{"id":"c04c567a-0e47-453f-a7dc-af97f93e3c55","arxiv_id":"2507.11848","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":7,"one_line_summary":"A parametric dual projection method and a visual analytics system, HybridLens, are introduced to help rice breeders identify regulatory genes and select promising hybrids faster.","lead":"This paper presents HybridLens, a visual analytics tool that lets rice breeders explore gene-trait relationships and select hybrid crosses interactively. Its core contribution is a fast parametric dual projection method using invertible neural networks, claimed to beat prior dual projection systems.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Theorem 2 silently assumes perfect autoencoder reconstruction; because Eq. 6 minimizes a weighted sum, ps is not an exact inverse map, so the 'always better than SIRIUS' claim is unproven as stated.","rationale":"The reader's weakest_assumption is exactly the hidden perfect-reconstruction condition in Theorem 2, and my reading of Sec. 5.1.1 confirms that this is the most load-bearing weakness. The quantitative projection results (Tab. 1) and running-time measurements support the practical utility of the method, and the source code is promised, but none of that validates the formal 'always better than SIRIUS' statement. Since the stated inequality depends on ps being an exact inverse and the actual architecture only provides an approximate inverse through the autoencoder, the theorem is not derivable from the stated assumptions without additional conditions. The concrete test above would settle the matter: if inverse consistency is exact in practice on the target data, the concern is minor and the theorem can be re-stated with a reconstruction-error condition; if not, the theoretical contribution must be weakened to an approximation guarantee. I keep the reader's CONDITIONAL verdict unchanged because the concern does not shift the verdict, but I would escalate to REJECT if the test shows the inequality failing for realistic edits and the authors do not supply a proof under weaker assumptions. The paper's own limitations sections are candid about the field-simulation issue and incremental-update limitation, so those are not grounds for a stronger verdict.","tokens_in":20199,"tokens_out":6025,"duration_ms":86628,"concrete_test":"Independently re-derive Theorem 2 from the stated assumptions without assuming l_recon=0; if the proof uses Encoder(Decoder(z))=z or ps^{-1}∘ps=id, the theorem is conditional on perfect reconstruction. Empirically, after training exactly as in Sec. 6.1 on the rice genomic dataset, compute the average reconstruction error of Eq. 8 and generate 100 simulated user edits S' (moving/contracting points as in the case study). For each edit, compute X'_inv=Decoder(f_inv^{-1}(S')), then compare Delta_inv=||ps(X'_inv)-S'||^2 with Delta_SIRIUS obtained by running SIRIUS on the same edit. If any Delta_inv exceeds Delta_SIRIUS, or if Delta_inv does not approach zero as lambda is decreased while projection quality is held fixed, Theorem 2 is false as stated.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central theoretical claim is Theorem 2: when one scatterplot is modified to S', the updated hybrids computed as X'_inv = Decoder(f_inv^{-1}(S')) satisfy ||ps(X'_inv)-S'||^2 <= ||ps(X'_SIRIUS)-S'||^2, i.e., the method 'can always be better' than SIRIUS. This is only a theorem if ps is an exact inverse of the construction used to recover X'. But ps(X)=f_inv(Encoder(X)) while X'_inv = Decoder(f_inv^{-1}(S')), so ps(X'_inv) = f_inv(Encoder(Decoder(f_inv^{-1}(S')))). Exact equality with S' requires Encoder(Decoder(z))=z for the relevant latent z, i.e., zero reconstruction loss in Eq. 8. The actual training objective, Eq. 6, is L = l_proj + lambda*l_recon with lambda set by multi-task weighting; nothing forces l_recon to vanish. With imperfect reconstruction, X'_inv is only an approximate solution to Eq. 3, and no argument is given that it is closer to S' than SIRIUS's update is. The issue is compounded by the fact that ps maps the original high-dimensional X to a 2D scatterplot, so ps cannot be injective/bijective as a map from X-space; p^{-1} is only a decoder-based approximation. The proof is deferred to a missing supplemental document, so the hidden assumption cannot currently be checked. This is a load-bearing gap because Theorem 2 is the paper's advertised formal guarantee and the basis for claiming superiority over the existing state of the art. A secondary concern, the field-experiment simulation that treats all uncultivated hybrids as undesired, is explicitly acknowledged by the authors and is not the main theoretical weakness.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a visual analytics system, HybridLens, for interactive hybrid rice breeding. The core technical contribution is a parametric dual projection method that uses an invertible neural network nested inside an autoencoder to project both hybrids and genes into two linked scatterplots, allowing a user's modification of one scatterplot to update the other in about 0.01 seconds. The authors claim two theorems: Theorem 1 states that the optimal updated data are obtained by a single backward pass through the inverse projection, and Theorem 2 states that the proposed update is always better than the existing SIRIUS method in terms of squared distance to the user-modified projection. The system also includes a hybrid recommendation view and a gene visualization, and the paper reports a quantitative comparison with t-SNE, PCA, and SIRIUS on MNIST, CIFAR-10, and rice genomic data, as well as a case study with rice breeders and a simulated field experiment.","tokens_in":20588,"tokens_out":6618,"duration_ms":76039,"significance":"If the theoretical and empirical claims are valid, the paper makes a valuable contribution to interactive dual projection and to visual analytics for genomic breeding. The system addresses a real and well-motivated problem, the case study with breeders is informative, and the source code is promised. The interactive speed of the proposed dual projection is a clear improvement over SIRIUS, and the approach is domain-agnostic. However, the advertised theoretical guarantee 'always better than SIRIUS' is currently not supported by a verifiable proof in the manuscript and appears to rely on an unstated perfect-reconstruction assumption. The quantitative evaluation is also weakened by single-run results and by metrics that measure the same neighborhood-preservation objective used in training. These issues are central to the paper's claimed contributions, but they are in principle fixable, so the paper merits revision rather than rejection.","major_comments":[{"comment":"Theorem 2 does not follow from the stated training objective. Since ps(X'_inv) = finv(Encoder(Decoder(finv^{-1}(S')))), the equality ps(X'_inv)=S' holds only if Encoder(Decoder(z))=z for z=finv^{-1}(S'), i.e., the autoencoder reconstruction loss in Eq. (8) is zero. The actual objective in Eq. (6) is a weighted sum L = ℓ_proj + λ ℓ_recon, and nothing forces ℓ_recon to vanish. As stated, the theorem is therefore unproven, and the inequality over SIRIUS is a conditional statement at best. The proof is deferred to a supplemental document that is not included, so the hidden assumption cannot be checked. Please either state and prove the theorem under an explicit perfect-reconstruction assumption, or derive a bound of the form ||ps(X'_inv)-S'||^2 ≤ B(ℓ_recon) that depends on the reconstruction error and show that this bound is smaller than the corresponding SIRIUS error under the settings of Table 1.","section":"Sec. 5.1.1, Eq. (6)-(8)"},{"comment":"Theorem 1 claims that when an invertible neural network is used as the projection function ps, the optimal updated data satisfy X'_opt = p^{-1}_s(S'). However, the architecture defines ps as ps(X)=finv(Encoder(X)), and the inverse is defined as p^{-1}_s(S')=Decoder(finv^{-1}(S')). Because the encoder maps the high-dimensional input to a low-dimensional latent space, ps is not invertible as a function on the input space; the decoder-based quantity is an approximate inverse only if the autoencoder is near-perfect. The statement of Theorem 1 is thus inconsistent with the actual architecture unless an explicit autoencoder-invertibility condition is added. Please clarify the exact invertibility assumptions and reconcile them with the dimensionality reduction performed by the encoder, or revise the theorem statement accordingly.","section":"Sec. 5.1.1, Theorem 1 and Fig. 5"},{"comment":"The quantitative comparison reports single values for trustworthiness T(30), continuity C(30), and Time, with no standard deviations, no repeated runs, and no significance tests. Since the proposed method is trained with stochastic optimization and the contrastive loss in Eq. (7), the reported margins (e.g., T(30) of 0.961 vs 0.781 on MNIST) cannot be assessed for statistical reliability. Additionally, trustworthiness and continuity measure the same kNN neighborhood-preservation objective that the contrastive loss ℓ_proj optimizes, so the comparison is partly self-confirming. An independent evaluation, such as a downstream classification task or a baseline trained without the contrastive objective, is needed to support the claim that the proposed projection better preserves neighborhoods.","section":"Table 1 and Sec. 6.1"},{"comment":"The simulated field experiment labels every hybrid that has not been cultivated in the private dataset as undesired. The manuscript explicitly acknowledges that this assumption is 'not highly rigorous' and that some uncultivated hybrids may actually be desirable. Under this labeling, the reported saving of 102 field cultivations is not a valid measure of workload reduction; it conflates 'not previously cultivated' with 'undesired.' The claim that HybridLens 'largely reduces the workload' of field cultivation is therefore overstated. Please re-analyze the simulations using only the subset of hybrids with known cultivated traits, or clearly label the result as anecdotal rather than as quantitative evidence.","section":"Sec. 6.2.3"},{"comment":"The proofs of Theorems 1 and 2 are deferred to a supplemental document that is not included in the manuscript. Because these theorems are the central advertised theoretical guarantees and their validity is in question for the reasons described above, the main text should contain at least a proof sketch with all assumptions explicitly stated. Without this, the reader cannot verify the claimed theoretical superiority over SIRIUS.","section":"Sec. 5.1.1, Theorems 1 and 2"}],"minor_comments":[{"comment":"The formula for W contains stray 'q' symbols and is difficult to parse; please rewrite the expression with clear notation for the distance terms.","section":"Eq. (5)"},{"comment":"The notation p^{-1}_s is used to define X' before the inverse map is formally introduced; please define p^{-1}_s as Decoder(finv^{-1}(·)) before its first use.","section":"Sec. 5.1.1"},{"comment":"Several subfigures are referenced as A-F in the text but are not clearly labeled in the figure; adding explicit labels would improve readability.","section":"Fig. 6"},{"comment":"The paper uses 'gene' to refer to a Single Nucleotide Polymorphism; although the term is defined, the unconventional usage is potentially confusing and could be highlighted at first use.","section":"Sec. 3"},{"comment":"The text quotes an update time of approximately 17 seconds for a dataset of 714 hybrids and 2,081 genes, which is consistent with the genomic Time value in Table 1, but the quotation and table should be cross-referenced explicitly to avoid ambiguity.","section":"Sec. 6.1"}],"recommendation":"major_revision","confidential_remarks":"The 'always better than SIRIUS' theorem is the main hook of the paper, but the missing proofs and the unstated perfect-reconstruction assumption make the theoretical claim overreaching as currently written. The single-run evaluation and the circularity of the projection metrics further weaken the empirical support. The visualization system itself appears well designed, the breeder feedback is encouraging, and the interactive runtime is a genuine improvement. If the authors can supply rigorous proofs with explicit assumptions and improve the evaluation, the paper could be publishable; otherwise, the theoretical claims should be removed or substantially qualified."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The interesting new thing here is the parametric dual projection built from invertible networks inside an autoencoder, and the rice-breeding application gives it a real testbed. But the advertised theoretical guarantee—that the method 'can always be better' than SIRIUS—does not follow from the stated equations. Theorem 2 requires p_s(X'_inv) to land on S', which needs Encoder(Decoder(z)) = z for the latent z you start from. The training loss only penalizes reconstruction in the original space; nothing forces the latent-space reconstruction to be exact, and p_s maps from high-dimensional space into 2D, so an exact inverse isn't even well-defined. The proof is deferred to a missing supplemental, so you can't check whether some extra assumption saves it. The same issue hangs over Theorem 1. That is load-bearing because the 'better than SIRIUS' claim is the paper's headline. What the paper does well: the system is well-scoped, developed with four breeders over 14 months, the case study is detailed and honest, and the limitation statements (simulated field trial, pair analytics, environmental factors) are right there in Sec. 6.2.3 and 7.2. The measured 0.01s update time versus SIRIUS's 17s is a real practical gain. On the projection evaluation, the method beats SIRIUS on trustworthiness and continuity, though those metrics measure the same neighborhood objective used in the contrastive loss, so the comparison is partially self-confirming. Table 1 has single runs with no variance, which is a minor weakness. The 126-to-24 field trial reduction rests on a simulation where uncultivated hybrids are counted as undesired; the authors acknowledge this, so it's a stated limitation, not a hidden one. The citation pattern is standard; WMDS, SIRIUS, CDR, and genomic prediction works are all discussed. Bottom line: this deserves a serious referee. The visual analytics side is solid, the application is meaningful, and the parametric dual projection idea is worth developing. But the theorem needs to be either repaired with explicit conditions or reframed as an empirical claim, and the proofs must be in the supplemental before the formal headline is taken at face value. A revision is the right next step, not a reject.","headline":"The parametric dual projection is a real new combination and the case study is solid, but the 'always better than SIRIUS' theorem silently assumes exact autoencoder reconstruction that the training loss doesn't enforce.","tokens_in":21144,"tokens_out":3385,"would_cite":true,"duration_ms":40183,"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":"A parametric dual projection method, built on invertible neural networks inside an autoencoder, makes interactive dual analysis fast enough for hybrid rice breeding while preserving neighborhoods better than the previous state-of-the-art…","keywords":["hybrid rice breeding","dual projection","parametric projection","invertible neural networks","genomic prediction","visual analytics","regulatory gene identification","neighborhood preservation"],"falsifier":"Measure the reconstruction error of the trained autoencoder on the rice dataset; if $\\mathrm{Decoder}(f_{\\mathrm{inv}}^{-1}(S'))$ is not close to $X'$ for typical edited scatterplots, then the inequality in Theorem 2 is not operational. Then, on many random edits, compute both $p_s(X'_{\\mathrm{inv}})$ and $p_s(X'_{\\mathrm{SIRIUS}})$ and check whether the claimed distance-to-$S'$ dominance actually holds in every case.","tokens_in":19952,"feed_emoji":"🌾","tokens_out":4410,"duration_ms":50993,"temperature":0.7,"pith_summary":"The paper is trying to establish that the two core tasks in hybrid rice breeding—identifying regulatory genes and selecting desired hybrids—are one dual-analysis problem, and that a parametric dual projection built from invertible neural networks solves it fast enough for interaction and faithfully enough to beat the prior state of the art. If true, breeders can combine genomic prediction with their own judgment, see the gene-level consequences of moving hybrids around a scatterplot in real time, and reduce the number of field cultivations needed. The payoff is concrete: in the paper's simulated field experiment, 24 hybrids instead of 126 needed cultivation to find a desired high-yield, high length/width-ratio hybrid.","feed_headline":"Interactive breeding tool cuts field trials from 126 hybrids to 24","feed_subtitle":"A dual scatterplot lets breeders edit hybrids and see gene weights respond, cutting cultivation from 126 to 24.","key_machinery":"The central object is a parametric dual projection network: an invertible neural network $f_{\\mathrm{inv}}$ nested inside an autoencoder, with the hybrid projection given by $S = f_{\\mathrm{inv}}(\\mathrm{Encoder}(X))$ and the gene projection obtained analogously from $X^T$. Because $f_{\\mathrm{inv}}$ is invertible, an edited scatterplot $S'$ is mapped back to data by one backward pass, $X' = \\mathrm{Decoder}(f_{\\mathrm{inv}}^{-1}(S'))$, which replaces SIRIUS's iterative stress-majorization optimization and makes updates take about 0.01 seconds. The training objective is a weighted sum of a contrastive loss that preserves neighborhoods and a reconstruction loss that lets the inverse exist; the paper's Theorem 2 claims this scheme always beats SIRIUS in matching the edited scatterplot.","core_discovery":"The paper claims that interactive dual analysis can be made parametric and near-instantaneous by using invertible neural networks as the projection functions, and that this is not only faster but also better at preserving neighborhoods than the existing state-of-the-art method SIRIUS. Applied to 714 Indica rice hybrids, the HybridLens tool lets a breeder move hybrids in one scatterplot, immediately see which genes' weights rise in the other, verify candidate regulatory genes against chromosome context and genotype-trait boxplots, and select hybrids whose genotypes match known regulators. In the case study the method rediscovers the known length/width-ratio gene GW7, points to a previously unreported candidate region on chromosome 4, and explains why GS3 appears ineffective due to extreme allele imbalance. The central theoretical assertion, stated as Theorem 2, is that when one scatterplot is modified, the updated other scatterplot produced by this method is always at least as close to the user's edit as the one produced by SIRIUS.","pith_inferences":["Because the projection is parametric, the same dual-analysis machinery should transfer to other paired datasets, such as protein features and organism traits or patient genotypes and drug responses, with only the visual encodings redesigned.","The theoretical guarantee is only as strong as the reconstruction term in the training loss; a natural next experiment is to measure how the claimed dominance over SIRIUS degrades as the reconstruction error grows or as the balancing weight $\\lambda$ changes.","The closed-form gene-weight update $W = \\mathrm{diag}(X^T X')/\\mathrm{diag}(X^T X)$ gives an interpretable per-gene importance score that could be validated against GWAS hits beyond the single case study.","The 126-versus-24 comparison treats uncultivated hybrids as undesired, so a true forward validation would require growing the 24 recommended hybrids to maturity and measuring their traits directly."],"forward_implications":["Dual projection updates complete in about 0.01 seconds versus roughly 17 seconds for SIRIUS on the rice dataset, making interactive exploration feasible on genomic-scale data.","The method achieves higher trustworthiness and continuity than SIRIUS on MNIST, CIFAR-10, and the rice genomic dataset, meaning the scatterplot neighborhoods better reflect true high-dimensional relationships.","Breeders can identify regulatory genes by interacting with hybrids, rediscovering the known gene GW7 and surfacing a candidate region on chromosome 4 for the target rice population.","Combining the hybrid recommendation with dual analysis selected 24 hybrids that met the desired traits, versus 126 hybrids for automatic genomic selection alone, reducing the field-cultivation workload.","The parametric projection is domain-agnostic, so the same dual-analysis machinery can be applied to other paired high-dimensional datasets once the visual encodings are redesigned."],"supporting_citations":[{"why":"Supplies the SIRIUS baseline, the prior state-of-the-art dual projection method that the new method is compared against and that Theorem 2 claims to dominate.","marker":"[12]"},{"why":"Provides the weighting-based dual projection formulation and stress-majorization optimization that SIRIUS inherits, motivating the need for a faster parametric alternative.","marker":"[14]"},{"why":"Supplies the invertible neural network construction (Glow) that makes the inverse mapping a single backward pass, the core mechanism of the proposed method.","marker":"[25]"},{"why":"Shows that contrastive dimensionality reduction achieves state-of-the-art projection quality and supplies the contrastive loss used for neighborhood preservation.","marker":"[60]"},{"why":"Provides the DNNGP genomic prediction model used to predict hybrid traits from genotype data, which the tool integrates with breeder interaction.","marker":"[54]"},{"why":"Identifies the reported regulatory genes GS3 and GW7 and the GWAS correlation method used to filter genes and validate the case-study findings.","marker":"[21]"}],"fun_headline_variants":["Breeder tool cuts hybrid rice trials from 126 to 24","Drag hybrids, watch gene weights shift—trials drop to 24","Parametric dual projection enables instant rice breeding edits","Rediscovers gene GW7 and cuts rice field tests to 24","Interactive dual scatterplot reduces rice trials to 24"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The theorem that the new method always beats SIRIUS assumes the autoencoder reconstructs the original gene data almost perfectly, but the training loss only balances reconstruction against projection quality, so perfect reconstruction is not guaranteed.","fun_headline_variants_meta":{"raw":{"variants":["Breeder tool cuts hybrid rice trials from 126 to 24","Drag hybrids, watch gene weights shift—trials drop to 24","Parametric dual projection enables instant rice breeding edits","Rediscovers gene GW7 and cuts rice field tests to 24","Interactive dual scatterplot reduces rice trials to 24"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000602,"raw_usage":{"total_tokens":2810,"prompt_tokens":946,"completion_tokens":1864,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":562,"completion_tokens_details":{"reasoning_tokens":1777}},"tokens_in":562,"tokens_out":1864,"duration_ms":19171,"temperature":1.0,"reasoning_tokens":1777,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T17:00:37.051892+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Measure the reconstruction error of the trained autoencoder on the rice dataset; if $\\mathrm{Decoder}(f_{\\mathrm{inv}}^{-1}(S'))$ is not close to $X'$ for typical edited scatterplots, then the inequality in Theorem 2 is not operational. Then, on many random edits, compute both $p_s(X'_{\\mathrm{inv}})$ and $p_s(X'_{\\mathrm{SIRIUS}})$ and check whether the claimed distance-to-$S'$ dominance actually holds in every case.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the invertible neural network construction (Glow) that makes the inverse mapping a single backward pass, the core mechanism of the proposed method."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the DNNGP genomic prediction model used to predict hybrid traits from genotype data, which the tool integrates with breeder interaction."},{"cited_title":"Huang, X","cited_arxiv_id":null,"evidence_quote":"Identifies the reported regulatory genes GS3 and GW7 and the GWAS correlation method used to filter genes and validate the case-study findings."}],"review_version":1}