illustrating how earlier edits alter subsequent targeting conditions. Figure 1a has been redrawn as a compact vector diagram, unnecessary margins have been reduced, and the two panels are aligned at equal displayed height.

The revised figure and §3 now consistently distinguish current-state catalog regeneration from root-operation replayability, including the possibility that a refreshed executable record retains an original operation despite having a new exact ID.

I appreciate the authors response, which answers most of my concerns regarding evaluation, and I acknowledge that the manuscript has been revised to improve wording. I choose retain my score.

Reviewer_2zJFOpenReview ↗
Rating
2
Confidence
1
Informativeness
1
Presentation
1
Soundness
2
Read review
Paper summary

The paper introduces L2L-GP, which treats multi-step CRISPR editing as planning over a dynamic graph, where each edit changes what can be done next. It updates candidate edits at each step, uses gradient ascent in a continuous feature space to guide the search, and maps the results back to discrete edits that satisfy computational constraints, with AlphaGenome predicting their effects. Across 50 RegEdit tasks, the median Jaccard similarity between pre- and post-edit action sets was just 0.10, and80% of tasks used actions unavailable at the start. Dynamic updates outperformed fixed initial candidates in the original setup, but showed no significant advantage once candidate evaluations and query budgets were matched. The takeaway is that new editing opportunities do not automatically produce better outcomes: using them effectively remains a challenge, and the findings rely entirely on model predictions, without biological validation.

Critical strengths

I appreciated the many visualizations throughout the paper, particularly those in Figures 2, 3, and 4.

Critical weaknesses

I struggled to understand the main contribution of L2L-GP. The paper presents a method and a set of experiments, but it was not clear to me what specific problem the method addresses or what it adds over existing approaches. I also found it difficult to distinguish the new contributions from the baselines and to understand how to judge the method’s success: which metrics matter most, and where do the results show an advantage?

Questions for authors

Could the authors clarify what is new in L2L-GP, what benefits it is intended to provide over the baselines, and which metrics and experimental results best support those benefits?

1 comment
AuthorsOpenReview ↗

Response to Reviewer 2zJF: Contributions and Supporting Evidence

Thank you for your questions about the specific contribution of L2L-GP and the evidence supporting its effectiveness. We have clarified the central research problem, methodological novelty, and the distinct roles of our experimental comparisons.

Q1. What is new in L2L-GP, and what problem does it address?

The central contribution of L2L-GP is a constraint-aware planning framework for multi-step genome editing, where each edit may change the targeting conditions and executable choices available to subsequent steps.

Our novelty lies in two connected aspects.

First, we formulate sequential editing over dynamically regenerated executable records and distinguish record identity from operation replayability. This distinction enables us to investigate how earlier edits reshape future editing opportunities without conflating refreshed records with newly available operations.

Second, L2L-GP integrates gradient-guided optimization in an engineered action-feature space, projection onto currently executable CRISPR edits, and budgeted multi-step program search. Its analytic proxy supplies proposal gradients without differentiating through AlphaGenome, while current-state targeting and coordinate-replay constraints determine which discrete edits are executable (§§3–4).

Q2. Which experiments demonstrate these contributions?

Our primary metric is the task-paired difference in absolute predicted target-transcript gene-body change relative to the unedited baseline (§5.3).

The experiments establish three complementary findings.

Dynamic editing opportunities. In historical K=192 L2L-GP trajectories, median root-to-next-state Jaccard overlap is 0.10 using exact record IDs, compared with 0.55 using replay keys. This distinguishes record refresh from changes in available operations (§6.1).

Consequences for planning. Under shared candidate-count schedules, Dynamic and FrozenRoot Greedy select identical trajectories on 35/50 tasks, with no detectable additional mean predicted utility from unrestricted catalog regeneration. This shows that changes in available operations do not necessarily produce different selected programs or better outcomes (§6.2).

Gradient proposal effectiveness. A new same-version 32/0-step ablation evaluates the contribution of ascent. L2L-GP achieves higher predicted utility on 39/50 tasks, with positive mean and median paired differences and 95% bootstrap intervals excluding zero (§6.3).

These comparisons address distinct questions: which operations are available, whether selected decisions change, and whether gradient-guided proposals improve search outcomes. Table 1 and the revised Introduction explicitly connect each contribution to its supporting evidence.

Beyond the observed predicted-utility improvements, our formulation provides a computational foundation for exploring constraint-aware multi-step CRISPR editing and prioritizing candidate programs for future experimental investigation.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.