acceptodds
Withdrawn from ICLR 2027

BasePrompt: Context as Prompt for Causal Genome Language Models in Zero-Shot RNA Fitness Prediction

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Causal genome language models (GLMs) are pretrained on long genomic context, whereas zero-shot RNA fitness benchmarks often expose isolated mature transcripts with little usable local context. We study whether a frozen causal GLM can use synthetic context to improve prediction without task-specific training. BasePrompt is a training-free procedure that constructs short synthetic 5′ and 3′ flanks around a target sequence. Its key mechanism is reverse-complement (RC) decoding, which opens a left-context channel for an otherwise forward-only model: we decode on the RC strand to synthesize an upstream flank, map back to the original strand, and then decode a downstream flank autoregressively. The main evaluation uses a shared fixed budget of three bases per side and reports matched RC-ensemble and direct-scoring results separately. On RNAGym, this fixed-budget comparison improves all three Evo2 backbones in the main text; for example, Evo2 7B rises from 0.276 to 0.292 Spearman ρ under matched RC scoring. Gains persist without RC ensembling. Assay-level and type-level decompositions are used as descriptive robustness views, making the type-averaged benchmark aggregation and the remaining heterogeneity explicit. Auxiliary DNA controls show that extension content matters in the tested settings, without establishing the causal mechanism of the RNA gains. We present BasePrompt as an inference-time context-construction method for causal genome language models, not as a reconstruction of true biological flanks.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
RNA fitness prediction, Prompting, Genome Language Model, DNA, Inference-time, Test-time

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_MwvCOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper investigates whether synthetic context can improve a frozen causal genomic language model’s RNA fitness predictions without task-specific training. The authors evaluate BasePrompt on 70 RNAGym assays using experimental variant-fitness measurements, with matched reverse-complement scoring and auxiliary DNA experiments. The results show modest improvements in aggregate variant-ranking performance.

Critical strengths

1.BasePrompt is a simple inference-time method that modestly improves agreement with experimentally measured RNA fitness without additional training.

2.Matched reverse-complement comparisons and direct-scoring controls help establish that the gains are not solely attributable to reverse-complement ensembling.

Critical weaknesses

1.The context-shortage motivation is insufficiently established. Short RNA sequences, absent genomic flanks, and insufficient context for fitness prediction are distinct issues. Table 3 includes sequences ranging from 45 to 5,592 nucleotides, and mature RNA boundaries do not necessarily indicate missing biological information. The paper does not establish which assays lack relevant context, whether that context is retrievable, or whether its absence impairs prediction.

2.Improved correlation with experimental measurements supports better benchmark ranking, but does not establish that synthetic flanks compensate for missing biological context. Although the authors appropriately acknowledge this distinction, the experiments do not adequately separate context-conditioning effects from changes introduced by scoring the synthetic bases themselves.

Questions for authors

1.Can you identify assays with genuinely missing biological context and evaluate whether gains depend on sequence length, available context, or controlled context truncation?

2.Do the improvements persist when scoring only the original RNA positions, excluding synthetic-flank likelihood terms, and when comparing against random or constant flanks on RNAGym?

3.How does BasePrompt compare with simpler alternatives, such as conservation or alignment-based scores, relevant thermodynamic scores, and sequence-feature models, with supervised methods evaluated separately under appropriate training protocols?

Reviewer_MER9OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

BasePrompt is a training-free method in which frozen Evo2 generates 3-nt synthetic flanks on both ends of a target RNA before scoring, with the left flank obtained by decoding on the reverse-complement strand. On zero-shot RNA fitness prediction in RNAGym, it yields modest gains across all three Evo2 models, e.g., raising Evo2 7B’s Spearman ρ from 0.276 to 0.292.

Critical strengths
  • The method is training-free and nearly cost-free, requiring no model changes or labels, generating flanks only once per assay on the reference sequence for reuse across all mutants, and lengthening inputs by only ~1.5% on average.
Critical weaknesses
  • Regarding writing and details, the main text repeats similar disclaimers (e.g., that the work does not establish a mechanism and the flanks are not biological) , and the Evo2 7B direct baseline is reported inconsistently (0.251 in Table 1 vs. 0.250 in Table 6).
Questions for authors
  • When stratified by mutation position (near the 5′ end vs. mid-sequence), are the gains concentrated at the beginning of the sequence?
  • What do the generated 3+3 flanks actually look like for each assay, and how similar are they across assays?
Reviewer_nCo4OpenReview ↗
Rating
1
Confidence
4
Informativeness
2
Presentation
3
Soundness
1
Read review
Paper summary

This manuscript presents BasePrompt, an inference-time context-construction method for zero-shot RNA fitness prediction using frozen causal genome language models. The manuscript is motivated by the limited flanking context available in RNA benchmarks and the directional constraint of causal models, which naturally extend sequences only downstream. BasePrompt maps each RNA reference sequence to the DNA alphabet and uses reverse-complement decoding to generate a synthetic 5′ flank, followed by autoregressive generation of a 3′ flank. These flanks are reused across variants from the same assay, and the complete augmented sequences are scored with the same frozen model. The authors evaluate the method on RNAGym, with the main comparison using three Evo2 backbones and three generated bases per flank, alongside auxiliary DNA pathogenicity and next-base prediction experiments. They report improvements in type-averaged benchmark scores across the three backbones, including an increase in Spearman correlation from 0.276 to 0.292 for Evo2 7B under matched reverse-complement scoring. The approach is straightforward to apply and requires neither backbone parameter updates nor training on assay-specific labels. However, the evaluation omits relevant mRNA-focused alternatives, limiting the assessment of its practical value on the corresponding RNA tasks. Overall, the manuscript presents a simple approach with reported improvements on the evaluated GLMs, but the broader significance of these gains for RNA fitness prediction remains insufficiently established.

Compared with the version submitted to ICLR 2026, the current manuscript clarifies its formulation and narrows several claims, but the concern about the adequacy of its mRNA-related comparisons remains unresolved. In their response during the ICLR 2026 review, the authors described RNA-FM, RNAErnie, and RiNALMo as “mRNA-specific baselines.” However, the original publications describe their pretraining corpora as ncRNA, so including these models does not address the absence of relevant mRNA-focused baselines. The previously suggested comparison with established biological RNA sequences used as fixed prediction contexts for the relevant assays has also not been provided.

Critical strengths

A simple inference-time procedure with reusable prompts

BasePrompt provides a concrete way to construct upstream and downstream context for a frozen causal genome language model, without updating model parameters or training on assay-specific labels. Generating the flanks once per reference sequence and reusing them across variants amortizes prompt generation over an entire assay, making the procedure straightforward to apply to large variant collections.

Positive results under matched evaluation protocols

Under the shared three-bases-per-flank budget and matched reverse-complement scoring, the manuscript reports improvements in all three aggregate metrics for the three evaluated Evo2 backbones. For example, Evo2 7B improves from 0.276 to 0.292 in type-averaged Spearman correlation. The direct-scoring controls also report improvements without reverse-complement ensembling, providing evidence that the observed benefit is not solely attributable to that evaluation choice.

Clearer formulation and more transparent reporting in the revision

The current manuscript explicitly defines the context-construction procedure and variant-effect score, and distinguishes synthetic conditioning sequences from recovered biological flanks. The assay-level and RNA-type breakdowns also make the heterogeneity of the gains visible, helping readers assess the scope of the aggregate results.

Critical weaknesses

Insufficient task-appropriate baselines for the mRNA-related evaluations

The comparisons establish improvements over the corresponding unprompted Evo2 backbones, but do not establish how the method compares with relevant mRNA-focused models on the corresponding assays. RNA-FM, RNAErnie, and RiNALMo are described in their original publications as pretrained on ncRNA corpora, so their inclusion does not resolve this gap. This limits the assessment of the practical significance of the reported gains for mRNA-related applications. A focused comparison with a representative mRNA-pretrained model on an appropriate subset of these assays, using a comparable zero-shot scoring protocol, would materially strengthen the evaluation.

Missing RNA context comparisons that directly assess the value of the generated prompts

The random and constant-sequence controls in Figure 6, and the native-versus-synthetic flank comparisons in Table 10, concern auxiliary DNA tasks. The primary RNA evaluation does not include the previously suggested comparison with established biological RNA sequences used as fixed prediction contexts. Consequently, the evidence does not establish whether generating reference-dependent flanks offers an advantage over suitable, straightforward context choices for the target RNA applications. A targeted comparison on biologically appropriate RNA assays, with context length, placement, and scoring controlled, would help resolve this question. Such an experiment would evaluate predictive utility and would not require demonstrating that the generated prompts function as regulatory elements in cells.

Questions for authors
  1. On RNAGym’s mRNA-coding assays, how does Evo2 7B with BasePrompt compare with mRNA-FM or CaLM in Spearman correlation when evaluated on the same variants using comparable zero-shot scoring and identical assay aggregation?
  2. Using Evo2 7B on a suitable mRNA assay, does a reference-generated 3′ extension achieve higher Spearman correlation than a fixed human β-globin 3′ UTR when the two extensions have the same length and the upstream context, variant set, and scoring procedure are held constant?

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

Related papers

Discussion (0)

Sign in to comment.