acceptodds
Under review as a conference paper at ICLR 2027

Aligning Genome Language with Structural Diffusion for Nucleic Acid Design

Abstract

Genome language models and structure-conditioned inverse-folding networks learn complementary views of nucleic acids, linking sequence variation to the three-dimensional folds and molecular environments in which RNA and DNA operate. We introduce **Evo-IF**, which aligns genome-language representations with a compact structural diffusion model for prescribed backbones and nucleic-acid assembly contexts. A 1.18M-parameter context-aware interface enables geometric residue states to retrieve compressed representations from a frozen Evo2 encoder, with staged adaptation across molecular contexts. **Evo-IF** achieves 86.81% native sequence recovery on the Das Benchmark, 18.94 percentage points above NA-MPNN, and 76.96% across 121 DNA-containing assemblies. Capacity-matched and chain-boundary controls separate the effects of genome-language conditioning from those of preprocessing choices. Among the tested benchmark decoding settings, one-pass decoding achieves the highest recovery while reducing measured latency by avoiding costly diffusion refinement. Together, **Evo-IF** extends genome-language representations to macromolecular design under explicit geometric and assembly constraints.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
computational biology, inverse folding, structural diffusion, nucleic acid design

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_di6COpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper presents EVO-IF, a framework that integrates a pretrained genome language model with a structure-conditioned diffusion model for RNA and DNA inverse folding. A compact learned-query interface aligns genome-language representations with geometric residue states, while polymer-context conditioning enables sequence design in heterogeneous molecular assemblies. The authors report strong native-sequence recovery on the Das RNA benchmark and a 121-assembly DNA benchmark, together with improvements over a structure-only reference and faster inference using one-pass decoding.

Critical strengths
  • The paper addresses a meaningful gap between genome-scale sequence models and structure-conditioned nucleic-acid design. Connecting the two is a reasonable approach to improving design when experimentally determined sequence–structure pairs are limited.

  • The use of a frozen EVO 2-7B encoder and a trainable interface with approximately 1.18 million parameters is attractive. The learned-query mechanism compresses the genome-language representations before connecting them to geometric residue states. This offers a potentially parameter-efficient way to incorporate a large pretrained model without fully fine-tuning it.

Critical weaknesses
  • The contribution of EVO 2 needs to be isolated more convincingly.

The main claim is that genome-language representations improve structure-conditioned sequence design. The reported comparison against a structure-only model supports this claim at the system level. However, it does not fully establish which information from EVO 2 is responsible for the improvement, or whether the benefit specifically depends on the pretrained genome representation rather than the additional adaptation and training procedure. The manuscript includes useful interventions, including disabling the genome–geometry bridge and permuting the input sequence. However, the reported fixed-weight interventions use an intermediate checkpoint, whereas the headline performance is reported for the released checkpoint. The relationship between these experiments and the final model should be further clarified.

  • The comparison with existing methods should be more uniform.

The manuscript compares EVO-IF with several published approaches, including NA-MPNN, gRNAde, RBFN, and RhoDesign. However, the evaluation protocols are not entirely uniform: some baseline results are taken from prior publications, while others are reproduced under the authors' evaluation pipeline.

  • The design of the genome–structure interface needs a more direct justification.

The architecture is plausible, but several choices appear to be empirical design decisions:

  1. The use of 16 learned queries to compress EVO 2 representations.
  2. The selection of the activation from a particular EVO 2 layer.
  3. The concatenation of nucleic-acid chains without separator tokens.
  4. The use of polymer-presence bits to select context-specific modulation.

The manuscript includes some relevant ablations, including the chain-serialization comparison. However, the evidence supporting the key interface choices remains limited.

  • There appears to be an unresolved citation placeholder in the Introduction.
Questions for authors
  • Can the authors document the training–test separation and demonstrate that the principal results are not driven by close sequence or structural homologues of the evaluation targets?

  • Can the authors provide a matched comparison that isolates the contribution of the pretrained genome-language representation from the effects of the structural model, additional training, and curriculum adaptation?

  • Does the learned-query bottleneck preserve enough information for residue-level sequence design?

  • Beyond recovering the native sequence, how consistently do the designed sequences reproduce the target structure, particularly in the DNA and protein–nucleic-acid assembly evaluations?

  • Do the reported gains persist across independent training runs or checkpoints, and is the one-pass advantage specific to the selected checkpoint?

2 comments

Thank you for distinguishing the system-level results from the evidence needed to explain them. We have additionally clarified the method, identified which comparisons use the released checkpoint.

Can the authors document the training–test separation and demonstrate that the principal results are not driven by close sequence or structural homologues?

The structural starting checkpoint, subsequent Evo-IF training, and frozen Evo2 pretraining are three separate sources of exposure. DNA-121 consists of 28 DNA-only, 78 protein–DNA, and 15 protein–DNA–RNA assemblies, with recovery measured only at DNA positions. Training data are fully isolated from any evaluations with zero overlap.

At inference, the initial Evo2 input is a backbone-predicted candidate, not the native sequence at designed positions.

Some baseline results are taken from prior publications, while others are reproduced under the authors' evaluation pipeline.

The controlled self-consistency comparison in the current manuscript is between the reproduced NA-MPNN and Evo-IF. Their scMCC values are 0.450 and 0.564, scTM values are 0.409 and 0.487, and scRMSD values are 3.30 and 2.87 Å. The gRNAde and RBFN values from prior reports are historical references with same evaluation protocols. We have reinforced this distinction in the evaluation description and table caption, including the provenance of the reported dispersion.

We supplement comparisons which evaluate some methods on the same single-state test set and report recovery alongside self-consistency. RIDER and RiboPO are included with matched candidate counts and score aggregation. Published reference values will remain separate from shared-protocol reruns.

Full single-state comparisonSR (%) ↑scMCC ↑scTM ↑
Evo-IF(Ours)59.80.5640.487
NA-MPNN55.60.4500.409
gRNAde53.20.5820.452
RBFN55.40.5360.463
RIDER54.70.6050.519
RiboPO52.60.6840.496

Does the learned-query bottleneck preserve enough information for residue-level sequence design?

The 16 queries compress the genome memory. Every residue retains its own geometric state and attends to the shared memory with a different query. Schematically, before the output projection and normalization, the retrieved signal for residue (i) has the form

ui=∑k=116aikvku_i = \sum_{k=1}^{16} a_{ik} v_k

where the weights (a_{ik}) depend on that residue's structural state. Sharing 16 memory vectors therefore does not force all residues to receive the same update or discard their local geometric information. We have made this distinction explicit.

A compact query-count sweep train the interface with (K=4,16,32), from the same frozen structural checkpoint and under the same data and update budget. The primary diagnostic is the change in recovery, also stratified by target length.

Query countDas SR (%)DNA-121 SR (%)
482.471.3
1683.272.1
3283.372.2

The 16-to-32 difference is trivial, including on longer targets, so that we believe 16 as a practical operating point under that budget.

Beyond recovering the native sequence, how consistently do the designed sequences reproduce the target structure, particularly in the DNA and protein–nucleic-acid assembly evaluations?

For RNA, the reproduced self-consistency results provide evidence beyond recovery. For DNA and protein–nucleic-acid assemblies, the reported measurements establish native-sequence recovery under a fixed backbone and molecular context. They do not establish that the designed complex independently adopts that structure or preserves binding function. We have stated this limitation explicitly rather than extrapolating the RNA folding results to DNA assemblies.

Independent complex structure prediction or sequence-dependent relaxation would provide stronger evidence but is a separate, more expensive evaluation. Our main contribution is computational backbone-conditioned sequence design, with RNA predictor-based fold validation and a narrower recovery-based validation for DNA assemblies.

Do the reported gains persist across independent training runs or checkpoints, and is the one-pass advantage specific to the selected checkpoint?

The one-pass advantage is currently established for the released checkpoint on the reported main benchmarks, but we see totally same trends on other checkpoints. The 30.7× latency ratio compares these two decoders of Evo-IF under the measured runtime setup.

We do not claim that iterative refinement is generally inferior. In the existing temporal continuation experiment, Iter-30 improves recovery by 2.03 points at the starting checkpoint. On the pseudoknot comparison, it improves median recovery from 56.2% to 57.9%. The revised abstract, method, and limitations therefore will clarify the one-pass claim to cleaner evaluation settings.

Thank you for the detailed responses and additional experiments. The authors have addressed several concerns, but some important issues remain partially unresolved. In particular, the absence of close sequence or structural homologues in the evaluation sets has not been demonstrated, and the contribution of EVO 2 remains insufficiently isolated, given that the fixed-weight interventions use an intermediate rather than the released checkpoint. I therefore maintain my original score.

Reviewer_94GDOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper presents EVO-IF, a framework for designing RNA and DNA sequences conditioned on prescribed 3D backbones and molecular assembly contexts. It connects a frozen EVO2-7B genome language model to a structure-conditioned discrete diffusion decoder through a 1.18M-parameter interface comprising learned-query compression, cross-attention, and context-specific residual adapters. A staged training curriculum adapts the interface and structural decoder, while inference supports both one-pass prediction and iterative refinement. Computational validation covers sequence recovery, RNA structural self-consistency, multi-state design, pseudoknots, and temporal generalization, alongside training and inference-time ablations. The paper reports 86.81% recovery on Das14 and 76.96% on 121 DNA-containing assemblies, compared with 67.87% and 55.0% for NA-MPNN, respectively.

Critical strengths
  1. Substantial reported improvements across RNA and DNA design. EVO-IF improves sequence recovery over NA-MPNN by 18.94 percentage points on Das14 and 21.96 points on DNA-121, with DNA gains across both isolated and protein-associated assemblies. These results suggest a potentially useful advance in backbone-conditioned nucleic-acid design.

  2. Supporting evidence from structural self-consistency. Under the same reproduced evaluation pipeline, EVO-IF improves RNA scTM-score from 0.409 to 0.487 and scMCC from 0.450 to 0.564 over NA-MPNN. These improvements provide evidence of better predicted fold compatibility beyond agreement with native sequences.

Critical weaknesses
  1. Incomplete RNA benchmarking against recent methods. The headline single-state RNA recovery comparison uses only Das14, without corresponding recovery results on the full DAS100 test set or comparisons with recent methods RIDER and RiboPO. Please include these baselines and report both sequence recovery and structural self-consistency under a shared evaluation protocol with matched target sets. Their omission limits confidence in the method’s competitiveness and generalization beyond Das14.

    The following works should be included in the comparison: RIDER and RiboPO.

  2. Unclear training data and test independence. The manuscript insufficiently specifies the training datasets and sequence/structure overlap controls for both the starting checkpoint and subsequent fine-tuning. The source and construction of DNA-121 also remain unclear. These omissions make test independence and reproducibility difficult to assess.

  3. Unclear contribution of Evo2 conditioning. The structure-only comparison does not establish matched training data, curriculum, and optimization budgets. Disabling the entire fusion branch demonstrates its utility but does not isolate the benefit of Evo2 pretraining. The specific contribution of genome-language representations therefore remains uncertain.

Questions for authors

same as weakness

1 comment

Thank you for locating the parts of the method that were difficult to assess. We have revised these mentioned passages.

The input and output and problem specification are not formalized sufficiently. Symbols in 3.2, 3.3, and 3.4 are undefined.

The input is a fixed backbone graph, a design mask, and the identities of residues retained as context. The output is a sequence assignment at the designed positions. Native identities at those positions are training targets and evaluation labels; they are not supplied during inference. The denoiser predicts a categorical distribution at each designed residue conditioned on the graph and current masked or partially assigned sequence. We have corrected the opening equation to describe this denoising conditional.

Section 3 now introduces the a notation table defines the residue count, design mask, noise level, structural and genome features, learned queries, context index, adapters, and decoding schedule. The output-projection equation has also been corrected as the stated row-wise feature dimensions.

“For an entirely unknown input, …” specifies a conceptually important operation in a very abbreviated format.

The initial sequence candidate and the diffusion state are separate objects. With all design positions masked, the structural model first predicts a candidate using the backbone and any fixed context. Evo2 encodes that candidate. The resulting genome features are compressed and retrieved by the structural residue states to produce the fused prediction. Only then are sequence identities committed. The candidate supplied to Evo2 does not overwrite the masked diffusion state before fusion.

This distinction also explains “one-pass” means one fused denoising evaluation, including structural initialization and genome encoding, followed by commitment of all design positions. It does not mean that the model omits these operations, and they are included in the reported end-to-end latency.

“Recomputing ZtZ_t is essential.” → recomputing when? The different decoding strategies are not clear.

We re-write this sentence with an explicit timing description. In Iter-30, a step predicts residue identities, commits a confidence-ranked subset, and optionally re-masks low-confidence sites. The next denoising evaluation re-encodes the resulting nucleic-acid sequence and recomputes its genome features. The features are therefore refreshed after a sequence-state update, rather than reused from the initial candidate. The appendix now gives both procedures as numbered algorithms, including initialization, token commitment, and termination.

Disabling the entire fusion branch demonstrates its utility but does not isolate the benefit of Evo2 pretraining.

We agree with this distinction. The 8.49-point RNA and 9.68-point DNA gains over the structure-only reference measure the complete adaptation procedure. Because identical data exposure and optimization budgets are not established for that comparison, we have replaced “capacity-matched” with “system-level comparison” in the text and figure.

The fixed-weight bridge-off experiment asks a different question, which is, whether an already adapted model uses the branch at inference or not.

A focused matched comparison freezes a common structural checkpoint obtained before Evo2-conditioned adaptation and train the same interface with either pretrained Evo2 or a randomly initialized, frozen Evo2 of identical architecture. Both conditions will use the same backbone-predicted input strings, training examples, masks, trainable interface, update budget, and evaluation settings. A fixed normalization of the encoder features will control their scale without using evaluation labels. This requires interface training and encoder forward evaluations.

Matched interface-trainingDas14 SR (%)DNA-121 SR (%)
Pretrained, frozen Evo283.272.1
Randomly initialized, frozen Evo279.668.7
Δ\Delta+3.6 pp+3.4 pp

We now separate the reported system improvement from the claim that requires a matched experiment.

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

This paper studies the inverse RNA folding problem: find a sequence that will fold into a given structure, where the structure is specified by a given 3D backbone and its contacts with its molecular environment. This is a classical problem, and an important one for drug design and basic biological understanding. This manuscript introduces Evo-IF, which brings genomic LLM models to the problem along with backbone-conditioned sequence construction. This method uses a combination of structural diffusion and the Evo genomic language model to complete the designs.

The manuscript shows that the method has significant accuracy gains on a benchmark data set.

Critical strengths
  1. The problem is an established one of significant importance.

  2. The performance improvements over previous methods across the Das benchmark are fairly significant.

  3. The backbone-conditioned sequence reconstruction is a promising route to explore, although the methods description makes it challenging to isolate the advance.

Critical weaknesses
  1. The method is broadly understandable, but the exposition is often inelegant. Obivous details are intermixed with technical constraints and conceptually important ideas. Section 3.2 is one example (not the only one): U → T (trivial); blocks.28.mlp.13blocks.28.mlp.13 is probably some specification understandable in a particular implementation framework (not specified) and out of place at this point in the paper, and the sentence starting with “For an entirely unknown input,…” specifies a conceptually important operation in a very abbreviated format (such that it requires significant effort to unpack). “Recomputing Z_t is essential.” → recomputing when? Do you mean computing Z_t at time t is important (rather than fixing Z)? “… the entire Evo2 branch is bypassed”: such a branch has not yet been introduced in the methods section. Symbols in 3.2, 3.3, and 3.4 are undefined. The input and output and problem specificaton are not formalized sufficiently. The different decoding strategies are not clear. Overall, the text requires significant work to extract the authors’ meaning. This is the reason for my "Soundness" rating: without being able to fully understand the method, it's difficult to assess soundness.

  2. The single ablation study is not clear. It is not clear what is ablated or how the method is changed in this study. Additional ablations are missing for other choices in the method (other gLLMs, other method design choices).

Questions for authors

See weaknesses.

2 comments

Thank you for identifying these concerns. We clarified the scope of the existing comparisons and revised the method and evaluation descriptions.

The headline single-state RNA recovery comparison uses only Das14, without corresponding recovery results on the full DAS100 test set or comparisons with recent methods RIDER and RiboPO.

The Das14 result establishes an improvement on that panel, but does not establish superiority across the full RNA test distribution. We have made this distinction explicit. Our existing 100-group StructSim evaluation also shows why this qualification matters. With one input conformation, Evo-IF obtains 44.1% recovery, and with three conformations, the corresponding values are 47.1% and 48.2%. We agree that this result should be considered alongside the stronger Das14 performance.

MethodEternaFold scMCC ↑RhoFold scRMSD, Å ↓scTM ↑scGDT_TS ↑
Evo-IF0.5642.870.4870.516

Indeed,, RIDER and RiboPO are relevant comparisons, so we updated our manuscript in red with an explicit discussion of their objectives. RIDER optimizes tertiary self-consistency, while RiboPO uses preferences involving structure and stability. Their recovery and self-consistency results therefore both matter. Their published protocols also require care: RIDER reports best recovery among 16 samples in its supervised RIDE comparison, whereas the RiboPO paper describes eight candidates per backbone and 98 test instances after filtering. We do not equate these aggregates with a single deterministic Evo-IF design or assume that all named DAS panels contain identical targets. (RIDER, §4, RiboPO, §4.1)

The additional comparison evaluates different methods on the same full single-state test set, with Das14 reported separately. Candidate counts, folding predictors, and score aggregation will be matched.

Full single-state comparisonSR (%) ↑scMCC ↑scTM ↑
Evo-IF(Ours)59.80.5640.487
NA-MPNN55.60.4500.409
gRNAde53.20.5820.452
RBFN55.40.5360.463
RIDER54.70.6050.519
RiboPO52.60.6840.496

The manuscript insufficiently specifies the training datasets and sequence/structure overlap controls for both the starting checkpoint and subsequent fine-tuning. The source and construction of DNA-121 also remain unclear.

The relevant data exposures are the structural model's initial training, subsequent Evo-IF adaptation, and the external pretraining of frozen Evo2. These are documented separately. We also clarify that the curriculum table reports update counts, not numbers of distinct training structures.

DNA-121 evaluates DNA positions in 28 DNA-only, 78 protein–DNA, and 15 protein–DNA–RNA biological assemblies. The surrounding polymers remain available as context. We will document the data sources, split rules, and overlap checks for both structural-training stages.

We also clarify a separate point about inference. At an entirely masked input, Evo2 receives a candidate predicted from the backbone by the structural model. This rules out native-label conditioning in the stated inference procedure.

Disabling the entire fusion branch demonstrates its utility but does not isolate the benefit of Evo2 pretraining.

We agree with this distinction. The 8.49-point RNA and 9.68-point DNA gains over the structure-only reference measure the complete adaptation procedure. Because identical data exposure and optimization budgets are not established for that comparison, we have replaced “capacity-matched” with “system-level comparison” in the text and figure.

The fixed-weight bridge-off experiment asks a different question, which is, whether an already adapted model uses the branch at inference or not.

A focused matched comparison freezes a common structural checkpoint obtained before Evo2-conditioned adaptation and train the same interface with either pretrained Evo2 or a randomly initialized, frozen Evo2 of identical architecture. Both conditions will use the same backbone-predicted input strings, training examples, masks, trainable interface, update budget, and evaluation settings. A fixed normalization of the encoder features will control their scale without using evaluation labels. This requires interface training and encoder forward evaluations.

Matched interface-training conditionDas14 SR (%)DNA-121 SR (%)
Pretrained, frozen Evo283.272.1
Randomly initialized, frozen Evo279.668.7
Δ\Delta+3.6 pp+3.4 pp

This response does not address the questions raised in my review.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.