{"id":"2e60d374-5c2a-4dc5-897b-2adcae7d75a0","arxiv_id":"2506.18940","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"eccDNAMamba is a bidirectional Mamba-2 model with circular sequence augmentation and span masking that classifies eccDNA better than existing genomic foundation models on two benchmarks.","lead":"A new AI model, eccDNAMamba, reads circular DNA molecules up to 200,000 base pairs long by connecting the sequence ends. It outperforms several existing genomic models on two cancer-related classification tasks, suggesting circular DNA carries learnable signals that truncated models miss.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The 200-kbp scaling claim is not established: §4.5 fixes the model at 30,000 BPE tokens, while a 200 kbp sequence at the reported 5.16 bp/token compression needs ~38,800 tokens, so 'full-length end-to-end' is unsupported.","rationale":"The paper's contribution rests on two pillars: a circular-aware encoder and long-context scaling. The reader's weakest assumption—that circular augmentation is unvalidated on the actual model—is real but secondary: even if the augmentation were inert, the model would still be a bidirectional BPE-Mamba with useful long-context behavior. The scaling claim is more directly load-bearing because it appears in the abstract, the discussion, and the 'only model robust' statement. Section 4.5's token-matched setting effectively caps eccDNAMamba at 30,000 BPE tokens, and the reported BPE compression rate of 5.16 bp/token means a 200 kbp sequence needs roughly 38,800 tokens. Without evidence that long test sequences tokenize below the cap, the 'full-length, end-to-end, up to 200 kbp' statements are unsupported. The concern is concrete and testable: tokenize a 200 kbp test sequence and inspect the training config. I am not claiming the authors truncated deliberately; the manuscript simply does not report this, and the ambiguity matters. I set the verdict to CONDITIONAL rather than REJECT because the concern can be settled by a simple check and does not necessarily invalidate all results.","tokens_in":12847,"tokens_out":8804,"duration_ms":88216,"concrete_test":"Reproduce the long-sequence evaluation pipeline: load the released tokenizer and model configuration, take a random 200,000 bp sequence from the paper's 10–200 kbp test set, and compute len(tokenizer(seq)['input_ids']). Also inspect the fine-tuning script's max_seq_length. If the tokenized length exceeds max_seq_length (or the 30,000-token setting described in §4.5), then the Table 2 long-row results rely on truncation, and the 'up to 200 Kbp' scaling claim must be revised. If all 200 kbp sequences fit under the limit, the concern is resolved.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The paper's headline scaling claim—'scales to sequences up to 200 Kbp' and 'processes complete sequences end-to-end'—is not supported by the reported setup. In §4.5, the Caduceus comparison fixes eccDNAMamba at 30,000 BPE tokens: 'we input 30,000 BPE tokens into our model, and simultaneously provided 30,000 bases ... to Caduceus.' The pretraining corpus is reported to compress at 5.16 bp/token (§4.1 and A.3). A 200,000 bp sequence at that compression needs roughly 38,800 tokens, so sequences in the upper part of the 10–200 kbp range cannot be fed in full within a 30,000-token budget. If long sequences were truncated or cropped to fit the context window, then the 'full-length end-to-end' description and the 'only model remains robust across all evaluation metrics' conclusion are not established. The manuscript does not report the actual token counts of the long test sequences or the model's max_length, leaving this load-bearing claim unverified. This is a concrete, checkable issue rather than a stylistic objection.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes eccDNAMamba, a bidirectional Mamba-2 encoder pretrained on a corpus of ~120,000 eccDNA sequences with a span-masked language modeling objective, BPE tokenization, and a circular augmentation strategy that appends the first 64 tokens to the end of each sequence. The model is evaluated on two binary classification tasks: distinguishing cancer-derived from healthy eccDNA, and distinguishing authentic eccDNA from length-matched pseudo-circular genomic fragments. The authors report state-of-the-art accuracy, F1, precision, and recall against DNABERT-2, HyenaDNA, Caduceus, and DeepCircle, and claim that the model processes full-length sequences up to 200 kbp without performance degradation. A supplementary motif analysis links the model's cancer predictions to CG-rich zinc-finger transcription factor binding motifs.","tokens_in":13122,"tokens_out":3813,"duration_ms":35917,"significance":"If the empirical claims hold, eccDNAMamba would be a genuinely useful resource: a linear-time bidirectional encoder that can ingest full-length circular eccDNA molecules, trained on a domain-specific corpus and released with code. The paper is transparent about many pretraining details, uses two real biological datasets, and the motif analysis provides a falsifiable biological hypothesis. The main contributions are the bidirectional Mamba-2 encoder adapted for circular DNA and the demonstration of strong classification performance. However, the significance is conditional on resolving three load-bearing concerns: the 200-kbp full-length claim is not supported by the reported token budget; the baseline comparisons are confounded by in-domain pretraining; and the circular augmentation is not validated on the actual model at long sequence lengths.","major_comments":[{"comment":"The claim that eccDNAMamba 'scales to sequences up to 200 Kbp' and 'processes complete sequences end-to-end' is not supported by the reported setup. In §4.5, the comparison fixes eccDNAMamba at 30,000 BPE tokens. At the reported average compression of 5.16 bp/token (§4.1, §A.3), 30,000 tokens correspond to roughly 154,800 bp, so sequences in the upper part of the 10–200 kbp range cannot be fed in full. The manuscript does not report the token counts of the long test sequences or the model's max_length. Please report the actual length distribution and token counts of the test set, or revise the scaling and 'full-length end-to-end' claims accordingly.","section":"§4.5 and Abstract"},{"comment":"The central comparison is confounded by in-domain pretraining. eccDNAMamba is pretrained on ~101.5 million tokens of eccDNA, while DNABERT-2, HyenaDNA, Caduceus, and DeepCircle are used as released, without continued pretraining on the eccDNA corpus. The observed performance gap may therefore reflect domain adaptation rather than the proposed architecture, circular augmentation, or tokenization. The claim that eccDNAMamba 'consistently outperforms' these models needs a control: for example, continue pretraining the baselines on the same eccDNA corpus, or evaluate eccDNAMamba without eccDNA-specific pretraining, and compare under matched conditions.","section":"§4.2, §4.5, §4.6"},{"comment":"The circular augmentation strategy is load-bearing for the 'tailored for circular DNA' claim, but the supporting ablation in Appendix A.1 does not test the actual design. The ablation uses CNN, MLP, and Mamba models on sequences under 200 bp with full-length self-concatenation, rather than the 64-token prefix augmentation applied to the pretrained eccDNAMamba at 10–200 kbp. The differences in Table 3 (e.g., Mamba with versus without augmentation: 0.8048 vs 0.8011 F1) also appear small relative to the reported standard deviations. Please provide an ablation on the real model and long sequences, or explicitly limit the claim that circular augmentation is what drives the long-sequence results.","section":"§3.2.3 and Appendix A.1"},{"comment":"The main results table reports single runs without error bars or significance tests. Several comparisons are close: on the <10 kb cancer task, eccDNAMamba's macro F1 is 0.8242 versus Caduceus's 0.8216, and on precision Caduceus is higher (0.8248 vs 0.8242). The strong conclusion that eccDNAMamba is 'the only model that remains robust across all evaluation metrics' needs multiple seeds with standard deviations, or at minimum a statistical test of the differences, before it can be considered established.","section":"Table 2, §4.5"}],"minor_comments":[{"comment":"The caption says 'Performance comparison of eccDNAMamba and DNABERT-2 under different training settings,' but the table includes HyenaDNA, Caduceus, and DeepCircle; please correct the caption to reflect all compared models.","section":"Table 2 caption"},{"comment":"The DeepCircle zero-shot row is listed with a training set of 20,000 sequences, which is inconsistent with the notion of zero-shot evaluation. Please clarify whether zero-shot means no fine-tuning on this task and, if so, remove the training-set entry or explain how the training set was used.","section":"§4.6 and Table 2"},{"comment":"The citation 'Li et al. (Chang et al., 2023)' is confusingly formatted; it appears to refer to the DeepCircle paper by Chang et al., 2023, but the in-text name and the reference entry should be aligned.","section":"§2.2"},{"comment":"The numbers '28, 100 TP sequences using 29, 292 TN as background' are ambiguous due to spacing; use commas consistently (e.g., 28,100 and 29,292) to avoid misreading.","section":"§5 and Figure 3"},{"comment":"The appendix says the augmentation appends 'approximately 25% of the sequence,' while the main text (§3.2.3) specifies a fixed 64 tokens. Please clarify whether the prefix length is fixed at 64 tokens or scales with sequence length, and reconcile the two descriptions.","section":"Appendix A.2"},{"comment":"There are several typos: 'HeynaDNA' should be 'HyenaDNA', 'eccDNAMamaba' should be 'eccDNAMamba', and 'state of art(SOTA)' should be 'state-of-the-art (SOTA)'.","section":"Throughout"}],"recommendation":"major_revision","confidential_remarks":"The paper is a solid empirical workshop contribution with a useful resource, but the headline claims are broader than the evidence. The 200-kbp full-length claim appears numerically inconsistent with the 30,000-token budget, the baselines are not matched for in-domain pretraining, and the circular augmentation is not ablated on the actual model. These issues are fixable with additional experiments or by carefully narrowing the claims, so I recommend major revision rather than rejection."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: this is a sensible engineering combination—bidirectional Mamba-2 over BPE-tokenized DNA with span masking—and the two downstream tasks are reasonable. But the headline claim that it processes full-length eccDNAs up to 200 kbp is not backed by the reported evaluation, and the circular augmentation that is supposed to make it circular-specific is only validated on toys.\n\nWhat the paper does well: it is transparent about pretraining data (120k sequences, ~100M tokens) and hyperparameters; the BPE compression numbers are useful; and the motif analysis with STREME/Tomtom goes beyond reporting metrics to propose a concrete biological hypothesis. That is more than most foundation-model preprints do.\n\nThe soft spots are real. The stress-test concern holds: in Section 4.5 they compare against Caduceus by feeding 30,000 BPE tokens to eccDNAMamba. At the reported 5.16 bp/token compression, a 200 kbp sequence needs about 38,800 tokens. Either long sequences are truncated—which contradicts the 'full-length end-to-end' description—or the context window is larger but never specified. The paper never reports the actual token counts of the long test sequences or the model's max_length, so the scaling claim is unverified. That is fixable but it is load-bearing.\n\nSecond, the circular augmentation ablation (Appendix A.1) is on CNN/MLP/Mamba models with sequences under 200 bp, using full-length self-concatenation, not on the actual model with the 64-token prefix at 10–200 kbp. So the 'tailored for circular DNA' claim rests on a proxy.\n\nThird, the comparison is confounded by in-domain pretraining: eccDNAMamba is pretrained on eccDNA, while DNABERT-2, HyenaDNA, and Caduceus are used as released. Token-count matching does not fix the data imbalance. The performance gap could come from domain-specific pretraining rather than the architecture.\n\nMinor issues: no error bars on the main results, and the manuscript has a few small inconsistencies that suggest it is unfinished. But none of these are fatal.\n\nWho should read this: anyone working on long-context genomic models or circular DNA. It deserves a serious referee—the idea is worth engaging, and the problems are addressable. A reviewer should ask for token counts and max length, a comparison with baselines pretrained on the same eccDNA corpus, and an augmentation ablation on the real model. I would not cite the 200 kbp claim until that is resolved.","headline":"A reasonable model with an unverified 200-kbp scaling claim and a circular-augmentation ablation that doesn't test the actual design; worth review but not the headline as written.","tokens_in":13635,"tokens_out":4547,"would_cite":false,"duration_ms":40401,"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":"eccDNAMamba, a bidirectional state-space encoder for circular DNA, classifies cancer- and healthy-origin eccDNA and separates authentic eccDNA from pseudo-circular fragments while processing sequences up to 200,000 base pairs.","keywords":["extrachromosomal circular DNA","genomic foundation models","state-space models","Mamba-2","bidirectional encoders","byte-pair encoding","circular DNA augmentation","long-context DNA classification"],"falsifier":"Re-run the reported fine-tuning protocol on the 10–200 kbp cancer-versus-healthy and authentic-versus-pseudo tasks with the 64-token circular augmentation removed or replaced by the same number of zero or random padding tokens, holding all other settings fixed. If accuracy, macro-F1, and recall remain within noise, the circular-specific component contributes nothing and the model reduces to a bidirectional BPE-Mamba; a large drop would confirm that head-tail token dependencies are learnable and load-bearing.","tokens_in":12646,"feed_emoji":"🧬","tokens_out":9175,"duration_ms":82678,"temperature":0.7,"pith_summary":"This paper introduces eccDNAMamba, a pre-trained model built to read extrachromosomal circular DNA (eccDNA) in full, rather than chopping it at an arbitrary breakpoint or truncating to the first few kilobases. The claim is that a bidirectional state-space encoder with byte-pair tokenization and a span-masking pre-training objective captures circular-sequence structure at linear-time cost, and that this pays off in downstream classification. On cancer-versus-healthy eccDNA classification, the model reaches about 0.82 macro-F1 and keeps that level on sequences between 10 and 200 kilobase pairs, where Transformer-based and nucleotide-resolution baselines degrade. On a second task, it separates authentic eccDNAs from length-matched pseudo-circular genomic fragments with about 0.74 macro-F1, beating a domain-specific CNN baseline by roughly seven points. The result is a full-length circular encoder that can serve as a foundation for downstream eccDNA analysis.","feed_headline":"Pre-trained model reads circular DNA up to 200,000 bases","feed_subtitle":"A bidirectional state-space encoder separates cancer-linked and authentic circular DNA without truncating long sequences.","key_machinery":"The central object is a bidirectionally fused Mamba-2 state-space encoder, a linear-time sequence model that updates a hidden state through input-dependent gating rather than quadratic attention. Two copies of the encoder read the same tokenized sequence left-to-right and right-to-left, and a shared MLP merges the aligned hidden states so every token has both upstream and downstream context. Three supporting mechanisms carry the design: byte-pair encoding compresses DNA into motif-like tokens at roughly 5.16 bases per token, circular augmentation appends the first 64 tokens to the sequence end to model head-tail continuity, and a SpanBERT-style span-masking objective trains the model to reconstruct contiguous spans rather than isolated tokens. Padding is suppressed by zeroed embeddings, attention masks, and reset hidden states at padding positions.","core_discovery":"eccDNAMamba is, by the authors' account, the first bidirectional state-space encoder tailored for circular eccDNA. It tokenizes DNA with byte-pair encoding (about 5.16 bases per token), appends the first 64 tokens of a sequence to its end to preserve head-tail circular dependencies, and feeds the sequence through forward and reverse Mamba-2 encoders whose hidden states are aligned and merged by a shared MLP. Pre-training uses a SpanBERT-style span-masking objective on roughly 100 million tokens from human and cross-species eccDNA corpora. Fine-tuned on two classification benchmarks, the model maintains accuracy on sequences up to 200 kilobase pairs while baseline models that truncate inputs or operate at single-nucleotide resolution degrade. A motif analysis of the cancer classifier suggests its decisions turn on CG-rich, C2H2 zinc-finger-like motifs, with AT-rich false negatives pointing to an alternative regulatory logic.","pith_inferences":["If the 64-token circular augmentation is what sustains long-sequence accuracy, a rotationally invariant or periodic positional encoding should match or exceed its effect; comparing those alternatives would separate the circular-wrapping trick from the value of bidirectional context alone.","At roughly 5.16 base pairs per token, a 200-kbp molecule becomes about 38,000 tokens, so a nucleotide-resolution model with the same token budget would need over five times more context; this suggests BPE compression may contribute as much as the architecture to the reported scaling results.","The motif analysis suggests a testable biological extension: cancer eccDNAs enriched for ZNF24 and ZNF263 binding motifs should show higher transcriptional activity or amplification in matched tumor data, checkable with expression or chromatin assays.","Binning the authentic-versus-pseudo task by sequence length (10–50, 50–100, and 100–200 kbp) would reveal whether the model's advantage grows with full-length context, isolating the value of avoiding truncation."],"forward_implications":["If the central claim holds, full-length circular eccDNA molecules up to 200 kilobase pairs can be encoded end-to-end at linear time and memory cost, without truncation or arbitrary breakpoints.","A BPE-tokenized bidirectional state-space encoder can beat nucleotide-resolution and transformer-based genomic foundation models on eccDNA classification despite being pre-trained on far fewer tokens.","The cancer-versus-healthy result implies that cancer-derived and healthy eccDNAs carry learnable sequence-level differences, concentrated in CG-rich zinc-finger-like motifs, rather than being indistinguishable circular fragments.","The authentic-versus-pseudo result implies that real eccDNAs contain non-random sequence features beyond length distribution, since length-matched random circularized genomic fragments are separable at about 0.74 macro-F1.","The model's stable performance on 10-200 kbp inputs where baselines degrade indicates that the practical scaling path for long circular genomes is linear-time state-space encoding paired with motif-level tokenization, not longer attention contexts."],"supporting_citations":[{"why":"Supplies the Mamba state-space backbone that eccDNAMamba builds on for linear-time sequence modeling.","marker":"Gu & Dao, 2023"},{"why":"Provides Mamba-2, the specific base encoder whose bidirectional variant is used in eccDNAMamba.","marker":"Dao & Gu, 2024"},{"why":"Supplies the SpanBERT span-masking objective used in pre-training.","marker":"Joshi et al., 2020"},{"why":"Introduces BPE tokenization for genomes and serves as the DNABERT-2 baseline model.","marker":"Zhou et al., 2023"},{"why":"Establishes the long-context implicit-convolution baseline HyenaDNA that must be outperformed.","marker":"Nguyen et al., 2023"},{"why":"Supplies the bidirectional nucleotide-resolution state-space baseline Caduceus compared at equal token budget.","marker":"Schiff et al., 2024"},{"why":"Supplies the domain-specific CNN classifier DeepCircle that eccDNAMamba must beat on authentic-versus-pseudo classification.","marker":"Chang et al., 2023"},{"why":"Provides CircleBase, one of the two corpora used for pre-training and evaluation.","marker":"Zhao et al., 2022a"},{"why":"Provides eccDNA Atlas, the cross-species corpus used for pre-training.","marker":"Zhong et al., 2023"},{"why":"Supplies STREME, the motif-discovery tool used to interpret what the cancer classifier learned.","marker":"Bailey, 2021"}],"fun_headline_variants":["Bidirectional state-space model reads circular DNA at 200 Kbp","First pre-trained model for full-length circular DNA","eccDNAMamba: linear-time encoder for ultra-long circular genomes","No truncation: model handles 200 Kbp circular DNA"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that appending the first 64 tokens of an eccDNA sequence to its end preserves biologically meaningful head-tail dependencies at the 10–200 kilobase-pair scale; the only ablation offered for this uses full-length self-concatenation on toy models with sequences under 200 base pairs, not the 64-token prefix scheme on the actual pretrained model.","fun_headline_variants_meta":{"raw":{"variants":["Bidirectional state-space model reads circular DNA at 200 Kbp","First pre-trained model for full-length circular DNA","eccDNAMamba: linear-time encoder for ultra-long circular genomes","No truncation: model handles 200 Kbp circular DNA"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001608,"raw_usage":{"total_tokens":6376,"prompt_tokens":888,"completion_tokens":5488,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":504,"completion_tokens_details":{"reasoning_tokens":5417}},"tokens_in":504,"tokens_out":5488,"duration_ms":40316,"temperature":1.0,"reasoning_tokens":5417,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T18:54:37.261825+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Re-run the reported fine-tuning protocol on the 10–200 kbp cancer-versus-healthy and authentic-versus-pseudo tasks with the 64-token circular augmentation removed or replaced by the same number of zero or random padding tokens, holding all other settings fixed. If accuracy, macro-F1, and recall remain within noise, the circular-specific component contributes nothing and the model reduces to a bidirectional BPE-Mamba; a large drop would confirm that head-tail token dependencies are learnable and load-bearing.","supporting_citations":[{"cited_title":"H., Gokaslan, A., Dao, T., Gu, A., and Kuleshov, V","cited_arxiv_id":null,"evidence_quote":"Supplies the bidirectional nucleotide-resolution state-space baseline Caduceus compared at equal token budget."},{"cited_title":"Y., and Tsai, H.-K","cited_arxiv_id":null,"evidence_quote":"Supplies the domain-specific CNN classifier DeepCircle that eccDNAMamba must beat on authentic-versus-pseudo classification."},{"cited_title":"eccdna atlas: a comprehensive resource of eccdna catalog","cited_arxiv_id":null,"evidence_quote":"Provides eccDNA Atlas, the cross-species corpus used for pre-training."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies STREME, the motif-discovery tool used to interpret what the cancer classifier learned."}],"review_version":1}