{"id":"4dd6af74-bdf6-4a7e-9688-235ed628bf81","arxiv_id":"2607.06224","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"unknown","formal_verification":"none","parameter_count":18,"one_line_summary":"Frozen embeddings from a pretrained heterogeneous graph transformer over a 6.9M-node metabolic-engineering knowledge graph predict fermentation titers at R²=0.41, outperforming tabular baselines (R²=0.24).","lead":"Canopy builds a 6.9M-node knowledge graph for metabolic engineering and pretrains a heterogeneous graph transformer on it, achieving R²=0.41 on fermentation titer prediction vs. 0.24 for tabular baselines. A smart generalist might read it because it tests whether graph-structured biological knowledge improves strain-design predictions over flat-feature ML.","discovery_kind":"unclear","skeptic_critique":{"model":"glm-5.2","headline":"The missing graph-free encoder control is the load-bearing gap: R²=0.41 vs 0.24 conflates graph structure with foundation-model features (ESM-2, MoLFormer) that tabular baselines never receive.","rationale":"The reader correctly identifies the most load-bearing concern: the absence of a graph-free pooled-encoder control. This is the single experiment that would directly test whether the heterogeneous graph structure — as opposed to the frozen foundation-model feature encoders (ESM-2, MoLFormer, PubMedBERT) — drives the R² improvement from 0.24 to 0.41. The authors themselves flag this in their limitations (Section 5, third point), which strengthens the case that this is the critical missing piece rather than an oversight the reader invented. The paper does provide partial evidence that graph architecture matters: Canopy (3B, R²=0.413) outperforms vanilla HGT (R²=0.308) and GraphSAGE (R²=0.241), all sharing the same node features. But these comparisons only show that the specific graph architecture and augmentations matter among graph-based methods; they do not establish that graph structure per se is necessary, because no method in Table 3 operates without message passing. The tabular baselines are not a fair control because they lack the foundation-model features entirely. I also note two secondary issues that compound the main concern: (1) No error bars are reported across the 5 CV folds. With n=410 test experiments, fold-to-fold variance could be substantial, and the gap between Canopy (0.41) and the best graph baseline (GraphSAGE+SN/JK/VN, 0.334) is 0.076 R² — potentially within noise. (2) The depth ablation (Table 6) shows the shallow L=2 model (R²=0.370) outperforms the headline L=6 configuration (R²=0.359) at 500M scale, yet the deeper config is retained for headline results. The authors justify this by noting the gap is small (ΔR²=0.011) and the deeper topology matches the 3B configuration, but this means the headline architecture choice is not the one that performs best in controlled comparison. These issues do not change the verdict from CONDITIONAL — the reader's assessment is accurate. The paper is well-engineered, the KG construction is thorough, and the ablations (Tables 4–7) are informative. But the central attribution claim rests on a comparison that is confounded by feature access, and the one experiment that would resolve this is explicitly acknowledged as not yet run. CONDITIONAL is the right call: the work is promising but the core claim is not yet sufficiently isolated.","tokens_in":19057,"tokens_out":2588,"duration_ms":206223,"concrete_test":"Construct the graph-free control the authors describe: for each Experiment node, concatenate (1) mean-pooled frozen ESM-2 embeddings of all genes with modification edges (knockout/knockin/overexpression) to that experiment's strain, (2) frozen MoLFormer embedding of the target compound, (3) the 429-dim raw condition vector. Feed this through the same two-layer MLP probe used for Canopy, on the identical MD5-hashed 5x CV split. Report mean R² ± std across folds. If the graph-free probe reaches R² ≥ 0.35, the graph structure's marginal contribution drops below 0.06 R² and the central attribution claim weakens. If it stays near the tabular baselines (R² ≈ 0.24), the graph structure claim is strongly supported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that the heterogeneous graph transformer over the KG produces representations encoding useful cross-organism metabolic structure. The evidence is R²=0.41 (Canopy 3B) vs R²=0.24 (best tabular baseline, XGBoost). But the tabular baselines operate on a 429-dim raw experiment-condition vector, while Canopy's Experiment node aggregates ESM-2 protein embeddings, MoLFormer chemical embeddings, and PubMedBERT text features from neighboring graph nodes. The comparison therefore conflates two factors: (a) graph structure and (b) access to pretrained foundation-model features. The graph-internal comparisons (Canopy vs vanilla HGT vs GraphSAGE, Table 3) partially address this since all use the same node features, but none isolates the graph-free case. The authors acknowledge this in Section 5, Limitations, third point: 'our baselines do not include a graph-free pooled-encoder control, e.g., concatenating frozen ESM-2 and MoLFormer embeddings of a strain's genes and target compound and applying an MLP.' If such a probe achieves R² near 0.41, the claim that graph structure drives the improvement would be substantially weakened — the improvement would be attributable to the feature encoders, not the heterogeneous graph transformer. This is the single most load-bearing gap because it directly tests the paper's core attribution claim. A secondary concern: no error bars are reported across the 5 CV folds (n=410 test experiments), so the statistical significance of R²=0.41 vs 0.24 is unknown.","agreement_with_reader":"agree"},"referee_report":{"model":"glm-5.2","summary":"The paper presents Canopy, a heterogeneous graph foundation model for metabolic engineering. The authors construct a knowledge graph of 6.9M nodes (13 types) and 34 edge types from ten data sources, encode node features using frozen pretrained models (ESM-2, MoLFormer, PubMedBERT), and pretrain a Heterogeneous Graph Transformer (HGT) with four self-supervised objectives (link prediction, masked node modeling, distance prediction, contrastive experiment clustering) combined via learned uncertainty weighting. On fermentation titer prediction using 4,791 experiments with a deterministic 5-fold CV split, frozen Canopy embeddings achieve R²=0.41 with a lightweight probe, outperforming tabular baselines (best R²=0.24) and homogeneous GNN variants. Ablations test pretraining objectives, architectural components, depth, and loss weighting.","tokens_in":19381,"tokens_out":1544,"duration_ms":267116,"significance":"The integration of a multi-modal metabolic-engineering knowledge graph at this scale is a genuine contribution to an underserved domain. The system design—combining BioCypher-based KG construction, schema-driven multi-modal feature dispatch, HGT with SignNet positional encodings and Jumping Knowledge, and multi-task self-supervised pretraining—is well-executed and technically sound. The deterministic MD5-hashed split protocol and the hold-out integrity design (excluding held-out Experiment nodes from all pretraining supervision while retaining them in the message-passing graph) are commendable. The ablation suite (Tables 4–7) is thorough in testing individual components. The authors are transparent about limitations, including the missing graph-free encoder control and data sparsity.","major_comments":[{"comment":"§4.2, Table 3: The central claim that graph structure drives the R²=0.41 improvement rests on the comparison between Canopy and tabular baselines (R²=0.24). However, the tabular baselines operate on a 429-dim raw experiment-condition vector, while Canopy's Experiment node aggregates ESM-2, MoLFormer, and PubMedBERT features from neighboring graph nodes. This conflates two factors: (a) the heterogeneous graph transformer and (b) access to pretrained foundation-model features. The authors acknowledge this in §5 (Limitations, third point) as a planned ablation. Without a graph-free control that concatenates frozen ESM-2 and MoLFormer embeddings of a strain's genes and target compound and applies an MLP probe, the attribution of improvement to graph structure rather than to the feature encoders cannot be established. This is load-bearing for the paper's core claim. The graph-internal basings","section":null},{"comment":"§4.2, Table 3: No error bars or standard deviations are reported across the 5 CV folds (n=410 test experiments). The R²=0.41 vs. 0.24 comparison lacks any statistical significance assessment. Given the modest absolute R² values and the relatively small test set, reporting per-fold variance and a significance test would strengthen the claim that the improvement is robust rather than an artifact of split variance.","section":null},{"comment":"§5, Limitations, fourth point: The 4,791-experiment benchmark and split files are stated to be released 'in a forthcoming publication,' and trained model weights and LIMS records are not released. For a paper claiming a foundation model and a benchmark, the inability to reproduce any of the headline results—no code, no data, no weights—is a significant gap. At minimum, the benchmark split files and probe evaluation code should be released to allow independent verification of the R²=0.41 claim.","section":null}],"minor_comments":[{"comment":"Table 4: The '∆pp' column header is ambiguous—it appears to mean percentage-point change, but this should be stated explicitly. The row showing '✓ ✓ ✓ ✓' with R²=0.359 and no ∆pp value is the reference; labeling it as such would improve clarity.","section":null},{"comment":"Table 7: The flat-weighting row reports '∆R² = −9.18', which appears to be in percentage points, but the learned-weighting row reports R²=0.359 with no ∆. Consistent units (decimal vs. percentage points) across tables would aid comparison.","section":null},{"comment":"§3.4, contrastive experiment-pair loss: The target similarity s_ij = 1 − d_ij/d_max is defined, but d_max is not specified. Is it the maximum graph distance within the batch, the subgraph, or the full graph? This should be clarified.","section":null},{"comment":"§3.5: The multi-anchor sampling strategy rebalances GenomicGene fraction 'from 78% to 15%'. It would help to state what the Experiment-node coverage actually is (the '13×' relative figure is given but the absolute percentage is not).","section":null},{"comment":"§4.3, Table 6: The depth ablation compares L=2 (h=224) vs. L=6 (h=128) at 500M scale. The authors note the shallow model performs better (R²=0.370 vs. 0.359) but retain the deeper default. The justification ('the gap is small and the deeper topology matches the 3B configuration') is reasonable but could note whether the 3B model itself was tested at L=2.","section":null},{"comment":"Figure references: Figures 2–5 are referenced in §3.3 but the figure captions themselves do not include figure numbers in the text provided, making cross-referencing difficult. Ensure figures are numbered consistently.","section":null},{"comment":"§3.3: The virtual node type is described as 'added with bidirectional edges to all other nodes.' Given the subgraph sampling approach (§3.5), it would be useful to clarify whether virtual node edges are added within each sampled subgraph or at the full-graph level before sampling.","section":null},{"comment":"Table 1: The total node count is listed as 6,863,526, but the abstract states 6.9M. Minor rounding inconsistency—consider aligning.","section":null},{"comment":"§4.1: The hyperparameter sweep used ~1,300 XPU-hours total (500 + 800). It would be useful to state the total compute including the final training runs at each scale.","section":null}],"recommendation":"major_revision","confidential_remarks":"The missing graph-free encoder control is the most substantive issue. The authors themselves acknowledge it as a planned ablation, which suggests they recognize its importance. I would not recommend rejection because the system design and KG construction are solid contributions, and the graph-internal comparisons (Canopy vs. vanilla HGT vs. GraphSAGE, all using the same node features) partially address the attribution question. However, the graph-free control is essential for the central claim and should be required before acceptance. The lack of any released code or data is also concerning for reproducibility, though this may be partly explained by the workshop-to-journal trajectory the authors describe."},"author_rebuttal":{"model":"glm-5.2","summary":"We thank the referee for a careful and constructive report. The referee correctly identifies that our central claim—graph structure drives the R²=0.41 improvement—requires a graph-free foundation-model control to disentangle the contribution of the heterogeneous graph transformer from the contribution of the frozen ESM-2/MoLFormer/PubMedBERT feature encoders. We agree this is load-bearing and will add the control. We also agree on error bars and will release benchmark split files and probe evaluation code. One genuine constraint remains: trained model weights and proprietary LIMS records cannot be released.","responses":[{"response":"The referee is correct that the current Table 3 comparison conflates two factors: the heterogeneous graph transformer and access to pretrained foundation-model features. We agree this is load-bearing for the core claim and will add the requested graph-free control in the revision. Specifically, we will implement an MLP probe on the concatenation of frozen ESM-2 mean-pooled embeddings (for each strain's gene set), frozen MoLFormer embeddings (for the target compound SMILES), and frozen PubMedBERT embeddings (for experiment text), using the same MD5-hashed 5-fold CV split. This control isolates whether the graph structure and message passing contribute beyond what the foundation-model encoders alone provide. We note that the existing graph-internal ablations in Table 3 (GraphSAGE vanilla at R²=0.241, HGT vanilla at R²=0.308, HGT+SN/JK/VN at R²=0.413) already provide partial evidence that graph architecture matters—these baselines share the same frozen encoder features but differ in backbone, and the improvement from 0.241 to 0.413 tracks architectural choices. However, we concede that none of these baselines controls for the foundation-model encoders in the way the referee describes, and the graph-free encoder control is the cleaner and more direct test. We will add it as a new row in Table 3 and revise the claims in §4.2 and the Abstract accordingly.","revision_made":"yes","referee_comment":"§4.2, Table 3: The central claim that graph structure drives the R²=0.41 improvement rests on the comparison between Canopy and tabular baselines (R²=0.24). However, the tabular baselines operate on a 429-dim raw experiment-condition vector, while Canopy's Experiment node aggregates ESM-2, MoLFormer, and PubMedBERT features from neighboring graph nodes. This conflates two factors: (a) the heterogeneous graph transformer and (b) access to pretrained foundation-model features. The authors acknowledge this in §5 (Limitations, third point) as a planned ablation. Without a graph-free control that concatenates frozen ESM-2 and MoLFormer embeddings of a strain's genes and target compound and applies an MLP probe, the attribution of improvement to graph structure rather than to the feature encoders cannot be established. This is load-bearing for the paper's core claim."},{"response":"We agree. We will report per-fold R² values, standard deviations across the 5 folds, and a paired statistical test (Wilcoxon signed-rank on per-fold R² values, given n=5 folds) for all methods in Table 3 in the revised manuscript. We acknowledge that with only 5 folds the statistical power of any significance test is limited, but reporting the variance is important for the reader to assess robustness. We will also add a note discussing the limitations of inference at this sample size.","revision_made":"yes","referee_comment":"§4.2, Table 3: No error bars or standard deviations are reported across the 5 CV folds (n=410 test experiments). The R²=0.41 vs. 0.24 comparison lacks any statistical significance assessment. Given the modest absolute R² values and the relatively small test set, reporting per-fold variance and a significance test would strengthen the claim that the improvement is robust rather than an artifact of split variance."},{"response":"We agree that the release scope described in the current manuscript is insufficient. In the revision we will release (1) the 4,791-experiment benchmark metadata with the deterministic MD5-hashed 5-fold CV split files, and (2) the probe evaluation code (linear and MLP probe training, R²/RMSE/Spearman computation) as a public repository. This will allow independent verification of the R²=0.41 claim given the released split and probe code. We will also release the subgraph sampling configuration and the schema YAML so that the knowledge graph construction pipeline is reproducible from public data sources. The trained model weights and proprietary LIMS records will remain constrained—the LIMS data is proprietary to our company and cannot be released, and the model weights encode learned representations over that proprietary data. We will state this constraint explicitly and adjust the 'foundation model' framing to be transparent about what is and is not reproducible from public components alone.","revision_made":"partial","referee_comment":"§5, Limitations, fourth point: The 4,791-experiment benchmark and split files are stated to be released 'in a forthcoming publication,' and trained model weights and LIMS records are not released. For a paper claiming a foundation model and a benchmark, the inability to reproduce any of the headline results—no code, no data, no weights—is a significant gap. At minimum, the benchmark split files and probe evaluation code should be released to allow independent verification of the R²=0.41 claim."}],"tokens_in":18982,"tokens_out":1238,"duration_ms":139023,"standing_objections":["Trained model weights cannot be released because they encode learned representations over proprietary LIMS fermentation records. This is a genuine constraint of the industrial setting and cannot be resolved for this submission.","Proprietary LIMS DBTL records (a subset of the 4,791 experiments) cannot be released. The literature-mined subset and split files will be released, but the in-house experimental records will remain constrained."]},"desk_editor":{"model":"glm-5.2","letter":"The headline: Canopy builds a genuinely new resource — a 6.9M-node, 13-type, 34-edge-type cross-organism knowledge graph for metabolic engineering, with a multi-modal HGT pretrained via four self-supervised objectives. The titer prediction result (R²=0.41 vs 0.24 for the best tabular baseline) is a real improvement. But the paper's central claim — that graph structure drives this improvement — is not yet isolated from the contribution of the frozen foundation-model encoders (ESM-2, MoLFormer, PubMedBERT) that the tabular baselines never see. The authors acknowledge this gap themselves in Section 5. That is the one thing you need to know before reading anything else. What is actually new and well done: the KG construction via BioCypher with ten adapters is solid engineering. The multi-anchor subgraph sampling strategy is thoughtful — the rebalancing of GenomicGene fraction from 78% to 15% and the 13× Experiment-node coverage increase over naive BFS shows they understood the graph-skew problem. The ablations are reasonably thorough: pretraining objectives (Table 4), architectural components (Table 5), depth (Table 6), and loss weighting (Table 7) are all tested. The learned uncertainty weighting matters — flat weighting collapses the probe, so the multi-task design is doing real work. The virtual node ablation showing a 10.3pp drop is the strongest single piece of evidence that graph topology matters. The soft spots, in proportion: The missing graph-free encoder control is the load-bearing gap, and the stress-test note is correct on this. If you concatenate frozen ESM-2 and MoLFormer embeddings of a strain's genes and target compound and feed them to an MLP, and that gets close to R²=0.41, the attribution to graph structure is substantially weakened. The authors say this is planned — good, but it should have been in the paper. The graph-internal comparisons (HGT vs GraphSAGE, Table 3) partially address this since all use the same node features, but they do not isolate the graph-free case. No error bars across the 5 CV folds is a secondary concern — with n=410 test experiments the R²=0.41 vs 0.24 gap is probably real, but we cannot confirm significance. The depth ablation showing the shallower model outperforms the headline deeper config (ΔR²=0.011) is minor but worth noting; the authors retain the deeper config for consistency with the 3B run, which is a defensible choice. The 'foundation model' framing is aspirational with only one downstream task. The authors acknowledge this and list planned multi-task probes. Code, data, and weights are not released, which limits reproducibility — though the literature-mined benchmark is promised for a forthcoming publication. Who this is for: researchers at the intersection of ML and metabolic engineering who want to see whether graph-based representations can improve strain-design triage. It is a workshop paper with clear limitations and honest self-assessment. The engineering is real, the result is real, but the attribution claim needs the one ablation they already identified. Recommendation: Accept for peer review. The missing control is addressable in revision, and the resource contribution (the KG, the sampling pipeline, the multi-task pretraining design) is substantial enough to warrant serious referee attention. A referee should specifically request the graph-free encoder ablation and per-fold variance estimates.","headline":"Well-engineered metabolic KG with a missing control that weakens the central attribution claim","tokens_in":20148,"tokens_out":790,"would_cite":false,"duration_ms":196030,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"glm-5.2","headline":"Graph foundation model nearly doubles titer prediction over tabular ML","keywords":["heterogeneous graph neural network","metabolic engineering","fermentation titer prediction","knowledge graph","self-supervised pretraining","graph foundation model","multi-modal representation learning","strain design"],"falsifier":"Concatenate frozen ESM-2 and MoLFormer embeddings of each strain's genes and target compound into a flat vector, apply the same MLP probe, and use the same train/test split. If this graph-free probe matches or exceeds R²=0.41, the heterogeneous graph transformer is not the source of Canopy's improvement.","tokens_in":19275,"feed_emoji":"","tokens_out":1311,"duration_ms":150622,"temperature":0.7,"pith_summary":"This paper argues that a pretrained heterogeneous graph neural network, operating over a unified knowledge graph spanning genes, proteins, metabolites, reactions, pathways, strains, and fermentation experiments across organisms, can learn representations that substantially improve prediction of fermentation product titers. The authors build a 6.9-million-node knowledge graph integrating ten data sources, encode each node type with domain-specific foundation models (ESM-2 for proteins, MoLFormer for chemicals, PubMedBERT for text), and pretrain a Heterogeneous Graph Transformer with four self-supervised objectives. On held-out fermentation experiments, frozen graph embeddings fed to a lightweight probe achieve R²=0.41, compared to R²=0.24 for the best tabular baseline and lower scores for homogeneous graph variants. The central claim is that the relational structure of a cross-organism, multi-modal knowledge graph encodes information useful for predicting biological production outcomes that flat feature vectors and single-organism graph models cannot capture.","feed_headline":"Graph foundation model nearly doubles titer prediction over tabular ML","feed_subtitle":"A 6.9M-node heterogeneous knowledge graph of genes, proteins, and fermentation experiments yields embeddings that predict microbial chemical","key_machinery":"The central object is Canopy's heterogeneous knowledge graph: 6.9 million nodes across 13 types (metabolites, reactions, UniRef protein clusters, InterPro domains, genes, chassis organisms, strains, pathways, GO terms, taxa, experiments, transcriptomic measurements) connected by 34 typed edge relations. Node features are dispatched by property prefix to frozen foundation-model encoders — ESM-2 (650M) for protein sequences, MoLFormer-XL for SMILES strings, PubMedBERT for biomedical text — and combined with SignNet Laplacian positional encodings and random-walk structural encodings. The encoder is a stack of Heterogeneous Graph Transformer convolution layers with per-type and per-relation注意力,","core_discovery":"The paper's central result is that a heterogeneous graph transformer pretrained on a multi-modal metabolic-engineering knowledge graph produces frozen embeddings that predict fermentation titer with R²=0.41 using only a lightweight probe, outperforming tabular baselines (best R²=0.24), homogeneous GraphSAGE (R²=0.24), and vanilla HGT without architectural augmentations (R²=0.31). The improvement holds across model scales from 80M to 3B parameters, with the largest model achieving the best R² and AUROC. Ablations show that virtual nodes contribute the most among architectural augmentations (removing them drops R² by 10.3 points), that learned uncertainty weighting of the four pretraining loss","pith_inferences":["The missing graph-free control (concatenating frozen ESM-2 and MoLFormer embeddings without any graph structure) is the critical falsifier: if an MLP on pooled encoder features matches or approaches R²=0.41, the heterogeneous graph transformer is not the source of the improvement, and the result reduces to 'foundation-model features are good for titer prediction.'","The shallow model (L=2) outperforming the deep model (L=6) in ablation suggests oversmoothing is already active at modest depth on this graph, which raises the question of whether the 3B model's advantage comes from depth-related representation quality or simply from wider per-head dimensions.","If the AUROC task is partly trivialised by fermentation-volume correlation (AUROC 0.65 from volume alone), the regression R² is the more trustworthy metric, and the binary classification gains should be interpreted cautiously.","The virtual-node ablation's large effect (−10.3 R²) may indicate that graph diameter reduction matters more than heterogeneous message passing on this particular graph, which would reframe the contribution from 'heterogeneous modeling' to 'graph connectivity engineering.'"],"forward_implications":["If the graph structure genuinely drives the improvement, the same pretrained embeddings could serve as conditioning for generative strain-design pipelines — the authors describe pairing them with flow-matching models and Bayesian optimization loops to propose multi-gene interventions scored by the frozen titer probe.","The 4,791-experiment benchmark with deterministic splits addresses a gap in metabolic-engineering ML evaluation: no existing resource combines experimental titer measurements, cross-organism coverage, and multi-omic context with held-out evaluation, which could make it a standard test bed if released.","Cross-organism transfer becomes testable: the heterogeneous graph naturally links homologs, pathways, and compounds across species, so pretraining on well-studied organisms and evaluating on underexplored ones is a direct extension the authors propose.","The modest scaling gain from 500M to 3B parameters (R² 0.38→0.41) suggests the current bottleneck is experimental data volume, not model capacity, implying that data expansion — not larger models — is the limiting factor for further improvement."],"fun_headline_variants":["Canopy: Heterograph foundation model predicts fermentation titer","Multi-modal metabolic KG doubles titer prediction vs tabular baselines","Frozen graph embeddings beat tabular ML at predicting fermentation titer","Canopy: Heterogeneous graph foundation model for metabolic engineering","Domain-specific foundation models power 6.9M-node metabolic KG"],"cache_read_input_tokens":0,"weakest_assumption_plain":"The claim that heterogeneous graph structure drives the titer-prediction improvement rests on comparisons against tabular baselines and homogeneous graph variants, but the paper has not yet run the decisive control: feeding the same frozen ESM-2 and MoLFormer embeddings into an MLP without any graph structure. Without that ablation, the improvement could come from the pretrained feature encoders rather than from the graph transformer.","fun_headline_variants_meta":{"raw":{"variants":["Canopy: Heterograph foundation model predicts fermentation titer","Multi-modal metabolic KG doubles titer prediction vs tabular baselines","Frozen graph embeddings beat tabular ML at predicting fermentation titer","Canopy: Heterogeneous graph foundation model for metabolic engineering","Domain-specific foundation models power 6.9M-node metabolic KG","Virtual nodes and learned loss weighting drive titer prediction gains","Frozen Canopy embeddings reach R²=0.41 on fermentation titer prediction","Knowledge graph foundation model outperforms homogeneous GNNs on titer"]},"model":"glm-5.2","effort":"high","cost_usd":0.0,"raw_usage":{"total_tokens":772,"prompt_tokens":597,"completion_tokens":175,"prompt_tokens_details":null},"tokens_in":597,"tokens_out":175,"duration_ms":14583,"temperature":1.0,"reasoning_tokens":44,"cache_read_input_tokens":0,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-07-08T12:42:56.124973+00:00","model_set":{"reader":"glm-5.2"},"falsifier":"Concatenate frozen ESM-2 and MoLFormer embeddings of each strain's genes and target compound into a flat vector, apply the same MLP probe, and use the same train/test split. If this graph-free probe matches or exceeds R²=0.41, the heterogeneous graph transformer is not the source of Canopy's improvement.","supporting_citations":[],"review_version":1}