{"id":"43e028f2-f5bf-40a0-8f3c-bc97617bfce9","arxiv_id":"2506.08936","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"high","formal_verification":"none","parameter_count":3,"one_line_summary":"Codon-level fusion of DNA, mRNA, and protein embeddings improves prediction on some molecular property tasks, but the reported consistent gains are not supported by the paper's own results.","lead":"This paper combines DNA, mRNA, and protein language model embeddings for molecular property prediction using three simple fusion methods aligned at the codon level. The best variant improves over single-modality models on four of five tasks, but the paper's claim of consistent gains is contradicted by its own table.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Table 1 contradicts the paper's central claim: concatenation and cross-attention underperform the best unimodal baseline on 3 and 2 of 5 tasks, so 'consistently outperform' is not supported as written.","rationale":"Chosen concern: the self-contained inconsistency between the central claim and Table 1. This is the most load-bearing because it directly targets what the paper promises, and it does not depend on assumptions about dataset processing. The codon-alignment issue raised by the reader is also serious and could invalidate the method's biological motivation, but the paper's datasets have max lengths divisible by 3 and the Methods mention start/stop codons, so a CDS-only (or ORF-trimmed) interpretation is not ruled out; the Table 1 inconsistency is decisive without any external knowledge. I therefore disagree with the reader's coded weakest_assumption, although the reader's broader rationale also mentions the table issue. Since the concern supports the reader's rejection, the verdict is unchanged.","tokens_in":12984,"tokens_out":15306,"duration_ms":149367,"concrete_test":"Take Table 1 and compute the 3x5 sign matrix D[i,j] = score(fusion_i, task_j) - score(best_unimodal, task_j). Any negative entry invalidates the literal 'consistently outperform' claim; the actual count from the table is 6 negatives. If the authors instead claim per-task best-variant selection, the paper must specify the selection procedure (e.g., validation-based) and report the variance of the selected model over multiple seeds; otherwise the claim is unfalsifiable.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central assertion, stated in §3 as 'Across all the tasks, fusion-based models consistently outperform the best single-modality baselines,' is not supported by the paper's own Table 1 under the natural reading that each fusion variant should beat the best unimodal baseline. Using the best unimodal baseline per column (CoV-Vac: RNA-FM 0.841; Fungal: Nucleotide Transformer 0.804; E. coli: RNA-FM 52.949; mRNA Stab.: RNA-FM 0.553; Ab1: RNA-FM 0.743), the three fusion variants produce losses in 6 of 15 comparisons: Concatenation is worse on CoV-Vac (0.831 vs 0.841), E. coli (50.280 vs 52.949), and mRNA Stab. (0.539 vs 0.553); Cross-Attention is worse on CoV-Vac (0.828 vs 0.841) and mRNA Stab. (0.550 vs 0.553); MIL+Entropy is worse on E. coli (52.107 vs 52.949). The only interpretation under which 'consistently' survives is per-task selection of the best fusion variant, but no such selection rule is stated, and the surviving margins (e.g., +0.010 Spearman on mRNA Stab.) are small with no error bars or significance tests. The headline therefore fails as written and needs at minimum a rewritten, weaker claim.","agreement_with_reader":"disagree"},"referee_report":{"model":"deepseek-v4-flash","summary":"BioLangFusion proposes fusing frozen embeddings from DNA, mRNA, and protein language models by first aligning them at a claimed codon-level resolution and then applying one of three fusion heads: concatenation, entropy-regularized gated attention pooling (MIL), or cross-modal multi-head attention. The method is evaluated on five molecular property prediction datasets (CoV-Vac, Fungal, E. coli, mRNA Stab., Ab1) against several unimodal baselines. The abstract, introduction, and Section 3 claim that the fusion variants consistently outperform the best single-modality baselines, with additional interpretability analysis of learned attention weights.","tokens_in":13279,"tokens_out":4940,"duration_ms":54954,"significance":"If the claims were fully supported, the paper would offer a lightweight, modular recipe for combining existing biological foundation models without retraining, which could be practically useful. The interpretability analysis through modality attention is a nice qualitative addition, and the paper explicitly lists the frozen backbones, fusion dimensions, and training hyperparameters. However, the central empirical claim is contradicted by the paper's own Table 1 under any straightforward reading, and the biological alignment assumption in Section 2.1 is questionable for mRNA sequences with untranslated regions. The paper provides no error bars, significance tests, or code, so the quantitative evidence is currently too weak to establish the claimed advantage.","major_comments":[{"comment":"The statement 'Across all the tasks, fusion-based models consistently outperform the best single-modality baselines' is not supported by the reported numbers. Taking the best unimodal baseline per column (RNA-FM for CoV-Vac, E. coli, mRNA Stab., Ab1; Nucleotide Transformer for Fungal), the fusion variants lose in 6 of 15 comparisons: Concatenation underperforms on CoV-Vac (0.831 vs 0.841), E. coli (50.280 vs 52.949), and mRNA Stab. (0.539 vs 0.553); Cross-Attention underperforms on CoV-Vac (0.828 vs 0.841) and mRNA Stab. (0.550 vs 0.553); MIL+Entropy underperforms on E. coli (52.107 vs 52.949). The claim can only survive if one selects the best fusion variant per task after seeing the test set, which is not stated as the evaluation protocol. The claim must be rewritten to reflect the observed pattern (e.g., some fusion variants help some tasks) and supported with uncertainty quantification.","section":"Section 3, Table 1"},{"comment":"The alignment procedure assumes that every mRNA position is coding: the protein frame is defined as T' = T/3, where T is the full mRNA length, and the mRNA embedding is downsampled by non-overlapping 3-nucleotide mean pooling. However, the datasets in Table 2 are full mRNA sequences, which include 5' and 3' untranslated regions; the actual protein-coding sequence is generally shorter than T/3 and is not aligned to the start of the mRNA. Consequently, the 'codon-level' correspondence between the pooled RNA tokens and the protein residues is biologically incorrect for these sequences. The authors need to either translate the coding sequence and align only the CDS, or explicitly model UTRs separately; otherwise the central alignment mechanism does not do what the paper claims.","section":"Section 2.1, Eq. (1)"},{"comment":"The preprocessing step truncates RNA sequences longer than 1000 tokens, yet Table 2 reports max lengths up to 3066 and shows large drops between raw and used counts (Fungal: 7056 to 3138; E. coli: 6348 to 4450; mRNA Stab.: 41123 to 23929). These numbers are inconsistent with a straightforward truncation to 1000 tokens (the post-truncation max length would be 1000), and the table caption does not state whether 'Max Length' refers to raw or used sequences. More importantly, the paper does not analyze how truncation interacts with the T/3 alignment or whether the discarded long sequences are systematically different. The authors should report post-truncation statistics, explain the discrepancy, and justify that the truncated subset remains representative.","section":"Section 2.1, Table 2"},{"comment":"No error bars, confidence intervals, or significance tests are reported, and the evaluation appears to use a single split with fixed training runs. Several reported gains are very small (e.g., +0.003 Spearman on Ab1 for concatenation, +0.010 on mRNA Stab. for MIL+Entropy), so without repeated seeds or statistical testing they cannot be distinguished from noise. The authors should add multiple random seeds, report mean and standard deviation, and, where appropriate, paired significance tests (e.g., Wilcoxon or bootstrap) against the best unimodal baseline.","section":"Section 3, Table 1 and Appendix A.4"}],"minor_comments":[{"comment":"There is a typo in the abstract: 'a an mRNA sequence' should be 'an mRNA sequence'; also 'can destabilizing an mRNA hairpin' in the Introduction should be 'can destabilize an mRNA hairpin'.","section":"Abstract and Section 1"},{"comment":"The text contains typos and grammatical errors, e.g., 'Moverove' should be 'Moreover', 'having shared projection layers accross modality' should be 'across modalities', and 'brings performance gain only on one (Ab1) out of the five data set' should be 'datasets'.","section":"Appendix A.3"},{"comment":"Several references are duplicated or incomplete: Dalla-Torre et al. appears as 2022, 2023, and 2024 entries; Elnaggar et al. as 2021a and 2021b; Lin et al. as 2022a and 2022b; and some entries contain placeholder-like DOIs (e.g., '10.1101/2023.01.01.123456'). The reference list should be consolidated and verified.","section":"References"},{"comment":"The transposed convolution for DNA upsampling is described inconsistently: Eq. (1) says k=2, s=2, while List 1 states ConvTranspose1D(kernel=3, stride=2, padding=2) and List 2 states ConvTranspose1D(kernel=3, stride=2, padding=2). Please reconcile the kernel size and padding so the implementation matches the method description.","section":"Eq. (1) and List 1"},{"comment":"The Nucleotide Transformer v2 100M model is listed with an embedding dimension of 4,107, which is unexpectedly large for a 100M-parameter model and differs from typical reported hidden dimensions for this model. Please verify this number and, if it is a concatenation of multiple representations, state that explicitly.","section":"Table 3"}],"recommendation":"major_revision","confidential_remarks":"I do not share the reader's circularity concern: the only fitted scalar hyperparameter (lambda) is selected on validation, and the attention weights are a post hoc readout, not a circular prediction. The serious issues are (1) the central 'consistent outperformance' claim is demonstrably false from Table 1, and (2) the codon-level alignment is biologically invalid for full mRNA sequences with UTRs. Both are fixable in principle, but the revision needs to be substantial: reword the claims, add statistical rigor, and either restrict the analysis to coding sequences or explicitly model UTRs. If the authors cannot address the alignment issue, the contribution as framed would no longer stand."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The paper puts a genuinely useful idea on the table: freeze DNA, mRNA, and protein LMs and align their embeddings at codon level before fusing. That alignment is the real contribution—nobody else in the cited literature does it—and the three fusion heads are standard but sensibly chosen. The ablation showing entropy regularization sharpens attention weights is also a nice practical detail. If the codon-level correspondence actually holds, the approach is a plug-and-play tool that costs no base-model retraining.\n\nSoft spots, in order. First, the central claim in Section 3—'Across all tasks, fusion-based models consistently outperform the best single-modality baselines'—does not survive its own Table 1. Under the natural reading, concatenation loses to the best unimodal on CoV-Vac, E. coli, and mRNA Stab.; cross-attention loses on CoV-Vac and mRNA Stab.; even the flagship MIL+Entropy loses on E. coli. The only way 'consistently' works is if you select the best fusion variant per task, but no such selection rule is stated. Second, no error bars or significance tests, and the winning margins are 0.01–0.02 Spearman; that is not enough to separate signal from noise. Third, the codon alignment assumes the protein frame length T/3, meaning every mRNA position is coding. The datasets are full transcripts with UTRs, so the alignment is biologically wrong for those inputs and the 'codon-level correspondence' claim is overstated. Fourth, fusion models have extra trainable layers; the comparison is not capacity-matched. The dataset preprocessing is also under-described: Table 2 shows used counts far below raw, but the text only says 'truncate', which does not explain the drop.\n\nNone of these are fatal to the underlying idea, but they are fatal to the paper as written. The experiments do not support the headline. With corrected claims, error bars, and a capacity-matched baseline, this could become a solid contribution.\n\nWho is this for? Researchers looking for cheap multimodal representations for molecular properties. It is a reasonable workshop-level contribution, and the alignment idea deserves scrutiny. I would send it to peer review rather than desk reject, but I would expect substantial revision.","headline":"The codon-level alignment idea is worth a look, but the paper's central claim of consistent gains is contradicted by its own Table 1 and needs major revision.","tokens_in":13827,"tokens_out":2904,"would_cite":false,"duration_ms":32125,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Fusing frozen DNA, mRNA, and protein language models at the codon level outperforms the best single-modality baseline on five molecular property prediction tasks.","keywords":["multimodal fusion","language models","DNA","mRNA","protein","molecular property prediction","codon alignment","attention pooling"],"falsifier":"Permute the order of codon-level positions in the DNA and protein embeddings relative to the mRNA embedding (or vice versa) while keeping each modality's own sequence order intact, and re-run the five tasks with the same fusion heads. If randomly misaligned fusion matches the codon-aligned fusion, the biological alignment is not the source of the gains; if performance drops, the alignment frame is confirmed as the carrier of the improvement.","tokens_in":12772,"feed_emoji":"🧬","tokens_out":15056,"duration_ms":152440,"temperature":0.7,"pith_summary":"BioLangFusion sets out to show that DNA, mRNA, and protein sequences carry complementary, jointly useful information that can be combined without retraining any large foundation model. It aligns frozen embeddings from a DNA model (6-mer tokens), an mRNA model (single-nucleotide tokens), and a protein model (amino-acid tokens) onto a shared codon grid, then attaches one of three small fusion heads. On five molecular property prediction tasks—vaccine mRNA degradation, fungal expression, E. coli expression classification, mRNA stability, and antibody expression—the paper reports that the fusion approach beats the strongest unimodal baselines, with entropy-regularized attention pooling the most consistently strong variant. The appeal is practical: no end-to-end retraining, a small parameter overhead, and attention weights that indicate which modality the model actually uses for each task.","feed_headline":"Fusing DNA, RNA, protein models beats single-modality baselines","feed_subtitle":"Small fusion heads on frozen DNA, mRNA, and protein language models lift molecular property prediction on five tasks.","key_machinery":"The load-bearing object is the codon-level alignment frame, in which every position $t$ of the three embedding tensors is taken to name the same biological codon. The DNA embeddings (originally over 6-mers) are upsampled with a stride-2 transposed convolution; the mRNA embeddings (per nucleotide) are downsampled by non-overlapping mean pooling over windows of three; the protein embeddings already sit on the frame. On top of that frame sit three fusion heads: (i) codon-level concatenation, with a modality-specific MLP projection for the 4107-dimensional DNA embedding so it does not dominate the 640-dimensional RNA and 320-dimensional protein vectors; (ii) entropy-regularized gated attention pooling, which treats the three modalities as a bag of instances, computes a softmax weight per modality from mean-pooled summaries, and adds a negative-entropy term to keep the weights from collapsing to uniform; and (iii) cross-modal multi-head attention, which concatenates the projected modalities along the time axis into a joint context $C = [H_{\\mathrm{DNA}}; H_{\\mathrm{RNA}}; H_{\\mathrm{Prot}}] \\in \\mathbb{R}^{3T' \\times d}$, lets each modality query that context, and merges the updated streams by residual averaging plus layer normalization. All three feed the same fixed TextCNN prediction head, so the reported differences come from the fusion mechanism itself.","core_discovery":"The paper's central claim is that a shared codon-level frame is enough to make frozen single-modality embeddings behave like a multimodal representation: after upsampling the DNA model's 6-mer tokens and downsampling the mRNA model's per-nucleotide tokens onto the protein frame of length $T' = T/3$, each aligned position nominally corresponds to the same biological codon across all three modalities. On that frame, the authors test three fusion heads and report that fusion beats the best unimodal baseline across the five tasks, with entropy-regularized attention pooling the most consistent winner (e.g., Spearman 0.864 vs 0.841 on CoV-Vac and accuracy 52.107% vs 49.017% on E. coli) while cross-modal multi-head attention takes the lead on E. coli classification. The claim is deliberately economical: the base models are frozen, no knowledge distillation or weight merging is needed, and the only trainable parts are the alignment projections, fusion head, and a fixed TextCNN predictor. Along the way, the attention weights are shown to track expected biology—mRNA emphasized for mRNA stability, protein emphasized for E. coli expression.","pith_inferences":["The paper does not restrict alignment to the annotated coding sequence; a natural follow-up is to compare the full-transcript $T/3$ alignment against a CDS-only alignment, which would reveal whether gains come from true codon correspondence or from pooling extra UTR-derived signal.","The recipe is model-agnostic: the same fusion heads could be applied to other frozen encoders (e.g., larger ESM-2 checkpoints or codon-aware mRNA models), a test the paper leaves open since it uses one checkpoint per modality.","The attention weights are presented as interpretation; a causal check would be to mask or corrupt one modality at inference and confirm that prediction accuracy drops most on tasks where that modality's learned attention is highest.","If the gains survive random permutation of codon positions across modalities, the 'biologically meaningful correspondence' explanation would be in doubt; if they vanish, the alignment frame is doing real work."],"forward_implications":["Adding any of the three fusion heads to frozen DNA, mRNA, and protein embeddings can improve molecular property prediction over the best single-modality baseline, with no end-to-end retraining of the base models.","Entropy-regularized attention pooling is the most consistently strong fusion head, so it is the natural default for low-budget multimodal setups on similar tasks.","The learned modality attention weights provide an interpretable diagnostic: they highlight mRNA on mRNA-stability prediction and protein on E. coli expression classification, suggesting the fused model can expose which biological layer drives a phenotype.","Codon-level alignment itself contributes to the gains: the paper's ablation shows vanilla concatenation without alignment scores 0.818 vs 0.831 on CoV-Vac and 46.208% vs 50.280% on E. coli, so the alignment step is not inert.","Cross-modal token-level attention is not uniformly better than sequence-level modality attention; it wins on E. coli classification but trails on most other tasks, implying token-level cross-talk is only sometimes worth its extra complexity."],"supporting_citations":[{"why":"Supplies the frozen Nucleotide Transformer DNA embeddings used as the DNA modality in every fusion variant.","marker":"Dalla-Torre et al., 2023"},{"why":"Supplies RNA-FM, the frozen mRNA encoder whose per-nucleotide embeddings are downsampled to the codon grid.","marker":"Chen et al., 2022"},{"why":"Supplies ESM-2, the frozen protein encoder whose amino-acid embeddings define the reference frame.","marker":"Lin et al., 2022a"},{"why":"Provides the gated attention multiple-instance-learning mechanism on which the entropy-regularized attention pooling head is built.","marker":"Ilse et al., 2018"},{"why":"Provides four of the five datasets and the train/test splits used for the evaluation.","marker":"Li et al., 2023b"},{"why":"Provides the Ab1 antibody expression dataset used as the fifth benchmark.","marker":"Yazdani et al., 2024"}],"fun_headline_variants":["Codon-level fusion of DNA, RNA, protein models wins on 5 tasks","Frozen language models fuse at codon level to beat unimodal baselines","BioLangFusion: align at codon, then fuse: better property prediction","Simple fusion heads on frozen bio models top five tasks","Codon frame aligns DNA, mRNA, protein models for higher accuracy"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing assumption is that every mRNA position is translated, so the protein sequence has exactly one-third the length of the transcript and the $T/3$ codon grid puts the same biological location in correspondence across DNA, mRNA, and protein; the datasets include untranslated regions where that correspondence breaks.","fun_headline_variants_meta":{"raw":{"variants":["Codon-level fusion of DNA, RNA, protein models wins on 5 tasks","Frozen language models fuse at codon level to beat unimodal baselines","BioLangFusion: align at codon, then fuse: better property prediction","Simple fusion heads on frozen bio models top five tasks","Codon frame aligns DNA, mRNA, protein models for higher accuracy"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000612,"raw_usage":{"total_tokens":2850,"prompt_tokens":952,"completion_tokens":1898,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":568,"completion_tokens_details":{"reasoning_tokens":1804}},"tokens_in":568,"tokens_out":1898,"duration_ms":15312,"temperature":1.0,"reasoning_tokens":1804,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T04:58:52.670352+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Permute the order of codon-level positions in the DNA and protein embeddings relative to the mRNA embedding (or vice versa) while keeping each modality's own sequence order intact, and re-run the five tasks with the same fusion heads. If randomly misaligned fusion matches the codon-aligned fusion, the biological alignment is not the source of the gains; if performance drops, the alignment frame is confirmed as the carrier of the improvement.","supporting_citations":[{"cited_title":"H., Oteri, F., Dallago, C., Trop, E., Sirelkhatim, H., Richard, G., et al","cited_arxiv_id":null,"evidence_quote":"Supplies the frozen Nucleotide Transformer DNA embeddings used as the DNA modality in every fusion variant."},{"cited_title":"Rna-fm: A foundation model and benchmark for functional rna embeddings","cited_arxiv_id":null,"evidence_quote":"Supplies RNA-FM, the frozen mRNA encoder whose per-nucleotide embeddings are downsampled to the codon grid."},{"cited_title":"Attention-based deep multiple instance learning","cited_arxiv_id":null,"evidence_quote":"Provides the gated attention multiple-instance-learning mechanism on which the entropy-regularized attention pooling head is built."},{"cited_title":"Helm: Hierarchical embeddings for language modeling of mrna sequences","cited_arxiv_id":null,"evidence_quote":"Provides the Ab1 antibody expression dataset used as the fifth benchmark."}],"review_version":1}