Frontiers in Bioinformatics

Peripheral whole-blood gene expression profiling offers a minimally invasive route to lung cancer detection, but high-dimensional transcriptomic data are prone to optimistic bias when preprocessing and model selection are not properly separated from performance evaluation. We applied L1-penalised (LASSO) logistic regression to 303 peripheral whole-blood microarray profiles (123 lung cancer cases and 180 healthy controls; Gene Expression Omnibus accession GSE252168; Illumina HumanHT-12 v4) within a leakage-free nested cross-validation framework (5 outer and 3 inner folds), in which all data-dependent steps (imputation, univariate feature screening by ANOVA F-test with k = 500, and standardisation) were confined strictly to training partitions. Statistical significance was assessed by permutation testing (B = 100), and feature selection stability was quantified across outer folds. LASSO was compared with ridge logistic regression, linear support vector machines, and random forest under the same framework. The LASSO model identified a sparse 29-probe signature with a pooled out-of-fold area under the ROC curve (AUC) of 0.990 (nested estimate 0.989 +/- 0.015), accuracy 97.4%, sensitivity 94.3%, and specificity 99.4% at a 0.50 threshold; permutation testing confirmed significance (p = 0.0099). Six probes, including CDC42, U2AF1, and RPS15A, were selected in all five outer folds, forming a stable core, and all classifiers exceeded AUC 0.987, indicating a strong, algorithm-independent signal. A leakage-free nested cross-validation framework enables unbiased performance estimation and reproducible feature selection in blood-based lung cancer classification. The 29-probe panel is an internally validated candidate requiring prospective, multicentre external validation before clinical use.

Classification of research articles according to translational research stages enables funding bodies, academic and medical institutes, and policymakers to objectively assess the distribution of resources across the research spectrum. We aim to utilise Large Language Models (LLM) to classify medical research papers into translational research levels based on their titles and abstracts, comparing performance across a range of LLMs, multiple runs and a bag of words (BoW) baseline. We quantify the performance of open-weight LLMs against a human-labelled data set of 318 medical research papers. Using a description of translational levels, the LLMs showed good performance with an F1 score of 0.83 ahead of a baseline BoW approach of 0.68. We show that LLMs can accurately classify titles and abstracts into translational levels within a fully automated pipeline.

Protein structure databases have grown from thousands of experimentally determined structures to hundreds of millions of AI-predicted models, creating an urgent need for search methods that combine high accuracy with practical scalability. Here, I present the third generation of Reseek, a protein structure search algorithm achieving the highest overall accuracy (median rank 1) according to diverse metrics among tested methods including DALI, Foldseek and TM-align. Improved accuracy is obtained by parallel sequence alignment of many discrete alphabets capturing primary, secondary and tertiary features, giving a combined space of[~] 1022 possible states. Separate statistical models are optimized for family, superfamily and fold discrimination, respectively, revealing distinct combinations of features that characterize each level. Hundreds of query structures can be searched a against a multimillion-structure database on a server computer in minutes, making large-scale structure search at state-of-the-art accuracy practical on commodity hardware.

MotivationSpurious protein sequences, resulting from gene prediction errors, theoretically should not yield folded structures. AlphaFold2 was previously shown to predict short spurious sequences with high pLDDT scores and was therefore unlikely to distinguish between real proteins and spurious proteins which are usually short. We evaluate whether newer structure prediction methods (ESMFold and AlphaFold3) similarly predict short sequences with high pLDDT or if they better discriminate between spurious and real proteins. ResultsAll three structure prediction methods (ESMFold, AlphaFold2, and AlphaFold3) predict short spurious sequences from AntiFam with unexpectedly high pLDDT scores, however the discrimination between spurious and real proteins improves beyond 100 amino acids. By analysing sequences with disparate pTM and pLDDT scores, we identified two likely spurious shadow ORFs in Swiss-Prot and one potentially non-spurious AntiFam entry. Using the structure prediction scores, we developed a Gaussian Process Model and evaluated its performance on AlphaFold DB, identifying potential spurious proteins at scale. While limited on its own, this model can increase confidence in spurious protein identification when combined with other methods. AvailabilityStructure predictions are available at https://doi.org/10.5281/zenodo.18390113. Model implementation and figure generation code are available at https://github.com/0rra/fold_unfold2.

External quality assessment (EQA) of multianalyte assays is commonly interpreted analyte by analyte, although many panels contain known relations among measured features that may reveal joint quality patterns. We propose PathEQA, a feature-graph-guided random forest framework in which a user-supplied graph can represent biochemical pathways, molecular interactions, shared measurement processes, or other domain relations. The same graph is allowed to influence feature representation, node-level candidate generation, and split selection, with an optional local grouped decision. We evaluated the framework in graph-aligned and graph-misspecified simulations and used a six-analyte catecholamine-related liquid chromatography-tandem mass spectrometry EQA data set as an illustrative case study (929 records from 58 laboratories and 117 complete multianalyte panels). In graph-aligned simulations, the grouped variant reduced test root mean squared error by 7.4-9.4% relative to ordinary random forest across training sizes of 60-240, whereas graph misspecification could worsen prediction. In the catecholamine case study, full PathEQA was comparable with ordinary random forest in laboratory-grouped cross-validation (RMSE 0.570 versus 0.569) and modestly better in the final-round temporal holdout (0.307 versus 0.318); a simpler static network-sampling baseline performed best. Dopamine-norepinephrine was the strongest pair, whereas dopamine-norepinephrine-epinephrine best estimated multianalyte failure burden. These results support a general conclusion: feature-graph guidance can improve small-sample multivariate quality assessment when the supplied structure is outcome-relevant, but graph relevance must be tested rather than assumed. Catecholamines serve here as a worked example rather than a restriction of the framework.

Short open reading frames (sORFs) within non-coding RNAs (ncRNAs) have arisen as a hidden layer of gene regulation, encoding small peptides that represent a new class of cancer regulators with diagnostic and therapeutic potential. However, inferring associations between sORFs to specific cancer types remains challenging and requires computational approaches for accurate prediction. Recently, the CoraL framework introduced the first computational approach for predicting cancer-associated peptides, focusing primarily on model architecture while overlooking how feature extraction strategies influence predictive accuracy. We present a systematic evaluation of machine learning models and feature extraction approaches to predict cancer-associated sORFs across 15 cancer types. We benchmarked seven traditional machine learning algorithms combined with three feature extraction methods: k-mer frequency, Word2Vec embeddings, and genomic language model (gLM)-based embeddings. To our knowledge, this is the first study applying gLM-derived embeddings to the prediction of cancer-associated sORFs in ncRNA. Our results show that traditional machine learning models with appropriate feature extraction outperform the CoraL baseline across all cancer types, achieving up to 10% higher accuracy in some of the 15 evaluated datasets. Interestingly, k-mer features consistently outperformed gLM embeddings without fine-tuning, suggesting that local sequence composition may provide more discriminative information for this task and that pre-trained genomic representations may require task-specific adaptation to fully capture these patterns. Additionally, we observed that the way sequences are tokenized, such as the k-mer length, can affect performance: longer fragments (e.g., k=7) sometimes reduced accuracy for Random Forest but had a smaller effect on MLP. Our findings suggest that appropriate feature engineering can provide greater improvements than increasing model complexity.

Machine learning workflows frequently incorporate data preprocessing to enhance predictive performance. However, the need for Super Learner ensembles made up only of preprocessing-invariant tree-based algorithms remains unexplored. Using three benchmark clinical classification datasets, this study examined how preprocessing affected the Super Learner's prediction performance, learner weight distribution, and oracle behavior. The Heart Disease (207 observations), Indian Liver Patient Dataset (583 observations), and Pima Indians Diabetes (768 observations) datasets were used to create a Super Learner ensemble model that included Classification and Regression Trees (CART), Random Forest, Ranger, and Extreme Gradient Boosting (XGBoost). Models were evaluated under raw and preprocessed data conditions using repeated cross-validation. Predictive performance was assessed using the area under the receiver operating characteristic curve (AUC), Matthews correlation coefficient (MCC), and Brier score. Learner weight allocation and Oracle Gap were compared using paired Wilcoxon signed-rank tests with Benjamini-Hochberg adjustment. Preprocessing produced negligible changes in predictive performance for the Heart Disease and Pima datasets. For the ILPD dataset, preprocessing significantly improved AUC (0.746 to 0.752; adjusted p = 0.0017) and reduced the Brier score (0.177 to 0.175; adjusted p < 0.001). Learner weights remained largely stable, although Random Forest replaced Ranger as the dominant learner for the Heart Disease dataset. Oracle Gaps remained extremely small (<0.002) across all datasets and did not differ significantly between preprocessing conditions. Preprocessing provides limited benefit for Super Learner ensembles composed of preprocessing-invariant learners and does not materially alter their oracle behavior. Preprocessing decisions should therefore be guided by dataset characteristics rather than adopted as a universal modelling practice.

Performance estimates in protein function prediction depend not only on model choice but also on upstream decisions that define the learning problem. Using antioxidant protein classification as a controlled case study, we evaluated how dataset harmonisation, protein representation, redundancy control, and partitioning strategy affect protein machine learning pipelines. We integrated 18,804 records from 12 publicly available dataset entries into a curated consensus dataset of 4,193 protein sequences. One-hot encoding and six pretrained protein language model representations were evaluated as model inputs and as similarity spaces for redundancy reduction and distance-aware splitting. Representation choice substantially altered dataset geometry, retained dataset size, class balance, and downstream evaluation. At representation-specific p90 thresholds, one-hot encoding retained the complete dataset, whereas pretrained embeddings retained between 5% and 25% of sequences. Distance-aware partitioning reduced apparent performance relative to random splitting by up to 0.15 MCC before redundancy control, while this difference narrowed after similarity filtering. Selected configurations nevertheless maintained high performance under stricter evaluation, reaching an MCC of 0.84. These findings show that performance estimates should be interpreted as outcomes of complete data-centric workflows rather than isolated properties of predictive models.

The N-linked glycans are structurally complex carbohydrate modifications that regulate protein folding, immune recognition, and cellular signaling, and their expression is extensively remodeled during cancer progression, making them promising biomarkers. In this study, prostate cancer-associated N-glycans from a range of relevant peer-reviewed studies were curated and digitized to develop a versatile computational framework that quantitatively encodes their spatial complexity across diverse biological systems. We invented two indices--the Distance & Connectivity Index (DCI) and the Position & Composition Index (PCI)--to capture the spatial information in N-glycans as layered architectures, enabling calculation of residue-level path lengths, branching structure, and compositional diversity. DCI summarizes glycan structure as both a scalar and matrix representation, while PCI does the same but also captures monosaccharide diversity, linkage heterogeneity, and cross-layer branching features. These metrics were computed with GlycoAssessor, an open-source platform that extracts information for the DCI and PCI from glycans drawn via Symbol Nomenclature for Glycans (SNFG) notation. Principal Component Analysis (PCA) was applied to evaluate whether glycans from prostate cancer tissues cluster distinctly in a disease-relevant manner. Results show that the spatial information in N-glycans: (1) increased in a multi-dimensional, non-linear manner, (2) objectively segregated structural themes, (3) could function as a potential prostate cancer biomarker that is distinct from mass-to-charge ratio and relative abundance, and (4) could objectively quantify novel subtype classifications of glycans associated with disease states and progression.

Missing data is a fundamental challenge in space biology, where high experimental costs, limited sample availability, and tissue allocation constraints produce datasets that are sparse, multimodal, and heterogeneous. We present a systematic four-stage framework for diagnosing, implementing, and validating data imputation strategies tailored to these characteristics, and demonstrate its application to retinal imaging and omics data from the NASA Rodent Research 9 (RR9) mission. Using logistic regression-based missingness diagnosis, we identify a Missing At Random (MAR) mechanism driven by experimental design constraints across nine assay modalities. We implement and optimize three imputation strategies: K-Nearest Neighbors (KNN), Multiple Imputation by Chained Equations with weak ElasticNet regularization (MICE-Elastic), and a per-column hybrid strategy, evaluated against a random sample imputer baseline. Validation across seven complementary metrics including supervised classification, unsupervised clustering, correlation structure preservation, masked value recovery, cross-dataset generalization, and permutation testing reveals that MICE-Elastic and the Hybrid strategy preserve genuine biological signal in both RNA-seq and TUNEL modalities, while KNN and the random sample imputer do not despite achieving comparable cross-validation accuracy. A critical finding is that imputation substantially improves supervised classification performance while consistently degrading unsupervised clustering structure, a trade-off researchers must understand before applying these methods. This framework provides practical, actionable guidance for space biologists and data scientists managing sparse multimodal datasets, and represents a foundational step toward digital twin development for space medicine.

The therapeutic landscape for hepatocellular carcinoma (HCC) is evolving rapidly, necessitating scalable approaches to synthesize the expanding scientific literature. We characterized thematic shifts in HCC treatment and prognosis research by conducting a retrospective bibliometric analysis of influential publications from 2023 and 2024. Using the OpenAlex database, we identified the 50 most highly cited papers from each year based on eighteen-month post-publication citation counts. Large language models were deployed to extract, normalize, and classify concepts from unstructured text into canonical topics and parent themes, enabling quantitative year-over-year frequency comparisons. Analysis of these 100 papers revealed a distinct maturation in research focus. Although broad categories like general immunotherapy remained prevalent, their relative frequency declined in favor of specific dual immune checkpoint regimens, notably CTLA-4 inhibition and the durvalumab plus tremelimumab combination. Concurrently, parent themes related to radiomics, imaging, and health systems exhibited significant growth in the 2024 cohort. These findings demonstrate a thematic transition in high-impact HCC research from foundational immuno-oncology toward optimized combination therapies and precision diagnostics. Furthermore, this study highlights the utility of artificial intelligence-driven bibliometrics for objectively tracking dynamic conceptual shifts in oncology. A web interface for exploring the data is available at https://pri.pepkio.com/.

SysBio FAIRplex is a Common Fund Venture Program that catalogs and indexes data from the Accelerating Medicines Partnership(R) (AMP(R)) Program through a federated model in which data hosts retain custody of their datasets. The central piece of this work is the SysBio Common Data Model (SysBio CDM). AMP is a precompetitive public-private partnership started in 2014 that unites the resources of NIH and private partners to improve our understanding of disease pathways and transform current models for developing new treatments by: - identifying new targets, biomarkers, and development paradigms; - developing leading-edge tools and technologies; - collecting large-scale datasets and supporting analytics for open analysis by the public; and - generating consensus platforms and procedures. A multidisciplinary Task Force was chartered to design the SysBio CDM by extending the Observational Medical Outcomes Partnership (OMOP) Common Data Model into the -omics domain. The Task Force produced a Minimum Viable Product comprising nine OMOP tables; four extension tables for assay and file metadata; and a Common Data Element (CDE) Registry to specify field semantics. This manuscript describes the deliverable: the underlying design choices, the criteria applied in selecting and constructing the extension tables, how the extended model supports multimodal data integration across AMP projects, and what further work to support additional -omics modalities would entail. As an auxiliary methodology, the paper also describes the AI-assisted CDE harmonization workflow used to populate the model.

Toxicogenomics can support regulatory toxicology, but its use is limited by the difficulty of translating molecular responses into mechanistic, decision-relevant interpretations. Adverse Outcome Pathways (AOPs) provide a framework for this translation, yet omics applications require scalable mapping of Key Events (KEs) to molecular features. Here, we present an AI-assisted, multi-step workflow for KE-to-gene mapping that uses embedding-based semantic retrieval to identify candidate ontology/pathway terms, large language model-assisted refinement to filter these candidates, and double-independent expert group curation with rule-based consolidation to finalize mappings and derive confidence scores. Compared with earlier NLP-based approaches, the workflow improves KE-to-ontology/pathway mapping performance and generates candidate annotations that better align with expert judgment while substantially reducing the need for manual augmentation. Explicit gene and protein mentions in KE titles were additionally grounded to improve specificity, and each curated mapping was assigned curator reason codes to support transparent, traceable, and confidence-aware reuse. Applied across AOP-Wiki, the workflow produced a comprehensive KE-to-gene set resource covering 1,254 KEs across 523 AOPs and linking 15,833 human genes. Utility is demonstrated through CTD-based AOP fingerprinting of curated reference chemical groups, highlighting expanded coverage and confidence-informed interpretation of chemical-associated gene signatures in an AOP context. The workflow and resulting resource provide a practical bridge between toxicogenomics and AOP-based mechanistic interpretation and support routine updating and future extension to additional omics layers within OECD Omics2AOP.

Purpose: Autosomal Dominant Polycystic Kidney Disease is the most common monogenic kidney disease and largely due to variants in PKD1. We aimed to assess pathogenicity evidence for PKD1 missense variants in disease databases and evaluate in silico pathogenicity prediction tool performance. Methods: PKD1 missense variants reported as pathogenic, likely pathogenic or likely benign were extracted from ClinVar and PKDB. Variants were re-classified using ACMG/AMP criteria to identify "truth sets" of pathogenic and benign variants. In silico scores were obtained from five tools (SIFT, PolyPhen-2, CADD, REVEL, AlphaMissense) and evaluated using established thresholds. A Receiver Operating Characteristic curve analysis was performed using the PKD1 variant truth sets. Results: 346/389 (89%) reported disease-causing missense variants in PKD1 were downgraded to Variants of Unknown Significance (VUS) using current classification criteria. Based on current thresholds, REVEL achieved the highest sensitivity of 62%, with specificity of 79%. AlphaMissense was the only tool not to misclassify any truth set variants, but many of the variant scores were between the pathogenic and benign thresholds. Conclusion: A large majority of PKD1 missense variants are classified as VUS with current pathogenicity criteria. Commonly used in silico tools, applied with established genome-wide thresholds, do not reliably distinguish pathogenic and benign missense variants in PKD1.

Hurthle cell carcinoma (HCC) is an aggressive form of thyroid cancer. While mitochondrial DNA mutations and chromosomal losses have been identified in HCC, isoform switching, and its functional consequences remain uncharacterized. This study reanalyzed NCBI GEO dataset GSE228870 (n = 32), using Salmon and IsoformSwitchAnalyzeR() to identify isoform switching. The analysis resulted in 371 switches across 335 genes showing functional consequences including loss of protein domains, shorter open reading frames (ORFs), loss of signal peptides and novel sub-cellular localizations. Most significant isoform switches (q-value < 0.05, |dIF| > 0.1) were observed in LAMA2, LSP1, MAD2L2, FBLN2 and CXCL12, implicating extracellular matrix dysregulation, DNA damage response, immune signaling and cytoskeleton regulation. These genes are expressed in normal thyroid (median TPM 20.69, 11.66, 14.79, 134.1 & 80.76). However, specific isoforms of LAMA2 and MAD2L2 are not expressed in normal thyroid, explaining tumor-specific expression in HCC. Alternative transcription termination site (ATTS) gain was significant, suggesting altered 3 end in HCC transcripts. TCGA SpliceSeq showed LSP1, FBLN2 and CXCL12 undergo alternative promoter (LSP1 exon1 PSI=94.5%, FBLN2 exon2 PSI=99.0%) and alternative termination (CXCL12 exon3.3 PSI=53.9%) in thyroid cancer, suggesting ATTS and alternative transcription start site (ATSS) as shared splicing dysregulation mechanisms. This is the first systematic characterization of isoform-level dysregulation in HCC.

Motivation: Graph Neural Networks (GNNs) have gained increasing interest in the biomedical domain, as the integration of prior knowledge and deep neural networks has the potential to enhance insights into molecular processes and disease mechanisms. However, a comprehensive and systematic assessment of model architectures, data modalities, graph structures, and their performance for graph signal classification in the biomedical domain is yet to be performed. In order to close this gap, we conducted a benchmarking study on multiple GNNs on a Protein-Protein Interaction (PPI) network for Kidney Renal Clear Cell Carcinoma and Breast cancer subtype prediction, performing an in-depth investigation of architectures, incorporating skip connections and various data modalities. Results: While none of the GNNs outperforms the structure-agnostic Multi-Layer Perceptron baseline, all of them can handle bimodal data (gene methylation and expression) and offer the ability to gain explainability based on PPIs. We offer practical guidelines for applying GNNs to graph signal processing tasks specifically for cancer classification. Depending on the underlying dataset and PPI structure employed, models on different data modalities outperform others. Overall, we suggest using ChebNet, which tends to outperform the Graph Convolutional Network and the Graph Attention Network in cancer subtype prediction. We recommend using GNN architectures that employ a simple flattening readout layer, as they provide better classification performance and faster training time than those with global average pooling. Additionally, we tested residual connections, but they had only an insignificant impact on classification performance.

The Protein Data Bank (PDB) is an ever-growing, open-access repository of structural data of biological molecules. This international database has been instrumental in the development of artificial intelligence and deep learning models for protein structure prediction and design. The PDB growth is a crucially important factor influencing further development of these models. Therefore, after analyzing the growth trend in PDB depositions since the archive's launch, we found that it is well fitted by the Gompertz function, a growth curve used across various disciplines. Furthermore, we observed that the function captures the "discovery of novel folds", i.e., the cumulative number of distinct folds among protein domains that constitute most of the PDB. Consequently, based on the fitting results, we estimated the likely numbers of PDB entries and protein folds. These findings provide insights into deceleration of growth in recent years and enable us to assess anticipated trends.

Motivation: Antimicrobial resistance (AMR) has been identified as a top global public health threat. Accurate AMR phenotype prediction from whole-genome sequencing data is an essential tool for accelerating clinical decision-making and mitigating resistance spread. Although many previous works have explored the use of tree-based machine learning (ML) models to predict resistance, the field lacks a systematic evaluation of the training pipeline across a variety of pathogenic species and antibiotics. Results: Using nine clinically relevant species-antibiotic combinations from the NCBI antimicrobial susceptibility testing database, we present a detailed analysis of the ML pipeline and identify key factors affecting model performance and evaluation. We begin by relabelling all isolates using current CLSI minimum inhibitory concentration breakpoints to resolve inconsistencies and increase available data, resulting in up to a 19% label swap and 56% data enlargement per species-antibiotic combination. We identify several key training parameters including k-mer length, which can increase classification F1 scores by over 20 points compared to commonly used k-values, feature matrix truncation, which can induce polynomial time reductions with limited performance reduction, and ML model class. By comparing 5-fold cross-validation with evaluation on an unseen clinical dataset, we show that random cross-validation splits--often criticized as overly optimistic--can act as a strong proxy for downstream clinical performance, yielding closer F1 scores than phylogeny-aware splits in all cases. We finally present an interpretability study which shows that over 95% of k-mers used by our models are associated with identifiable genomic features. Our results highlight the importance of feature design, evaluation protocol, and biological analysis in genomic AMR prediction, and support tree-based models as a robust and interpretable method.

Summary: Prot2Surf is a software tool designed for the characterization and prediction of protein association to surfaces. In this application note, Prot2Surf was tested using catalytic domains of the lytic polysaccharide monooxygenases (LPMOs), interacting with native surfaces. The results show that the software can efficiently analyze key binding features, including protein-surface distances, distances between catalytically reactive atoms, and the orientation angle between surface chains and the protein. These features are essential for distinguishing productive binding poses in these protein-surface systems and for understanding interaction patterns that provide guidance on protein engineering. Prot2Surf performs these analyses within seconds to a few minutes, providing a fast and accessible framework to post-process and characterize protein-surface encounter complexes. Availability and implementation: Prot2Surf, which is written in Fortran90, is documented and freely available as open source on GitHub: https://github.com/TUNNELING-GROUP/Prot2Surf. In order to run Prot2Surf, users should also install the SDA software package which is freely available at https://www.h-its.org/downloads/sda7/.

Discrete Ricci curvature is an appealing descriptor for single-cell trajectory graphs, but its practical value depends on task validity, graph-topology controls, and whether the biological target is a local transition region or a broader fate decision. We present a controlled empirical study of when curvature features add information to single-cell lineage graphs after repairing unsupported task definitions and preserving strong graph baselines. Paul15 is used as a benchmark-repair and boundary-setting dataset: on the repaired annotation-informed branch-region proxy, graph-plus-Forman and graph-plus-Ollivier improve modestly over graph topology alone (exact AUPRC 0.652 and 0.635 versus 0.607). Pancreas provides the clearest positive transition-region ranking result. On a Fev+ endocrine transition-region benchmark, graph-plus-Ollivier reaches exact AUPRC 0.761 versus 0.669 for the graph-feature stack (five-split canonical evaluation, paired mean +0.092, 95% CI [0.076, 0.114]), and a restricted preterminal endocrine fate task improves under both curvature hybrids. The primary evaluation is transductive node ranking; graph-attachment analyses provide supporting out-of-sample robustness checks. Zebrafish provides a realism check: branch-region ranking again benefits from a hybrid model, with graph-plus-Forman strongest (0.704 versus 0.593 for graph topology), whereas a valid early Notochord versus Prechordal Plate task is graph-topology dominated on the canonical graph. Supplementary checks, including a stricter zebrafish sample-token holdout, biologically grounded bottleneck proxy, negative controls, and pairwise transfer, sharpen the same conclusion without expanding the claim set. Curvature can add useful, dataset-dependent hybrid signal for lineage transition-region tasks, but curvature-only models are weak and graph topology remains essential and sometimes sufficient.