acceptodds
Under review as a conference paper at ICLR 2027

KRAIG: Knowledge-Guided RNA Adaptive Inference and Generation under Biological and Protocol Uncertainty

Abstract

RNA therapeutic design requires a sequence of coupled decisions under heterogeneous biological evidence, assay-dependent supervision, and protocol shift: selecting an admissible therapeutic mechanism, generating and prioritizing candidate oligonucleotides, and determining whether a prediction is sufficiently supported to act upon. Existing approaches typically optimize these decisions independently and propagate point predictions even when the evidence required by a downstream model is unavailable. We formulate RNA therapeutic design as compositional selective inference and introduce KRAIG, an end-to-end framework in which every stage communicates an evidence certificate rather than an unconditional decision. KRAIG first converts typed biological evidence into dependence-robust mechanism intervals and returns a certified mechanism, an admissible set, or abstention. This certificate governs a capability-aware design component that activates mechanism-conditioned generation, within-experiment ranking, and conformal candidate selection only when their modality-specific validity, data, and calibration requirements are satisfied. A support-selective protocol component permits residual adaptation only within supported protocol regions and otherwise exactly recovers an invariant predictor. Under controlled experiments with known latent mechanism scores, certification incurs zero selective error when it returns a singleton, whereas forced point-decision baselines err on 15.7–21.9% of near-tie cases. Across three supported therapeutic modalities, constrained decoding increases chemistry-valid generation from 0–88.8% to 100%. Modality-stratified conformal selection attains 0.92 coverage at 0.90 nominal coverage in the adequately powered stratum while withholding unsupported guarantees elsewhere. Learned components are evaluated under a frozen five-seed protocol with leakage-resistant splits and paired controls. By composing certification, generation constraints, calibrated selection, specialist routing, and invariant fallback, KRAIG makes the decision to abstain or defer an operational part of RNA therapeutic design.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
RNA therapeutics, selective inference, knowledge-guided machine learning, conditional generation, uncertainty quantification, conformal prediction, domain adaptation, abstention

Reviews

Among:
Shape:
1234this paper, Bottom 25%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.75 over 4 reviews, above 25% of the others.
Reviewer_AvHxOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper frames RNA therapeutic design as selective inference, where each pipeline stage returns an evidence certificate and is allowed to abstain. The framing, leakage controls, and exact-fallback proof are exemplary. The problem is the empirical core, which is weak for ICLR. For example, the ranker anti-correlates with real activity out of distribution, the ranking gain is small and partly just comes from ensembling, and several "certificates" rest on data reuse the authors themselves admit violates their independence assumptions.

Overall, it reads more like a reliability framework with honest negative results than a genuine methods advance.

Critical strengths
  • Every stage of the RNA design pipeline returns an evidence certificate and is allowed to abstain instead of forcing a point call. Turning "defer" into a first-class operation is the right instinct for a noisy, high-cost domain.

  • Frozen SHA-256 manifests, machine-checkable leakage audits, and a real exact-fallback proof. The constrained-decoding result is concrete too. Chemistry-validity jumps to 100% across all three modalities with novelty preserved.

Critical weaknesses
  • The learned ranker is the headline but it falls apart out of distribution. This paper claims that rank correlation with true activity goes negative on every external set, often worse than plain GC content. For a "ranker" contribution, that's the main event, not a footnote.

  • The +0.110 over OligoAI only shows up when you stack on the baseline you're comparing against, and a capacity-matched control basically matches it. Algorithmic novelty is light. For example, the stats are all classical and the authors themselves call KRAIG-SR "a simple diagnostic baseline."

  • The independence premises rely on weights/scores/grid that reuse the dev families, and the agreement stat was inspected after seeing outcomes.

  • Power is also low (five seeds, sign-test p=0.0625, Holm 0.44).

Questions for authors

please see Weakness section.

Reviewer_j5ZwOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

I'm rating this paper as 2, but I'm open to increasing my score, if my concerns are addressed.

The paper builds on a system for RNA therapeutic design KRAIG. The system has 5 stages

  • TRIAGE: the therapeutic mechanism is decided
  • GENERATE: a conditional VAE proposes sequences
  • RANK: several predictors (i.e. experts) rank the candidates
  • SELECT: select a shortlist via conformal prediction
  • ADAPT: a protocol-adjustment term is added to a base predictor

One of the main contributions is the contract idea: each stage passes along a certificate, and abstention is built in. The Fréchet–Hoeffding bounds are used to turn different biological evidence into support intervals.

The experimental results are mostly convincing, and show the effectiveness of the proposed approach.

Critical strengths
  • The multi-stage certification with guarantees is highly valuable in RNA therapeutic design
  • The experimental results are mostly convincing
Critical weaknesses
  • The paper is not easy to read, and some information of vaguely described. For example, I failed to understand what is meant by 'biological evidence' concretely (e.g., what data modalities does it include? Does it include meta-data as well?...)
  • According to the paper

KRAIG is being integrated into the broader KRAIG RNA Therapeutics Platform, an actively developed system spanning knockdown, upregulation, editing, and splicing strategies.

Therefore, KRAIG has been previously developed, and the paper does not explicitly mention what the current research adds to the existing pipeline. For example, is the existing KRAIG pipeline totally missing a certificate mechanism or the described KRAIG-SR training?

  • When presenting the theoretical findings, the distinction of the finding to the previous knowledge is not clearly mentioned. For example,Proposition 1 seems to be the classical Frechet–Hoeffding bound, and Proposition 2 is a trivial statement.
Questions for authors

Please see my comments in "Critical Weaknesses".

Reviewer_EtEUOpenReview ↗
Rating
1
Confidence
4
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

KRAIG organizes RNA therapeutic design into stages that pass explicit evidence and support records: mechanism triage, constrained generation, activity ranking, calibrated candidate selection, and protocol adaptation with fallback. The paper combines conditional guarantees with synthetic checks, internal multi-seed comparisons, and retrospective ASO evaluations. Its strongest ranking result augments public OligoAI scores with additional within-assay rank views. The broader contribution is an interface that makes unsupported decisions and abstention explicit throughout the workflow.

Critical strengths
  1. The problem is well motivated. Separating missing biological evidence, sequence validity, predictive ranking, and protocol support is useful for an application where a plausible score can otherwise be mistaken for sufficient evidence to proceed. The explicit reporting of external failures and unsupported guarantees makes the proposed interface more informative about its limits.
  2. The ranking evaluation includes matched assay comparisons, patent-family resampling, ablations, and sensitivity analyses. Recalculation from the supplied assay-level metrics recovers the main EF20 improvement, 1.539 to 1.649, and the paired interval for the difference, 0.110 [0.030, 0.190]. This is a meaningful, checkable result within the evaluation represented by those files.
Critical weaknesses
  1. The supplied training path does not establish the complete-stack isolation claimed in Section 4.1 and Table 2. The manuscript states that the 39 strict families contribute no labels to any fitted component. However, kraig_v2_listwise_ranker.py generates base predictions by cross-fitting over the full family pool, training on folds != fold; kraig_v2_alignment_ablation.py does the same for bag-raw. The recorded pool contains 86 families, and none of the five held-out folds contains more than 27 families. Thus these base-model fits necessarily include labels from some of the 39 strict families. kraig_v2_stacked_expert.py imports those predictions before separating development and strict groups for combiner fitting. A row's own family can be excluded from its base predictor while strict-family labels still affect the development scores used to fit the combiner. This is different from holding out the entire strict block throughout training. It does not establish how much the reported gain is affected, but it leaves its stated generalization interpretation unresolved. The paper's retrospective-development caveat does not reconcile this specific discrepancy with Table 2.
  2. The method description and the supplied implementation disagree on the experts underlying the main result. Appendix C.8 describes bag-raw as a weighted average of measurements and LambdaRank as an MLP over frozen RiNALMo embeddings. The files above instead produce these views with a LightGBM regressor and a LightGBM ranker on engineered features, and the stacking script explicitly consumes their output columns. These are substantively different algorithms. The authors need to identify which implementation produced Table 4 and align its description and prediction provenance. Verification beyond the main ranking table is also limited by the provided bundle: 14 of the 19 source paths required by submission_claim_audit.py are absent, including generation and conformal results, and no model checkpoints were found. This is narrower than the reproduction scope stated on page 10.
Questions for authors
  1. Which exact base-model training-family lists and prediction files produced Table 4? How do they satisfy Table 2's exclusion of strict outcomes? If the supplied cross-fitting path was used, can the comparison be repeated with an outer family holdout for the entire stack, with combiner-training predictions generated only within the outer training portion?
  2. Are the bag-raw and LambdaRank experts the methods in Appendix C.8 or the LightGBM models in the supplied scripts? Please reconcile the description, code, and result provenance, and identify the missing source artifacts or their versioned release location.
  3. What event does the RNase-H conformal set cover? Please specify the nonconformity score, calibration unit and sample size, candidate-set sizes, and whether the guarantee applies before or after upstream selection.
  4. The paper explicitly treats the 38.6% acceptance result as diagnostic, and Appendix E.3 reports act-all or refuse-all at the tested risk targets. Is there an independently calibrated operating point with useful partial action and controlled loss on an existing cohort, compared with a baseline allowed to abstain? Such a risk/action comparison, following the selective-prediction perspective of Geifman and El-Yaniv (2019), would help assess the broader system contribution beyond the ranking result. If none is established, please clarify that limitation in the scope of the contribution.
Reviewer_5U4hOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper recasts RNA therapeutic design as compositional selective inference: five stages (triage, generation, ranking, selection, adaptation) each emit an evidence certificate that downstream stages may refuse. Triage uses Fréchet–Hoeffding conjunction bounds over mechanism marginals to certify a unique mechanism or abstain; generation is a constrained conditional VAE with a finite-state decoder mask; ranking combines a shared multi-modality ranker (KRAIG-SR) with an assay-local percentile stack; selection applies split conformal top-k; adaptation is a gated FiLM residual with an exact λ=0 fallback. Validation is a mix of synthetic triage simulations, generation validity audits, a retrospective ASO transport benchmark under patent-family-disjoint splitting, one external FANTOM6-iPSC transfer test, and several deliberately reported negative external stress tests.

Critical strengths
  1. Explicit claim contract with disclosed failures. Table 1 scopes every claim, and Appendix E.2 reports external panels where KRAIG-SR is anti-correlated (H3-3A ρ = −0.500/−0.579; ASO Atlas 2.0 −0.092) and loses to plain GC content. Most submissions would bury this. It materially raises my trust in the reported positives.
  2. A serious leakage-control protocol for cross-assay transport. Patent-family grouping plus exact/Hamming-1 exclusion, pre-registered success rule, family-clustered intervals, and ablations of every view. This is the right way to evaluate this task, and the protocol is reusable.
Critical weaknesses
  1. The compositional claim — the paper's actual thesis — is never tested. No experiment compares the certified pipeline against the same predictors without certificates. The router is reported at 0.406 ± 0.032, identical to specialists (0.406 ± 0.032), i.e. it never routes on the evaluation set. Conformal coverage (0.92 vs 0.90 nominal, 100 groups) is reported standalone, with no selective-risk contrast against an uncertified baseline. Generation validity of 1.000 is enforced by a hard decoder mask, hence true by construction, while the authors' own audit finds no target-conditioned active-candidate recovery. Adaptation shows no accuracy gain and only the trivial λ=0 identity is verified. Each stage is therefore individually vacuous or unvalidated, and nothing establishes that the composition does more than the sum of its parts — which is precisely the contribution claimed.
  2. The headline ranking result is not attributable to the proposed method, and does not transfer. The strict-family EF20 of 1.649 beats OligoAI's 1.539 only in an ensemble in which OligoAI receives the largest weight (1.46, vs 0.30–0.55 for KRAIG views); removing it collapses the gain to +0.013 [−0.110, 0.145]. The authors' own stored head-to-head result has decision_rule_passed: false. On every outcome-unseen external panel the learned ranker is null or anti-correlated (KIF1A −0.257…−0.086; FANTOM6 consensus ΔEF20 −0.019 [−0.244, 0.207]), so the positive result is a retrospective within-distribution ensembling effect. Secondary: Table 2 states the 39 strict families provide no labels to any fitted component, but fit_predict_arm() trains each base view on folds != fold over the full 86-family frame — other strict families do contribute labels (each assay is still out-of-fold, so this is a claim/implementation mismatch rather than same-assay leakage); and the extension from 5 to 10 seeds after observing p = 0.0625 raises an optional-stopping concern, with effective support of 5 genes / 2 studies by the authors' own admission.
Questions for authors
  1. What is the measured benefit of the certificate layer itself — reported as certified pipeline vs. identical predictors with certificates removed — on any endpoint?
  2. Since the router's score equals the specialists' to three decimals, does it ever route or refuse on the evaluation set? If not, what evidence supports routing as a contribution?
  3. If OligoAI's contribution is removed from the stack, does any KRAIG-specific component retain a positive interval on the strict set? Please report that contrast directly.
  4. Why does the strict-family protocol exclude OOF labels from the combiner but not from the three base rankers, contrary to Table 2?
  5. Given cross-seed agreement of 0.909 internally versus −0.11 to −0.14 externally, what would a deployable abstention rule look like, and can it be pre-registered rather than fitted post hoc?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.