<?xml version="1.0" encoding="UTF-8" ?>
<rdf:RDF xmlns:admin="http://webns.net/mvcb/" xmlns="http://purl.org/rss/1.0/" xmlns:rdf="http://www.w3.org/1999/02/22-rdf-syntax-ns#" xmlns:prism="http://purl.org/rss/1.0/modules/prism/" xmlns:taxo="http://purl.org/rss/1.0/modules/taxonomy/" xmlns:content="http://purl.org/rss/1.0/modules/content/" xmlns:dc="http://purl.org/dc/elements/1.1/" xmlns:syn="http://purl.org/rss/1.0/modules/syndication/">
<channel rdf:about="https://biorxiv.org">
<admin:errorReportsTo rdf:resource="mailto:biorxiv@cshlpress.edu"/>
<title>bioRxiv Subject Collection: Bioinformatics</title>
<link>https://biorxiv.org</link>
<description>
This feed contains articles for bioRxiv Subject Collection "Bioinformatics"
</description>

<items>
<rdf:Seq>
<rdf:li rdf:resource="https://www.biorxiv.org/content/10.64898/2026.11.16.766906v1?rss=1"/>
<rdf:li rdf:resource="https://www.biorxiv.org/content/10.64898/2026.11.16.766980v1?rss=1"/>
<rdf:li rdf:resource="https://www.biorxiv.org/content/10.64898/2026.11.15.766634v1?rss=1"/>
<rdf:li rdf:resource="https://www.biorxiv.org/content/10.64898/2026.11.15.766670v1?rss=1"/>
<rdf:li rdf:resource="https://www.biorxiv.org/content/10.64898/2026.11.15.766769v1?rss=1"/>
<rdf:li rdf:resource="https://www.biorxiv.org/content/10.64898/2026.11.14.766416v1?rss=1"/>
<rdf:li rdf:resource="https://www.biorxiv.org/content/10.64898/2026.11.15.766815v1?rss=1"/>
<rdf:li rdf:resource="https://www.biorxiv.org/content/10.64898/2026.11.15.766754v1?rss=1"/>
<rdf:li rdf:resource="https://www.biorxiv.org/content/10.64898/2026.11.15.766718v1?rss=1"/>
<rdf:li rdf:resource="https://www.biorxiv.org/content/10.64898/2026.11.13.766595v1?rss=1"/>
<rdf:li rdf:resource="https://www.biorxiv.org/content/10.64898/2026.11.13.764046v1?rss=1"/>
<rdf:li rdf:resource="https://www.biorxiv.org/content/10.64898/2026.11.15.765655v1?rss=1"/>
<rdf:li rdf:resource="https://www.biorxiv.org/content/10.64898/2026.11.12.766252v1?rss=1"/>
<rdf:li rdf:resource="https://www.biorxiv.org/content/10.64898/2026.11.13.766601v1?rss=1"/>
<rdf:li rdf:resource="https://www.biorxiv.org/content/10.64898/2026.11.14.766684v1?rss=1"/>
<rdf:li rdf:resource="https://www.biorxiv.org/content/10.64898/2026.11.13.766457v1?rss=1"/>
<rdf:li rdf:resource="https://www.biorxiv.org/content/10.64898/2026.11.12.766202v1?rss=1"/>
<rdf:li rdf:resource="https://www.biorxiv.org/content/10.64898/2026.11.06.764786v1?rss=1"/>
<rdf:li rdf:resource="https://www.biorxiv.org/content/10.64898/2026.11.07.764304v1?rss=1"/>
<rdf:li rdf:resource="https://www.biorxiv.org/content/10.64898/2026.11.13.766490v1?rss=1"/>
<rdf:li rdf:resource="https://www.biorxiv.org/content/10.64898/2026.11.13.765983v1?rss=1"/>
<rdf:li rdf:resource="https://www.biorxiv.org/content/10.64898/2026.11.13.765998v1?rss=1"/>
<rdf:li rdf:resource="https://www.biorxiv.org/content/10.64898/2026.11.10.765731v1?rss=1"/>
<rdf:li rdf:resource="https://www.biorxiv.org/content/10.64898/2026.11.13.766406v1?rss=1"/>
<rdf:li rdf:resource="https://www.biorxiv.org/content/10.64898/2026.11.13.766391v1?rss=1"/>
<rdf:li rdf:resource="https://www.biorxiv.org/content/10.64898/2026.11.13.766436v1?rss=1"/>
<rdf:li rdf:resource="https://www.biorxiv.org/content/10.64898/2026.11.13.766356v1?rss=1"/>
<rdf:li rdf:resource="https://www.biorxiv.org/content/10.64898/2026.11.13.766422v1?rss=1"/>
<rdf:li rdf:resource="https://www.biorxiv.org/content/10.64898/2026.11.13.766327v1?rss=1"/>
<rdf:li rdf:resource="https://www.biorxiv.org/content/10.64898/2026.11.13.766344v1?rss=1"/>
</rdf:Seq>
</items>
<prism:eIssn/>
<prism:publicationName>bioRxiv</prism:publicationName>
<prism:issn/>

<image rdf:resource=""/>
</channel>
<image rdf:about="">
<title>bioRxiv</title>
<url>https://www.biorxiv.org/sites/default/files/bioRxiv_article.jpg</url>
<link>https://www.biorxiv.org</link>
</image>
<item rdf:about="https://www.biorxiv.org/content/10.64898/2026.11.16.766906v1?rss=1">
<title>
<![CDATA[
Pharmacophore-Inspired Virtual Receptor: Generating Realistic Peptide Binding Data for Machine Learning 
]]>
</title>
<link>
https://www.biorxiv.org/content/10.64898/2026.11.16.766906v1?rss=1
</link>
<description><![CDATA[
VReceptor is a virtual receptor simulation framework that generates realistic, sequence-dependent binding data for peptide-based therapeutics. It combines pharmacophore-inspired weighting schemes with amino acid similarity metrics to evaluate peptide design strategies in a controlled environment. Validation using the Prolactin-Releasing Peptide (PrRP) dataset confirms VReceptor's ability to approximate binding behaviours and optimize experimental planning. This framework empowers researchers in computational peptide design and active learning workflows while highlighting areas for further refinement.
]]></description>
<dc:creator><![CDATA[ Conde-Frieboes, K., Stahlhut, C. ]]></dc:creator>
<dc:date>2026-11-17</dc:date>
<dc:identifier>doi:10.64898/2026.11.16.766906</dc:identifier>
<dc:title><![CDATA[Pharmacophore-Inspired Virtual Receptor: Generating Realistic Peptide Binding Data for Machine Learning]]></dc:title>
<dc:publisher>Cold Spring Harbor Laboratory</dc:publisher>
<prism:publicationDate>2026-11-17</prism:publicationDate>
<prism:section></prism:section>
</item>
<item rdf:about="https://www.biorxiv.org/content/10.64898/2026.11.16.766980v1?rss=1">
<title>
<![CDATA[
Sequence-based models predict TF binding with negligible contributions from motif cooperativity 
]]>
</title>
<link>
https://www.biorxiv.org/content/10.64898/2026.11.16.766980v1?rss=1
</link>
<description><![CDATA[
How transcription factors (TFs) selectively recognize their genomic targets among numerous sites containing similar DNA motifs remains an open question. Cooperative binding at neighboring motif sites is widely considered a key mechanism of TF specificity, yet its genome-wide contribution remains unclear. Machine learning models can now predict TF binding directly from DNA sequence, providing opportunities to investigate its underlying determinants. Here, we leveraged BPNet models trained on ~2,000 ENCODE TF binding profiles to systematically assess motif cooperativity. Although detected at individual loci, cooperativity made a negligible contribution to genome-wide binding predictions. Nevertheless, the models captured differences in binding preferences between TFs with similar motif preferences, even within the same cellular environment. These findings highlight the need to critically assess prevailing assumptions and explore alternative mechanisms underlying selective TF binding.
]]></description>
<dc:creator><![CDATA[ Mindel, V., Barkai, N. ]]></dc:creator>
<dc:date>2026-11-17</dc:date>
<dc:identifier>doi:10.64898/2026.11.16.766980</dc:identifier>
<dc:title><![CDATA[Sequence-based models predict TF binding with negligible contributions from motif cooperativity]]></dc:title>
<dc:publisher>Cold Spring Harbor Laboratory</dc:publisher>
<prism:publicationDate>2026-11-17</prism:publicationDate>
<prism:section></prism:section>
</item>
<item rdf:about="https://www.biorxiv.org/content/10.64898/2026.11.15.766634v1?rss=1">
<title>
<![CDATA[
Casanovo-TMT: deep learning-based de novo sequencing of TMTpro-labeled peptides and phosphopeptides 
]]>
</title>
<link>
https://www.biorxiv.org/content/10.64898/2026.11.15.766634v1?rss=1
</link>
<description><![CDATA[
De novo peptide sequencing reads peptides directly from tandem mass spectra, but none of the released deep learning models can sequence peptides labeled with isobaric tags such as TMTpro, which are widely used for multiplexed quantitative proteomics. We extended the residue vocabulary of Casanovo with TMTpro tags and phosphorylation, fine-tuned it on 21 million spectra from our laboratory, most of them TMTpro-labeled, and re-ranked the candidates in the beam by their agreement with the measured precursor mass. The resulting model, Casanovo-TMT, predicted the exact peptidoform for 84.7% of the TMTpro-labeled spectra in our test set and for 87.4 and 79.6% of the spectra in two public global-proteome datasets from other laboratories, whereas the extended vocabulary without fine-tuning gave 0.3%. Discarding the lowest-scoring predictions raised precision to 95% while retaining 81.7% of the test spectra. Casanovo-TMT also predicted the exact peptidoform for 69.2% of the spectra in a public phosphopeptide-enriched dataset and for 66 to 72% of the phosphopeptide spectra in three phosphoproteomic studies from our laboratory that were not used for training. In these studies, it identified numbers of phosphopeptides comparable to or larger than those of database search engines, and proportionally more where database search identified fewer. For the phosphopeptides identified only by Casanovo-TMT, entrapment estimated that more than 90% were correct, and an independent search engine assigned the same sequence to 95 to 98% of the spectra that it identified with confidence. Casanovo-TMT can therefore extend database search in TMTpro-based quantitative proteomics, particularly in studies in which database search identifies few phosphopeptides.
]]></description>
<dc:creator><![CDATA[ Kim, H. B., Kim, K.-e. ]]></dc:creator>
<dc:date>2026-11-17</dc:date>
<dc:identifier>doi:10.64898/2026.11.15.766634</dc:identifier>
<dc:title><![CDATA[Casanovo-TMT: deep learning-based de novo sequencing of TMTpro-labeled peptides and phosphopeptides]]></dc:title>
<dc:publisher>Cold Spring Harbor Laboratory</dc:publisher>
<prism:publicationDate>2026-11-17</prism:publicationDate>
<prism:section></prism:section>
</item>
<item rdf:about="https://www.biorxiv.org/content/10.64898/2026.11.15.766670v1?rss=1">
<title>
<![CDATA[
SRAscout: Semantic discovery of public sequencing data through harmonized SRA and BioProject metadata 
]]>
</title>
<link>
https://www.biorxiv.org/content/10.64898/2026.11.15.766670v1?rss=1
</link>
<description><![CDATA[
The rapid growth of public sequencing data has created great opportunities for data-driven discovery in the life sciences, but it has also made it harder to find relevant datasets. The NCBI Sequence Read Archive (SRA) hosts more than 45 million high-throughput sequencing runs, and the BioProject database, shared by NCBI, the European Nucleotide Archive (ENA) and the DNA Data Bank of Japan, describes the projects behind them. However, their metadata are often incomplete, inconsistently annotated and difficult to search using keywords alone. We present SRAscout, a web-based platform that enables semantic dataset retrieval by integrating SRA run metadata with the corresponding BioProject records. SRAscout collects and indexes metadata for about 714,000 BioProjects, and harmonizes sample attributes such as cell line, tissue, age, sex and treatment into a compact and consistent schema. For semantic embedding, SRAscout employs a biomedical sentence embedding model based on BioBERT, which generates dense vector representations from the title, name, organism and description of each project. Users can search in three modes, with the DOI or PubMed ID of an article, with free text, or with BioProject and SRA study accessions, and SRAscout finds the most related projects in each case. Free-text queries are processed through a hybrid of semantic and full-text search, which finds conceptually relevant studies even when different terms are used to describe the same biology, while exact terms such as gene names are still matched. Structured filters allow refinement by organism, assay type (for example RNA-seq and miRNA-seq), harmonized sample attributes such as tissue, cell type and disease, tag words, and many additional attributes. For each search, users can download a harmonized metadata table and ready-to-use scripts that fetch the sequencing files from ENA. SRAscout also offers a public API for programmatic access. By combining harmonized metadata with context-aware search in a single interface, SRAscout facilitates dataset discovery, meta-analysis and data reuse. The application is freely available at www.srascout.com.
]]></description>
<dc:creator><![CDATA[ Grigoriadis, D., Perdikopanis, N., Chan, C. S., Pavlopoulos, G., Georgakopoulos-Soares, I. ]]></dc:creator>
<dc:date>2026-11-17</dc:date>
<dc:identifier>doi:10.64898/2026.11.15.766670</dc:identifier>
<dc:title><![CDATA[SRAscout: Semantic discovery of public sequencing data through harmonized SRA and BioProject metadata]]></dc:title>
<dc:publisher>Cold Spring Harbor Laboratory</dc:publisher>
<prism:publicationDate>2026-11-17</prism:publicationDate>
<prism:section></prism:section>
</item>
<item rdf:about="https://www.biorxiv.org/content/10.64898/2026.11.15.766769v1?rss=1">
<title>
<![CDATA[
Savant: A Fast, Structure-Free Binding Affinity Predictor For Protein-Ligand Drug Screening 
]]>
</title>
<link>
https://www.biorxiv.org/content/10.64898/2026.11.15.766769v1?rss=1
</link>
<description><![CDATA[
Protein-ligand binding affinity determines which of many candidate molecules is worth synthesizing and assaying, making it the central measurement in early drug discovery. Binding affinity is commonly expressed as the dissociation constant, an equilibrium measure that indicates how easily a complex separates into its components. AI-based generative chemistry and make-on-demand chemical catalogs now produce molecules faster than rigorous method can evaluate them. Leading computational methods co-fold the complex before scoring the binding affinity between the protein and the ligand, which provides high accuracy but requires a graphics processing unit (GPU) and substantial time for every pair evaluated.Here we present Savant, a machine learning model that estimates the dissociation constant using only the protein's amino acid sequence and the ligand's structure written as a line of text (SMILES). Savant needs no 3-dimensional folded structure, no docking, and no GPU. To test Savant fairly, we built a benchmark designed to prevent any training data to leak into the testing set, drawing only on measurements published after the training data was assembled and on molecules chemically unlike anything the model had seen. On an independent benchmark, Savant ranks ahead of every other 3-dimensional structure-free method.Savant is designed to rank compounds against a single protein, and performs as well as other methods that require a GPU, while running only on a central processing unit (CPU). Thus, Savant can screen libraries of 100,000 compounds in minutes with commodity hardware.
]]></description>
<dc:creator><![CDATA[ Guirales-Medrano, S., Ford, C. T., Janies, D. ]]></dc:creator>
<dc:date>2026-11-17</dc:date>
<dc:identifier>doi:10.64898/2026.11.15.766769</dc:identifier>
<dc:title><![CDATA[Savant: A Fast, Structure-Free Binding Affinity Predictor For Protein-Ligand Drug Screening]]></dc:title>
<dc:publisher>Cold Spring Harbor Laboratory</dc:publisher>
<prism:publicationDate>2026-11-17</prism:publicationDate>
<prism:section></prism:section>
</item>
<item rdf:about="https://www.biorxiv.org/content/10.64898/2026.11.14.766416v1?rss=1">
<title>
<![CDATA[
A Biologically Interpretable Single-Cell Pathology Foundation Model for Molecular Decoding of H&E Images 
]]>
</title>
<link>
https://www.biorxiv.org/content/10.64898/2026.11.14.766416v1?rss=1
</link>
<description><![CDATA[
The microscopic structure of human tissues encodes molecular information that is not directly accessible by visual inspection. Current pathology foundation models improve histological interpretation but operate at multicellular resolution. It remains unclear whether routine histology contains structured molecular and cellular information recoverable at single-cell resolution, which biological properties are associated with this recoverability, and where its limits lie. Here, we introduce Loki2, a single-cell pathology foundation model trained on UniSeg, a pan-tissue resource integrating 15 million cells with segmentation, lineage annotations and measured or histology-inferred transcriptomic profiles. We examine the above question across cellular representation, molecular recoverability, and patient-level relevance. Loki2 captures cellular identity and continuous morphological variation, with interpretable latent dimensions linked to physical nuclear phenotypes through DAPI validation. Transcriptomic retrieval and in silico protein staining test molecular recoverability, with gene-level analyses revealing the strongest recovery for genes associated with morphologically distinct cellular identities and functions. Then, patient staging and biomarker discovery connect cellular morphology representations to clinical relevance. Together, these findings establish Loki2 as a biologically interpretable single-cell pathology foundation model and provide the opportunities of recovering cellular and molecular information from routine H&E histology at single-cell resolution.
]]></description>
<dc:creator><![CDATA[ Chen, W., Li, Q., Cheng, H., Zhang, P., Chang, Y., Wang, X., Tran, T. N., Li, C., Yang, Y., Xiao, Y., Zhao, X., Wu, W., Ma, A., Mora, A., Rojas, M., Aggarwal, S., Lin, S. H., Li, S., Ma, Q., Wang, G. ]]></dc:creator>
<dc:date>2026-11-17</dc:date>
<dc:identifier>doi:10.64898/2026.11.14.766416</dc:identifier>
<dc:title><![CDATA[A Biologically Interpretable Single-Cell Pathology Foundation Model for Molecular Decoding of H&E Images]]></dc:title>
<dc:publisher>Cold Spring Harbor Laboratory</dc:publisher>
<prism:publicationDate>2026-11-17</prism:publicationDate>
<prism:section></prism:section>
</item>
<item rdf:about="https://www.biorxiv.org/content/10.64898/2026.11.15.766815v1?rss=1">
<title>
<![CDATA[
Transferring CRISPRi knockdown responses to new cell lines requires screening part of each line 
]]>
</title>
<link>
https://www.biorxiv.org/content/10.64898/2026.11.15.766815v1?rss=1
</link>
<description><![CDATA[
Perturb-seq screens measure knockdown responses in one cell line at a time. Models that transfer responses to new lines rarely beat simple averages, and standard metrics often ignore noise. It is unclear how much of a response is shared and what a new line must contribute. We studied CRISPR interference Perturb-seq of essential genes in four human cell lines, using noise ceilings and calibrated metrics. We also introduce CASTER, a set-based model that adapts the transferred mean to the target line. After noise correction, responses correlated at a median of 0.52, highest between the TP53 wild-type lines. With 30% of a line screened, a linear model and CASTER scored 0.84 and 0.80 versus 0.44 for the transferred mean (1 equals a technical duplicate). CASTER led only at 2.5% screened (0.62 versus 0.47) and fell below the transferred mean without a screen. Response size tracked general essentiality more than line-specific dependency.
]]></description>
<dc:creator><![CDATA[ Kawsar, M., Dip, S. A., Paul, B. K., Noman, A. A., Zhang, L., Moni, M. A. ]]></dc:creator>
<dc:date>2026-11-17</dc:date>
<dc:identifier>doi:10.64898/2026.11.15.766815</dc:identifier>
<dc:title><![CDATA[Transferring CRISPRi knockdown responses to new cell lines requires screening part of each line]]></dc:title>
<dc:publisher>Cold Spring Harbor Laboratory</dc:publisher>
<prism:publicationDate>2026-11-17</prism:publicationDate>
<prism:section></prism:section>
</item>
<item rdf:about="https://www.biorxiv.org/content/10.64898/2026.11.15.766754v1?rss=1">
<title>
<![CDATA[
Infernal-GPU: accurate E-values and sensitive covariance-model search for structured RNAs 
]]>
</title>
<link>
https://www.biorxiv.org/content/10.64898/2026.11.15.766754v1?rss=1
</link>
<description><![CDATA[
Covariance models (CMs) find structured RNAs by scoring sequence and base pairing together, and they underlie the Infernal software and the Rfam database. Their dynamic programming is expensive, so Infernal relies on two approximations: E-values are calibrated on a short stretch of random sequence, and fast sequence-only filters decide which parts of a target the CM examines. How much accuracy these approximations cost, and where, is largely unknown, because removing them on CPUs is prohibitively slow. We present Infernal-GPU, a GPU implementation of Infernal's CM dynamic programming that matches Infernal's results. Its key step evaluates the one serial recurrence of CM scoring, the self-loop of insert-left states, as two parallel prefix scans. One GPU calibrates all 4,227 Rfam models overnight, which makes calibrations on ten times more random sequence routine. These longer calibrations show that the E-values distributed with Rfam are off by more than twofold for one model in eight. An exhaustive search reference shows that sequence filters discard remote homologs whose similarity lies almost entirely in base pairing; behind permissive filters, Infernal-GPU recovers them with a third of the CPU work that Infernal needs at the same setting. Infernal-GPU is freely available at https://github.com/void-echo/infernal-gpu and from PyPI, together with Rfam 15.0 recalibrated on ten times more random sequence.
]]></description>
<dc:creator><![CDATA[ Wang, D., Jin, J., Qiao, J., Wei, L., Wu, S., Liu, Q. ]]></dc:creator>
<dc:date>2026-11-17</dc:date>
<dc:identifier>doi:10.64898/2026.11.15.766754</dc:identifier>
<dc:title><![CDATA[Infernal-GPU: accurate E-values and sensitive covariance-model search for structured RNAs]]></dc:title>
<dc:publisher>Cold Spring Harbor Laboratory</dc:publisher>
<prism:publicationDate>2026-11-17</prism:publicationDate>
<prism:section></prism:section>
</item>
<item rdf:about="https://www.biorxiv.org/content/10.64898/2026.11.15.766718v1?rss=1">
<title>
<![CDATA[
Decompose and Recompose: Reference-Centered Multimodal Adaptation of Protein Language Models for Enzyme Turnover Prediction 
]]>
</title>
<link>
https://www.biorxiv.org/content/10.64898/2026.11.15.766718v1?rss=1
</link>
<description><![CDATA[
Predicting enzyme turnover numbers ($k_{mathrm{cat}}$) from protein sequences and reaction chemistry helps prioritize candidates in virtual screening. Reliable ranking requires distinguishing activity differences between enzyme--reaction pairs. Recent interaction models combine enzyme and chemical information, but two representation challenges remain. First, reactants and products have distinct roles in catalysis: reactants are recognized and transformed, whereas products are formed and released. Yet these models merge the two sides before enzyme interaction, leaving the enzyme's response to each role implicit. Second, enzyme representations produced by cross-attention mix reaction-dependent changes with an enzyme-specific background shared across chemical conditions. This mixture can obscure activity-relevant differences between samples. To address these challenges, we introduce EnzRef, a lightweight multimodal framework that adapts frozen protein language models by decomposing reaction chemistry into reactant and product conditions, expressing each conditioned enzyme response relative to that enzyme's background, and recomposing the responses for activity prediction. Reference-Centered Cross-Attention uses each enzyme's average residue response as an internal reference to expose role-specific conditional deviations. Symmetry-Aware Semantic Recomposition encodes the mean deviation and directed role contrast into a joint activity representation. With 1.72M trainable parameters beyond the frozen encoder, EnzRef achieves state-of-the-art source-domain prediction on CatPred-DB and zero-shot transfer to TurNuP-DB. On unseen enzymes, it improves mean $R^2$ by 19.8% relative to MCKcat, the strongest baseline, without target-domain adaptation.
]]></description>
<dc:creator><![CDATA[ Ding, A., Ma, T., Lin, X., Lin, S., Liu, X., Wei, L. ]]></dc:creator>
<dc:date>2026-11-17</dc:date>
<dc:identifier>doi:10.64898/2026.11.15.766718</dc:identifier>
<dc:title><![CDATA[Decompose and Recompose: Reference-Centered Multimodal Adaptation of Protein Language Models for Enzyme Turnover Prediction]]></dc:title>
<dc:publisher>Cold Spring Harbor Laboratory</dc:publisher>
<prism:publicationDate>2026-11-17</prism:publicationDate>
<prism:section></prism:section>
</item>
<item rdf:about="https://www.biorxiv.org/content/10.64898/2026.11.13.766595v1?rss=1">
<title>
<![CDATA[
Harmonized sample metadata and processed cortical and hippocampal transcriptomes from wild-type and Alzheimer's disease model mice 
]]>
</title>
<link>
https://www.biorxiv.org/content/10.64898/2026.11.13.766595v1?rss=1
</link>
<description><![CDATA[
Public mouse brain single-cell and single-nucleus RNA sequencing (sc/snRNA-seq) datasets are difficult to reuse across studies because sample annotations are heterogeneous and incomplete. We harmonized sample-level metadata for 2,005 cortex and hippocampus samples from 389 studies, covering wild-type (WT) controls and Alzheimer's disease (AD)-related mouse models. Usable count data from 1,863 samples in 360 studies were uniformly processed into 12,628,942 cell and nuclear profiles with cell-type annotations. Expression-based indicators were consistent with the annotated sequencing protocol and sex, and integrated microglial profiles recovered established AD-associated signatures. To illustrate cross-study use, we asked whether microglial AD--WT expression differences varied with age and between cortex and hippocampus, using covariate-adjusted expression reconstructed with Kanade. Genes separated into groups with predominantly negative or positive AD--WT contrasts, including a group whose age-associated contrasts differed between cortex and hippocampus. These findings demonstrate the utility of harmonized metadata for investigating age- and region-dependent AD-associated transcriptional differences across studies. The resource supports explicit, reproducible cohort definitions across public mouse brain datasets.
]]></description>
<dc:creator><![CDATA[ Ijiri, H., Kariyayama, H., Ozaki, H. ]]></dc:creator>
<dc:date>2026-11-17</dc:date>
<dc:identifier>doi:10.64898/2026.11.13.766595</dc:identifier>
<dc:title><![CDATA[Harmonized sample metadata and processed cortical and hippocampal transcriptomes from wild-type and Alzheimer's disease model mice]]></dc:title>
<dc:publisher>Cold Spring Harbor Laboratory</dc:publisher>
<prism:publicationDate>2026-11-17</prism:publicationDate>
<prism:section></prism:section>
</item>
<item rdf:about="https://www.biorxiv.org/content/10.64898/2026.11.13.764046v1?rss=1">
<title>
<![CDATA[
Optimal Predictivity of Random Signatures: Causes and Good Modeling Practices 
]]>
</title>
<link>
https://www.biorxiv.org/content/10.64898/2026.11.13.764046v1?rss=1
</link>
<description><![CDATA[
In high-throughput omics data, randomly-selected sets of markers ("signatures") have large and often optimal phenotypic predictivity. This creates problems of over-interpretation of biological and predictive importance of signatures, of bioinformatics algorithms and expert knowledge-driven selection. We conduct first an empirical study establishing that the phenomenon occurs across a variety of types of datasets/omics types and algorithms (224 combinations, >120 million models fitted). Random signatures are systematically reproducible across datasets. We develop a comprehensive theoretical framework, encompassing definitions, feasibility and impossibility theorems, and analytical formulas. The theory reveals as necessary and sufficient causes of these anomalous behaviors, the existence of large equivalence classes of optimally predictive feature sets, and their structure of information equivalences. We validate the theory with additional experiments and propose good practices intended to strengthen scientific rigor. Our findings apply to both algorithmic discovery and expert knowledge-driven signature selection, and to both existing and future algorithms and expert criteria. They also extend to all types of data with similar equivalence classes of optimal predictors.
]]></description>
<dc:creator><![CDATA[ Ma, S., Tourani, R., Wang, L., Ugurbil, M., Kirsh, T., Wang, J., Winterhoff, B., Statnikov, A., Aliferis, C. ]]></dc:creator>
<dc:date>2026-11-17</dc:date>
<dc:identifier>doi:10.64898/2026.11.13.764046</dc:identifier>
<dc:title><![CDATA[Optimal Predictivity of Random Signatures: Causes and Good Modeling Practices]]></dc:title>
<dc:publisher>Cold Spring Harbor Laboratory</dc:publisher>
<prism:publicationDate>2026-11-17</prism:publicationDate>
<prism:section></prism:section>
</item>
<item rdf:about="https://www.biorxiv.org/content/10.64898/2026.11.15.765655v1?rss=1">
<title>
<![CDATA[
Chimeric and Outward Reads in Short-Read RNA Sequencing Data: A Unified Circular RNA Assembly Framework 
]]>
</title>
<link>
https://www.biorxiv.org/content/10.64898/2026.11.15.765655v1?rss=1
</link>
<description><![CDATA[
Circular RNAs (circRNAs) are RNA molecules that play important roles in regulating gene expression, yet detecting and assembling them from short read RNA sequencing (RNA-seq) data remains challenging. Most methods rely on chimeric reads that cross the back-splicing junction (BSJ), but for lowly expressed circRNAs, these reads are often absent, so many circRNAs go undetected. In this work, we propose TERRACE 2.0, a method that addresses this limitation by leveraging an additional but underutilized signal: outward reads. These are paired-end reads that point away from each other when aligned to the reference genome, which is a direct consequence of sequencing a fragment from a circRNA. Together, this multi-signal framework enables circRNAs to be detected and assembled more accurately from short-read RNA-seq data. TERRACE 2.0 infers candidate BSJs from outward reads by testing which ones give a fragment length consistent with the library, reconstructs each full-length circRNA from a splice graph, and combines outward and chimeric reads in a single representation that selects among the candidates jointly. On eight human tissue samples, TERRACE 2.0 has the best average F-score with and without gene annotations. On an alignment that retains more chimeric reads, it is the most accurate assembler in every sample, while its predecessor TERRACE loses much of its precision. Most notably, outward reads are involved in 36% to 56% of the true circRNAs that TERRACE misses, and on simulated data TERRACE 2.0 also has the highest recall and F-score.
]]></description>
<dc:creator><![CDATA[ Sen, A., Zahin, T., Shao, M. ]]></dc:creator>
<dc:date>2026-11-17</dc:date>
<dc:identifier>doi:10.64898/2026.11.15.765655</dc:identifier>
<dc:title><![CDATA[Chimeric and Outward Reads in Short-Read RNA Sequencing Data: A Unified Circular RNA Assembly Framework]]></dc:title>
<dc:publisher>Cold Spring Harbor Laboratory</dc:publisher>
<prism:publicationDate>2026-11-17</prism:publicationDate>
<prism:section></prism:section>
</item>
<item rdf:about="https://www.biorxiv.org/content/10.64898/2026.11.12.766252v1?rss=1">
<title>
<![CDATA[
Infectious Disease Knowledge Graph (IDKG) for Antiviral and Antibacterial Drug Discovery. 
]]>
</title>
<link>
https://www.biorxiv.org/content/10.64898/2026.11.12.766252v1?rss=1
</link>
<description><![CDATA[
Infectious diseases remain a leading cause of global morbidity and mortality, while therapeutic discovery is hindered by fragmented biomedical data and antimicrobial resistance. To address this challenge, we developed the Infectious Disease Knowledge Graph (IDKG), a queryable platform that integrates curated chemical, biological, pharmacological, and clinical information relevant to infectious-disease drug discovery. Using a customized ORION pipeline, we integrated 20 data sources, including DrugCentral, SMACC, Heli-SMACC, NPASS, ClinicalTrials.gov, and host-virus interaction databases, into a Neo4j graph containing more than 4.6 million nodes and 26 million edges. IDKG supports graph-based reasoning through Clinical Outcome Pathways (COPs) and incorporates curation workflows for small molecules, biologics, and complex mixtures. Illustrative applications include reconstruction of a literature-supported remdesivir pathway for COVID-19 and graph-based mechanistic interpretation of experimentally observed quorum-sensing inhibitor-antibiotic synergies in multidrug-resistant bacteria. The platform provides an interoperable framework for drug repurposing, target prioritization, and mechanism-based hypothesis generation. IDKG is freely available at https://idkg.mml.unc.edu.
]]></description>
<dc:creator><![CDATA[ Martin, H.-J., Beasley, J.-M., Kaneva, V., Tieghi, R., Causey, K., Tropsha, A., Muratov, E. ]]></dc:creator>
<dc:date>2026-11-17</dc:date>
<dc:identifier>doi:10.64898/2026.11.12.766252</dc:identifier>
<dc:title><![CDATA[Infectious Disease Knowledge Graph (IDKG) for Antiviral and Antibacterial Drug Discovery.]]></dc:title>
<dc:publisher>Cold Spring Harbor Laboratory</dc:publisher>
<prism:publicationDate>2026-11-17</prism:publicationDate>
<prism:section></prism:section>
</item>
<item rdf:about="https://www.biorxiv.org/content/10.64898/2026.11.13.766601v1?rss=1">
<title>
<![CDATA[
Machine Learning Prediction of Antibiotic Resistance Genes: Performance, Interpretability, and Cross-Database Redundancy 
]]>
</title>
<link>
https://www.biorxiv.org/content/10.64898/2026.11.13.766601v1?rss=1
</link>
<description><![CDATA[
Antibiotic resistance gene (ARG) classifiers are typically developed and evaluated within a single curated database, leaving open how much two independently curated databases agree and what that implies for apparent cross-database generalization. We trained an XGBoost classifier on amino acid trigram frequencies and physicochemical descriptors using the Comprehensive Antibiotic Resistance Database (CARD), the field's largest curated ARG resource. The model separated resistance genes from non-resistance proteins with near-ceiling accuracy (AUROC = 0.982, MCC = 0.886); a label-shuffling null baseline collapsed to chance (AUROC = 0.508) and bootstrap confidence intervals were narrow, confirming genuine signal rather than data leakage or overfitting. SHAP analysis attributed this performance to biologically interpretable features, notably the trigram KTG, isoelectric point, and aromaticity, each consistent with known catalytic motifs in resistance enzymes. Before testing the classifier against ResFinder, one of CARD's most widely used counterparts, we assessed database independence by clustering sequences with CD-HIT, finding that 96.8-98.7% of ResFinder sequences shared substantial protein identity with CARD even at a permissive 50% identity threshold, indicating the two databases largely represent the same sequence population, with direct implications for how cross-database generalization should be measured. On the small subset of ResFinder sequences genuinely non-redundant with CARD (40-100 sequences, depending on threshold), the classifier failed completely, never classifying any as resistance genes; SHAP traced this failure to sequence length, as non-redundant sequences are nearly three times longer than training proteins, and length is among the model's most influential features. The same pattern emerged by resistance mechanism: the model achieved perfect accuracy on beta-lactamases, compact single-domain enzymes well represented in training (100%, n = 351), but correctly identified only 38.5% of efflux pump components (n = 13), a mechanism built from large, multi-domain transporters rather than compact catalytic enzymes. Together, these results describe a capable, interpretable ARG classifier alongside a quantified methodological caution: cross-database redundancy must be measured rather than assumed, and when a sequence-composition classifier fails, the failure is often explained by protein architecture rather than database origin.
]]></description>
<dc:creator><![CDATA[ Ruthbah, C. A., Sadi, T. H., Jahan, N. E. S., Tanzim Adib, A. N. M. ]]></dc:creator>
<dc:date>2026-11-17</dc:date>
<dc:identifier>doi:10.64898/2026.11.13.766601</dc:identifier>
<dc:title><![CDATA[Machine Learning Prediction of Antibiotic Resistance Genes: Performance, Interpretability, and Cross-Database Redundancy]]></dc:title>
<dc:publisher>Cold Spring Harbor Laboratory</dc:publisher>
<prism:publicationDate>2026-11-17</prism:publicationDate>
<prism:section></prism:section>
</item>
<item rdf:about="https://www.biorxiv.org/content/10.64898/2026.11.14.766684v1?rss=1">
<title>
<![CDATA[
GatorAnchor: Cross-Modal Uncertainty-Aware Learning of Reliable Cell Neighborhoods in Single-Cell Multi-Omics 
]]>
</title>
<link>
https://www.biorxiv.org/content/10.64898/2026.11.14.766684v1?rss=1
</link>
<description><![CDATA[
Cross-modal uncertainty challenges single-cell multi-omics integration because molecular modalities differ in sparsity, technical noise, and context-dependent biological informativeness, producing discordant cell-cell relationships. Treating conflicting evidence as equally reliable can propagate uncertain relationships and obscure biologically meaningful cellular heterogeneity. GatorAnchor is a cross-modal uncertainty-aware framework that learns modality-specific probabilistic representations, quantifies disagreement across modalities and neighborhood evidence sources, and refines cell-cell relationships according to their reliability. Across ten paired transcriptome-protein and transcriptome-chromatin datasets, GatorAnchor achieves the highest mean performance across four evaluation metrics among nine representative integration methods, demonstrating improved resolution of cellular heterogeneity. Candidate relations with greater composite unreliability show lower consistency with reference cell-type structure, whereas filtering highly unreliable relations increases same-type edge purity across all datasets. Modality conflict is associated with transcript-protein measurement discordance, while transcriptome-chromatin analyses reveal lineage-dependent contributions of molecular and decoder-derived evidence. A three-modality transcriptome-chromatin-protein analysis characterizes cluster of differentiation 4 (CD4)-related communities with distinct molecular profiles and consistent within-donor contrasts across 12 donors. Under incomplete and mosaic measurements, GatorAnchor preserves informative cell-population structure, while cross-modal correspondence increasingly depends on paired-cell anchors. These findings support uncertainty-aware neighborhood refinement for reliable and interpretable characterization of cellular heterogeneity.
]]></description>
<dc:creator><![CDATA[ Zhang, Z., Tang, Y., Bian, J., Li, X., Liu, Y. ]]></dc:creator>
<dc:date>2026-11-17</dc:date>
<dc:identifier>doi:10.64898/2026.11.14.766684</dc:identifier>
<dc:title><![CDATA[GatorAnchor: Cross-Modal Uncertainty-Aware Learning of Reliable Cell Neighborhoods in Single-Cell Multi-Omics]]></dc:title>
<dc:publisher>Cold Spring Harbor Laboratory</dc:publisher>
<prism:publicationDate>2026-11-17</prism:publicationDate>
<prism:section></prism:section>
</item>
<item rdf:about="https://www.biorxiv.org/content/10.64898/2026.11.13.766457v1?rss=1">
<title>
<![CDATA[
Diffusion geometry and topology reveal Alzheimer's proteomic states 
]]>
</title>
<link>
https://www.biorxiv.org/content/10.64898/2026.11.13.766457v1?rss=1
</link>
<description><![CDATA[
ABSTRACT INTRODUCTION: Alzheimer's disease (AD) exhibits substantial molecular heterogeneity. Identifying biologically distinct proteomic states may improve understanding of disease heterogeneity and support precision diagnostics. METHODS: Cerebrospinal fluid proteomic data from ADNI were analyzed. Among 6,361proteins tested between cognitively normal and AD groups, 407 remained significant after Benjamini-Hochberg correction (FDR < 0.05), with 404 retained after missingness filtering. Following preprocessing and outlier removal, 130 individuals were included. Diffusion geometry and persistent homology were applied to characterize intrinsic proteomic structure without predefined clusters. The framework was further evaluated in an independent mass spectrometry cohort. RESULTS: Persistent homology revealed multiscale organization of the AD proteomic landscape. At the selected H0 filtration scale, two dominant proteomic states (n = 71 and n = 42) contained most individuals, alongside smaller components. H1 persistence identified nontrivial cyclic structure, indicating continuous higher-order organization beyond discrete state partitions. The dominant states exhibited distinct proteomic profiles and differed substantially in tau-related biomarkers. Elevated tau pathology in one state was reproduced in the independent cohort, whereas A{beta}42 differences were less consistent across cohorts. DISCUSSION: AD proteomic heterogeneity is characterized by multiscale organization combining dominant group structure with continuous higher-order relationships.
]]></description>
<dc:creator><![CDATA[ Shahu, K., Lutz, M. W., Moghekar, A., O'Brien, R., Songdechakraiwut, T. ]]></dc:creator>
<dc:date>2026-11-17</dc:date>
<dc:identifier>doi:10.64898/2026.11.13.766457</dc:identifier>
<dc:title><![CDATA[Diffusion geometry and topology reveal Alzheimer's proteomic states]]></dc:title>
<dc:publisher>Cold Spring Harbor Laboratory</dc:publisher>
<prism:publicationDate>2026-11-17</prism:publicationDate>
<prism:section></prism:section>
</item>
<item rdf:about="https://www.biorxiv.org/content/10.64898/2026.11.12.766202v1?rss=1">
<title>
<![CDATA[
Mathematical genomics of hidden signatures in NCF1: implications for chronic granulomatous disease 
]]>
</title>
<link>
https://www.biorxiv.org/content/10.64898/2026.11.12.766202v1?rss=1
</link>
<description><![CDATA[
Inborn errors of immunity (IEIs), such as Chronic Granulomatous Disease (CGD), impose a disproportionately severe clinical burden in populations with high rates of consanguinity. CGD is caused by defects in the phagocyte NADPH oxidase complex, frequently due to mutations in the NCF1 gene encoding the p47phox protein. While the functional domains of NCF1 are well characterized, the quantitative architectural constraints governing its structural stability remain elusive. This study aimed to decipher the multidimensional quantitative signatures of NCF1, across diverse organisms in order to understand its conserved spatial sequence architecture and the structural basis of variant severity. We analyzed 41 NCF1 sequences using ten distinct quantitative features, integrating primary sequence metrics, biophysical chemistry, and abstract topological geometry. The synthesis of these features revealed a lack of a unified proximity matrix, demonstrating that NCF1 is defined by feature-dependent, multidimensional individual signatures. Primary sequence metrics exhibited rigid phylogenetic conservation, revealing invariant zero-entropy blocks that correlated strictly with severe CGD-causing mutations, while topological metrics exposed a deep spatial architectural convergence. These multidimensional baselines confirmed that the protein requires a highly specific spatial packing to assemble the NADPH oxidase complex; pathogenic mutations therefore disrupt precise structural signatures rather than globally destroying the protein. Consistent with this, the AlphaFold model places the folded PX and tandem SH3 domains in a high-confidence core and the regulatory C-terminal tail in an exposed, low-confidence region, supporting a tail-mediated autoinhibitory mechanism. Ultimately, this quantitative framework offers a useful perspective on NCF1 architecture and may serve as a supportive tool for evaluating the severity of clinical variants, contributing to precision medicine efforts for NCF1-associated diseases.
]]></description>
<dc:creator><![CDATA[ Hassan, S. S., Hungyo, K., Uversky, V. N. ]]></dc:creator>
<dc:date>2026-11-17</dc:date>
<dc:identifier>doi:10.64898/2026.11.12.766202</dc:identifier>
<dc:title><![CDATA[Mathematical genomics of hidden signatures in NCF1: implications for chronic granulomatous disease]]></dc:title>
<dc:publisher>Cold Spring Harbor Laboratory</dc:publisher>
<prism:publicationDate>2026-11-17</prism:publicationDate>
<prism:section></prism:section>
</item>
<item rdf:about="https://www.biorxiv.org/content/10.64898/2026.11.06.764786v1?rss=1">
<title>
<![CDATA[
Vecmatch: Balancing Covariates in Multiple Groups Using Vector Matching Algorithm and Generalized Propensity Scores 
]]>
</title>
<link>
https://www.biorxiv.org/content/10.64898/2026.11.06.764786v1?rss=1
</link>
<description><![CDATA[
Covariate imbalance is a major source of bias in observational studies. This issue arises when confounding variables (e.g., age, sex) are distributed unevenly across compared groups, potentially influencing estimates of between-group differences in other dependent variables. Matching techniques based on propensity scores are widely used to address covariate imbalance and improve the accuracy of causal inferences. These methods work by calculating propensity scores based on imbalanced covariates and then matching observations with the closest propensity scores across compared groups. However, existing matching methods primarily focus on comparisons between at most two groups, and there is currently no widely implemented approach for matching across more than two treatment groups in popular programming languages. In this paper, we present an R implementation of the vector matching algorithm for matching across multiple treatment groups. The implementation includes the core algorithm itself along with several extensions aimed at improving usability and flexibility in applied research. These features include tools for visualizing covariate distributions, functions for estimating generalized propensity scores, and methods for defining rectangular regions of common support. Users can choose among different matching methods and adjust various parameters to improve covariate balance. To support this process, we provide an optimization workflow based on the random search algorithm that identifies parameter combinations yielding the best matching quality and highest percentage of matched samples. In addition, the package includes functions for assessing post-matching quality, offering a comprehensive and flexible framework for robust causal inference in complex observational settings. Keywords: covariate imbalance, bias reduction, vector matching, generalized propensity scores
]]></description>
<dc:creator><![CDATA[ Kolek, M. F., Vogl, T. ]]></dc:creator>
<dc:date>2026-11-17</dc:date>
<dc:identifier>doi:10.64898/2026.11.06.764786</dc:identifier>
<dc:title><![CDATA[Vecmatch: Balancing Covariates in Multiple Groups Using Vector Matching Algorithm and Generalized Propensity Scores]]></dc:title>
<dc:publisher>Cold Spring Harbor Laboratory</dc:publisher>
<prism:publicationDate>2026-11-17</prism:publicationDate>
<prism:section></prism:section>
</item>
<item rdf:about="https://www.biorxiv.org/content/10.64898/2026.11.07.764304v1?rss=1">
<title>
<![CDATA[
WormIE: an Integrated Explorer for multi-source Caenorhabditis elegans annotations and transcriptomic data interpretation 
]]>
</title>
<link>
https://www.biorxiv.org/content/10.64898/2026.11.07.764304v1?rss=1
</link>
<description><![CDATA[
Interpreting gene lists and expression data from Caenorhabditis elegans experiments usually means writing scripts, searching scattered databases or uploading unpublished data to a remote server. WormIE (Worm Integrated Explorer, https://wormie.bio) is a free, open-source tool for analyzing user-provided data and exploring C. elegans annotations in one browser tab. Starting with gene lists from -omics experiments, genetic screens or other approaches, users can explore what genes have in common and how they interact, test over-representation or ranked-list enrichment in built-in or user-provided gene sets and compare lists by their genes or enriched categories. Of WormIE's nine collections, which include Gene Ontology and CeNGEN neuron types, two are curated sets found in no other worm resource: neuronal genes and synaptic genes. WormIE counts aligned reads from RNA sequencing and tests for differential expression with the same R package as conventional pipelines, then checks the genes for enrichment in those collections. Clicking any gene in a result page opens the gene page gathering seventeen domains from WormBase, the Alliance of Genome Resources and other sources, such as phenotypes, interactions and expression. Everything runs on the user's computer without uploading data and works offline. On simulated data, every analysis agreed with its R counterpart (correlation 0.9999), and in a reanalysis of a germline study of the co-regulator SIN-3, WormIE recovered 99% of the reported differentially expressed genes and all nine WormCat categories. Together, these results show that WormIE reliably links gene lists and expression data to community-curated knowledge without programming or a server.
]]></description>
<dc:creator><![CDATA[ Wijaya, S. P., Sinurat, R. A. O., Richardson, C. E. ]]></dc:creator>
<dc:date>2026-11-17</dc:date>
<dc:identifier>doi:10.64898/2026.11.07.764304</dc:identifier>
<dc:title><![CDATA[WormIE: an Integrated Explorer for multi-source Caenorhabditis elegans annotations and transcriptomic data interpretation]]></dc:title>
<dc:publisher>Cold Spring Harbor Laboratory</dc:publisher>
<prism:publicationDate>2026-11-17</prism:publicationDate>
<prism:section></prism:section>
</item>
<item rdf:about="https://www.biorxiv.org/content/10.64898/2026.11.13.766490v1?rss=1">
<title>
<![CDATA[
Genomic tessellations 
]]>
</title>
<link>
https://www.biorxiv.org/content/10.64898/2026.11.13.766490v1?rss=1
</link>
<description><![CDATA[
Periodicity is ubiquitous in genomic data. Assembled sequences are dense with homopolymers, microsatellites, and large satellite arrays, which contribute to phenotype, variation, and disease. Third-generation sequencing technologies, such as Oxford Nanopore (ONT) and PacBio HiFi, has enabled longer and increasingly accurate reads, but have brought with them systematic error biases in repetitive regions, where insertions and deletions can distort array lengths and mimic or obscure biological variation. This periodicity is not benign: local repetitiveness makes sequence alignment ambiguous and genome assembly error-prone. In sequence analysis, these two manifestations of periodicity are simultaneously at play, each compounding the challenges of the other; a unified representation of repetitive and non-repetitive sequence is necessary to account for their combined effects consistently. We propose a paradigm under which (either repetitive and non-repetitive) DNA is split into its identity (its irreducible components) and its geometry (how its components are realized). A genomic tessellation {tau} is a canonical factorization of any sequence S of length n over the DNA alphabet into a sequence of tiles partitioning S, where each t in {tau}S is described as a quasi-periodic run of a canonical unit identity ut with geometry consisting of span st, phase {varphi}t, and a set of exceptions Xt. We show that, for a prescribed error bound k [&ge;] 1 and fixed maximum unit length P, S can be tessellated in O(n) time such that every tile has a unique identity whose periodic expansion differs from the tiled sequence by at most k edits, and that tiles sharing an identity are related by a bounded number of geometric adjustments. We introduce the tessellated Burrows-Wheeler transform (tBWT), which indexes {tau}S by concatenating, over all tiles in order, the identity ut of each tile. For a tessellated query q of m tiles, the tBWT counts the repeat-agnostic occurrences of q in O(m) time, and for sampling interval r locates its h occurrences in {tau}S in O(m + hr) time. We demonstrate that tessellations avoid the well-known speed-accuracy tradeoff and outperform state-of-the-art long-read mappers in repeat-dense regions in both speed and accuracy. On a 20x simulated set of both ONT and HiFi reads on the HG002 maternal X chromosome with synthetically-expanded repeat arrays, locating and chaining exact tile identity matches obtained from the tBWT is up to 57x and 83x faster, respectively, than the state-of-the-art mappers minimap2 and Winnowmap2, with comparable accuracy. Furthermore, by abstracting away differences in geometry, tessellations of a set of input sequences onto itself (an autotessellation) reveals homology otherwise obscured by array-length variation and local sequence divergence. A geometry-only polishing of an an ONT-only assembly of the HG002 maternal X chromosome with [~]50x PacBio HiFi reads improves assembly QV from 45.4 to 60.6 in under a minute using less than 10GB of memory. Tessellated-read evidence over the autotessellation of this assembly organizes 167 mitochondrial contigs totaling [~]4 Mbp into a single family with a [~]16.5kbp representative, which aligns to the HG002 Q100 mitochondrial genome with 100% identity. Genomic tessellations provide a canonical representation of repetitive and non-repetitive DNA and enable efficient sequence analyses while preserving homology.
]]></description>
<dc:creator><![CDATA[ Ekim, B., Cheng, H. ]]></dc:creator>
<dc:date>2026-11-16</dc:date>
<dc:identifier>doi:10.64898/2026.11.13.766490</dc:identifier>
<dc:title><![CDATA[Genomic tessellations]]></dc:title>
<dc:publisher>Cold Spring Harbor Laboratory</dc:publisher>
<prism:publicationDate>2026-11-16</prism:publicationDate>
<prism:section></prism:section>
</item>
<item rdf:about="https://www.biorxiv.org/content/10.64898/2026.11.13.765983v1?rss=1">
<title>
<![CDATA[
Expanded ME-model analysis capabilities enable genetic recoding, proteostasis, RNA decay, and cofactor essentiality in bacterial models 
]]>
</title>
<link>
https://www.biorxiv.org/content/10.64898/2026.11.13.765983v1?rss=1
</link>
<description><![CDATA[
Genome-scale metabolic models (M-models) are widely used to understand and manipulate microbial metabolism in silico, and are characterized by an extensive availability of automated reconstruction, analysis tools, and simulation software. The scope and accuracy of M-models was greatly improved by the development of metabolic and gene expression models (ME-models) aimed to infer the optimal allocation of individual macromolecules such as proteins, RNA, and enzymatic complexes which constrain metabolic fluxes to their availability and abundance. Although ME-models reconstruction and simulation is currently aided by open-source software such as coralME, COBRAme, ECOLIme, and solveME, they lag behind their M-model counterparts and lack formal analysis tools such as gene essentiality analysis, that would facilitate a broader adoption of ME-models in research. Moreover, the rigid and incomplete formulation of ME-models limit advanced analyses they are theoretically capable of, including the simulation of genomically recoded organisms, protein homeostasis, enzyme cofactor essentiality, and RNA decay. Here, we present coralME v1.3, a vastly improved software version aimed to expand the applicability of ME-models by adding support for translation of non-canonical amino acids, protein and RNA degradation reactions, use of arbitrary coupling coefficients and parameter values, and formalized gene and cofactor essentiality analyses. This work expands, formalizes, and streamlines ME-model reconstruction and analyses for applied research.
]]></description>
<dc:creator><![CDATA[ Santibanez-Palominos, R. A., Tibocha-Bonilla, J. D., Tec-Campos, D., Neal, M. J., Zengler, K. ]]></dc:creator>
<dc:date>2026-11-16</dc:date>
<dc:identifier>doi:10.64898/2026.11.13.765983</dc:identifier>
<dc:title><![CDATA[Expanded ME-model analysis capabilities enable genetic recoding, proteostasis, RNA decay, and cofactor essentiality in bacterial models]]></dc:title>
<dc:publisher>Cold Spring Harbor Laboratory</dc:publisher>
<prism:publicationDate>2026-11-16</prism:publicationDate>
<prism:section></prism:section>
</item>
<item rdf:about="https://www.biorxiv.org/content/10.64898/2026.11.13.765998v1?rss=1">
<title>
<![CDATA[
Paired multiomic profiling defines the healthy human synovium infrapatellar fat pad stromal continuum 
]]>
</title>
<link>
https://www.biorxiv.org/content/10.64898/2026.11.13.765998v1?rss=1
</link>
<description><![CDATA[
The synovium and infrapatellar fat pad (IPFP) are anatomically contiguous tissues implicated in osteoarthritis, yet their cellular and molecular organization in the healthy human knee remains poorly defined. We combined bulk RNA sequencing with paired single-nucleus transcriptomic and chromatin accessibility profiling to establish a quantitative multiomic reference of healthy human synovium and IPFP. Analysis of 54,716 nuclei from six cadaveric donors identified 23 synovial and 20 IPFP cell populations and indicated that the two tissues share one transcriptionally continuous, largely quiescent stromal compartment. Stromal populations shared core transcriptional programs but differed primarily in extracellular matrix output, with transcripts of inhibitory regulators predominating over those of their functional opponents. TIMP3 exceeded MMP14 in 42 of 43 single-nucleus populations across both tissues and by 13- to 39-fold in bulk tissue. Chromatin accessibility primarily distinguished stromal lineages and resolved transcriptionally defined subtypes, separating synovial lining from sublining fibroblasts and adipocytes from fibroblasts. Fibroblast and adipocyte populations occupied a common transcriptional trajectory accompanied by declining expression of organelle-turnover programs and little proliferation, with no dividing adipocytes and fewer than 0.05% of fibroblasts in the cell cycle. Despite this shared organization, the IPFP exhibited greater inferred intercellular signaling, attributable in part to BMP5, which was not detected in synovium. These findings describe healthy synovium and IPFP as regionally specialized parts of a transcriptionally continuous stromal compartment and provide a quantitative reference for identifying cellular and regulatory changes in joint disease.
]]></description>
<dc:creator><![CDATA[ Swahn, H., Gomez, I. L., Janssen, L., Kenvisay, C., Olmer, M., Maerz, T., Miller, R. E., Malfait, A.-M., Orange, D. E., Mehta, B., Otero, M., Price, T. J., Lotz, M. K. ]]></dc:creator>
<dc:date>2026-11-16</dc:date>
<dc:identifier>doi:10.64898/2026.11.13.765998</dc:identifier>
<dc:title><![CDATA[Paired multiomic profiling defines the healthy human synovium infrapatellar fat pad stromal continuum]]></dc:title>
<dc:publisher>Cold Spring Harbor Laboratory</dc:publisher>
<prism:publicationDate>2026-11-16</prism:publicationDate>
<prism:section></prism:section>
</item>
<item rdf:about="https://www.biorxiv.org/content/10.64898/2026.11.10.765731v1?rss=1">
<title>
<![CDATA[
Revealing the transcriptomic organization of extrasomatic interfaces using SCOUBI 
]]>
</title>
<link>
https://www.biorxiv.org/content/10.64898/2026.11.10.765731v1?rss=1
</link>
<description><![CDATA[
Spatial transcriptomics (ST) generates substantial transcript signal outside segmented cell bodies, which is generally discarded by analysis tools. This is especially problematic for brain ST data, as neuronal mRNAs can be localized far from somata - in axons, dendrites, and synaptic compartments - and up to half of the transcripts in ST data may lie outside cell bodies. Here, we present SCOUBI, a computational framework that identifies axon-dendrite interfaces from extrasomatic RNA in ST data, revealing a hitherto overlooked layer of molecular function in neural tissues. It first partitions extrasomatic space into axon- and dendrite-enriched compartments based on known marker genes, using a neural network trained to maximize marker specificity and inter-compartment communication signals. It then defines interfaces as loci where predicted axonal and dendritic compartments converge and exhibit ligand-receptor colocalization. Applied to mouse brain, SCOUBI creates maps of axonal and dendritic compartments and their interfaces that align with known spatial patterns of neurite and synaptic distribution. It identifies transcriptomic differences between interfaces and the surrounding extrasomatic space, revealing region-specific signatures of neuronal signaling and synaptic plasticity. By comparing interface transcript enrichment with local somatic expression, SCOUBI identifies genes that preferentially localize at sites of intercellular communication rather than simply reflecting nearby cell bodies. SCOUBI reveals interface-associated transcripts beyond established synaptic genes, with molecular signatures reproducible across independent samples and imaging platforms. In summary, SCOUBI transforms a substantial and often-discarded aspect of ST data into an interpretable readout of subcellular organization and neural signaling.
]]></description>
<dc:creator><![CDATA[ Aggarwal, B., Jain, S., Sinha, S. ]]></dc:creator>
<dc:date>2026-11-16</dc:date>
<dc:identifier>doi:10.64898/2026.11.10.765731</dc:identifier>
<dc:title><![CDATA[Revealing the transcriptomic organization of extrasomatic interfaces using SCOUBI]]></dc:title>
<dc:publisher>Cold Spring Harbor Laboratory</dc:publisher>
<prism:publicationDate>2026-11-16</prism:publicationDate>
<prism:section></prism:section>
</item>
<item rdf:about="https://www.biorxiv.org/content/10.64898/2026.11.13.766406v1?rss=1">
<title>
<![CDATA[
ProtCOVER: Family-Aware Conformal Prediction for Variant Effect Predictors on Unseen Protein Families 
]]>
</title>
<link>
https://www.biorxiv.org/content/10.64898/2026.11.13.766406v1?rss=1
</link>
<description><![CDATA[
Protein language models and related variant effect predictors (VEPs) increasingly classify missense variants in proteins evolutionarily distant from those used for calibration. Split conformal prediction over variants is valid for its own target, a random variant dominated by well-studied families, but deployment needs coverage for a variant of a new family. On ProteinGym clinical substitutions with five zero-shot predictors, Mondrian conformal comes within 0.005 of that target for four predictors, yet family-weighted pathogenic coverage is only 0.823-0.873 at a 0.90 target for all five, without any shift. We present ProtCOVER, a conformal procedure that treats homology families, rather than individual variants, as the exchangeable unit. ProtCOVER (i) calibrates at the family level over label-free homology clusters, giving marginal coverage for a new family under family exchangeability; (ii) adapts to descriptor shift by calibrating conditionally on an estimated density ratio of label-free descriptors (sequence similarity, alignment depth and embedding distance; a variant reweights by the ratio instead), with finite-sample family-weighted coverage, valid as computed, whenever the true ratio is a nonnegative affine function of the estimate; and (iii) can abstain at low effective calibration support, which proved no better than random abstention. Under a shift of its descriptors, ProtCOVER's mean worse-class shortfall below the target is 0.000, against 0.010 for group-weighted and 0.029 for hierarchical conformal prediction, with larger sets (about equal at matched coverage); on orphan families and held-out embedding regions it matches group weighting (0.002 against 0.003). In evaluations following pre-registered protocols, every shift-aware method met the target on new ClinVar variants of benchmark proteins, which ESM-1b separated better; on 12,225 proteins outside the benchmark, every method fell short for pathogenic variants with either of two predictors, ProtCOVER and a variant conditioned on the descriptors the least, within the 0.01 margin of group weighting.
]]></description>
<dc:creator><![CDATA[ Fahmy, W., Krikun, E., Alkhateeb, A. ]]></dc:creator>
<dc:date>2026-11-16</dc:date>
<dc:identifier>doi:10.64898/2026.11.13.766406</dc:identifier>
<dc:title><![CDATA[ProtCOVER: Family-Aware Conformal Prediction for Variant Effect Predictors on Unseen Protein Families]]></dc:title>
<dc:publisher>Cold Spring Harbor Laboratory</dc:publisher>
<prism:publicationDate>2026-11-16</prism:publicationDate>
<prism:section></prism:section>
</item>
<item rdf:about="https://www.biorxiv.org/content/10.64898/2026.11.13.766391v1?rss=1">
<title>
<![CDATA[
ProtFlap: Structure-based Deep Learning Prediction of Protein Flexibility from Atomic Energy Networks 
]]>
</title>
<link>
https://www.biorxiv.org/content/10.64898/2026.11.13.766391v1?rss=1
</link>
<description><![CDATA[
Protein flexibility is a key determinant of molecular recognition, catalysis, and regulation, yet remains difficult to characterise at scale. Molecular dynamics (MD) simulations provide detailed descriptions of time-dependent dynamics but are computationally expensive. Although recent AI-based structure prediction methods have substantially improved access to high-quality static protein structures, they inherently lack information regarding the dynamic fluctuations essential for functional understanding. There is therefore a need for scalable, structure-based computational approaches that can infer intrinsic protein flexibility without explicit dynamical sampling. We present ProtFlap, a graph neural network framework for predicting relative residue-level flexibility directly from static protein structures. ProtFlap represents proteins as residue interaction graphs and integrates empirical atomic energy features extracted from a single energy-minimised snapshot with structural and sequence-derived descriptors. Central to our approach is the hypothesis that the local atomic energy network encoded in a frozen protein structure contains sufficient information to estimate structural flexibility. Trained using atomic mobility labels, quantified via normalised root-mean-square fluctuation (RMSF) derived from short MD simulations, ProtFlap achieves robust and accurate flexibility prediction across diverse protein domains. Feature attribution analysis reveals that van der Waals-related energy terms are the dominant contributors to the prediction.
]]></description>
<dc:creator><![CDATA[ Lubna, H., Choi, Y. ]]></dc:creator>
<dc:date>2026-11-16</dc:date>
<dc:identifier>doi:10.64898/2026.11.13.766391</dc:identifier>
<dc:title><![CDATA[ProtFlap: Structure-based Deep Learning Prediction of Protein Flexibility from Atomic Energy Networks]]></dc:title>
<dc:publisher>Cold Spring Harbor Laboratory</dc:publisher>
<prism:publicationDate>2026-11-16</prism:publicationDate>
<prism:section></prism:section>
</item>
<item rdf:about="https://www.biorxiv.org/content/10.64898/2026.11.13.766436v1?rss=1">
<title>
<![CDATA[
Evaluating Language Models as Orchestrators in a Tool-First Biomedical Discovery Platform 
]]>
</title>
<link>
https://www.biorxiv.org/content/10.64898/2026.11.13.766436v1?rss=1
</link>
<description><![CDATA[
Large language models (LLMs) increasingly orchestrate agentic biomedical discovery systems, but cloud deployment raises concerns about inference cost, data privacy, and governance. Small language models (SLMs) that run on-premises are an alternative, provided they orchestrate scientific tools about as well as larger models. We built a tool-first, multi-agent system for scientific discovery in which an orchestrator coordinates six specialized tool systems: differential expression, literature retrieval, pathway enrichment, network-based gene prioritization and knowledge-graph analysis, perturbation signature matching, and structural/druggability assessment. In a tool-first design, specialized tools supply the domain knowledge and computation, so the model's main task is coordination, e.g. choosing tools, passing the right parameters, calling them in the right order, and carrying results across turns. With the tools, prompts, workflow, and benchmark held constant, we compared three orchestrators, Gemma 4 (26B, open-weight and locally deployable), Gemini 2.5 Flash, and Claude Sonnet 5. Evaluation used a custom benchmark of 104 single-turn and multi-turn tests, an LLM-as-judge protocol (groundedness and evidence validity), tool-use accuracy and efficiency metrics, and blind expert review. On single-turn tasks, the three models performed comparably, and reviewers did not rate the SLM below the larger models. In multi-turn tasks, Claude Sonnet 5 showed marginally higher evidence validity, Gemma 4 was weakest when parameters had to be assembled across turns, and all three models struggled with tool ordering on memory-dependent tasks. These differences were illustrated with case studies. Finally, we propose strategies to help SLMs address the multi-turn limitations and support cost- and privacy-aware deployment.
]]></description>
<dc:creator><![CDATA[ Sapireddy, V., Chakrabarty, B., Bung, N., Roy, A. ]]></dc:creator>
<dc:date>2026-11-16</dc:date>
<dc:identifier>doi:10.64898/2026.11.13.766436</dc:identifier>
<dc:title><![CDATA[Evaluating Language Models as Orchestrators in a Tool-First Biomedical Discovery Platform]]></dc:title>
<dc:publisher>Cold Spring Harbor Laboratory</dc:publisher>
<prism:publicationDate>2026-11-16</prism:publicationDate>
<prism:section></prism:section>
</item>
<item rdf:about="https://www.biorxiv.org/content/10.64898/2026.11.13.766356v1?rss=1">
<title>
<![CDATA[
A conserved matrix-centred core links fibrosis across human organs and mouse models 
]]>
</title>
<link>
https://www.biorxiv.org/content/10.64898/2026.11.13.766356v1?rss=1
</link>
<description><![CDATA[
Fibrosis develops in many organs through different initiating injuries, yet repeatedly engages inflammation, stromal activation and extracellular-matrix remodelling. Whether these recurring features represent a conserved disease programme, and whether commonly used mouse models reproduce it, remain unresolved. We developed ScarX, an R/Shiny workflow that analyses organ-resolved bulk RNA sequencing, identifies directional overlap across tissues, and ranks human-mouse conserved signals while retaining organ-restricted results. Public RNA-sequencing datasets representing fibrotic or fibrosis-associated disease in human heart, kidney, liver and lung were compared with matched mouse injury models. Differential expression was estimated independently within each organ using DESeq2, followed by cross-organ intersection, functional enrichment, matched-organ effect-profile comparison and a human-first conserved modulation (HCM) score that gives greatest weight to the breadth and magnitude of human disease change, adds support for same-direction mouse effects and penalises discordance. In human disease, 202 genes increased and 35 decreased in the same direction across all four organs. Mouse models contained 409 pan-organ increases and 46 decreases. After gene-symbol harmonisation, 25 gene-direction signatures were conserved across all four organs in both species. These included AEBP1, CDH11, COL1A1, COL1A2, COL3A1, COL5A2, COL6A2, EFEMP2, EMILIN1, LOXL1, LOXL2, TGFB2 and VIM. Matched human-mouse effect profiles varied considerably: Spearman rho was 0.42 for kidney, 0.31 for heart, 0.16 for liver and -0.06 for lung. Despite this uneven gene-level fidelity, 62 same-direction pathways were shared across species, led by extracellular-matrix organisation, collagen fibril organisation, cell-substrate adhesion, wound healing and ECM-receptor interaction. The HCM ranking concentrated structural matrix genes near the top but also identified cell-surface, secreted and regulatory candidates. Among the top 500 candidates, 376 carried at least one local tractability annotation. ScarX therefore separates a compact conserved matrix-remodelling core from a much larger organ- and species-dependent response. The results provide a transparent starting point for target and biomarker selection, while also showing why expression conservation alone cannot establish therapeutic dependence.
]]></description>
<dc:creator><![CDATA[ Di Carmine, S., Cunningham, R., Mylonas, K. J. ]]></dc:creator>
<dc:date>2026-11-16</dc:date>
<dc:identifier>doi:10.64898/2026.11.13.766356</dc:identifier>
<dc:title><![CDATA[A conserved matrix-centred core links fibrosis across human organs and mouse models]]></dc:title>
<dc:publisher>Cold Spring Harbor Laboratory</dc:publisher>
<prism:publicationDate>2026-11-16</prism:publicationDate>
<prism:section></prism:section>
</item>
<item rdf:about="https://www.biorxiv.org/content/10.64898/2026.11.13.766422v1?rss=1">
<title>
<![CDATA[
Spatial causal discovery reveals novel mechanisms of cellular senescence in human liver 
]]>
</title>
<link>
https://www.biorxiv.org/content/10.64898/2026.11.13.766422v1?rss=1
</link>
<description><![CDATA[
The complex network of mechanisms that surround cellular senescence, such as its precise causes and effects, are not entirely understood at present. Senescence occurs when a cell permanently ceases dividing, halts apoptosis, stops performing its normal functions, and begins secreting chemicals that can induce other cells in its environment to also become senescent. Senescent cells are present in both healthy and unhealthy tissue, but their abundance is positively correlated with disease and age. A better understanding of senescence could lead to healthier aging, improved recovery from chemotherapy, or increased lifespan. This study used novel machine learning algorithms - with guaranteed properties of correctness under broad assumptions - to investigate senescence in human liver tissue by analyzing Visium spatial transcriptomics data. This produced data-driven models of the causes and effects of senescence, both locally and from nearby tissue. This report includes an assessment of whether the data meet key requirements for causal analysis, the transcripts identified by causal analyses as causes and effects of senescence marker genes, and pathway analyses of those gene sets.
]]></description>
<dc:creator><![CDATA[ Kummerfeld, E., Wang, Y., Williams, L., Peters, S. T., Schmidt, E. L., DuFresne-To, M., Robbins, P., Anderson, M., Ikramuddin, S., Adeyi, O., Haynes, D., Johnson, S., Wang, J., Temiz, N. A., Niedernhofer, L., Nelson, A., Aliferis, C. ]]></dc:creator>
<dc:date>2026-11-16</dc:date>
<dc:identifier>doi:10.64898/2026.11.13.766422</dc:identifier>
<dc:title><![CDATA[Spatial causal discovery reveals novel mechanisms of cellular senescence in human liver]]></dc:title>
<dc:publisher>Cold Spring Harbor Laboratory</dc:publisher>
<prism:publicationDate>2026-11-16</prism:publicationDate>
<prism:section></prism:section>
</item>
<item rdf:about="https://www.biorxiv.org/content/10.64898/2026.11.13.766327v1?rss=1">
<title>
<![CDATA[
Bounded by compartment, composition and time: efferocytic transcriptional readouts after ischaemic stroke 
]]>
</title>
<link>
https://www.biorxiv.org/content/10.64898/2026.11.13.766327v1?rss=1
</link>
<description><![CDATA[
Background. Efferocytosis after ischaemic stroke is usually inferred from bulk tissue transcripts and condensed into one score, which cannot separate a tissue containing more phagocytes from one whose phagocytes are more active; the gene sets behind such scores also mix clearance machinery with the inflammatory output that follows. Methods. We built a decoupled framework scoring apoptotic demand (DEMAND), clearance machinery (CAPACITY), immunogenic output (OUTCOME) and a brake. We assembled 38 datasets across brain, blood, PBMC and sorted parenchyma, admitted a score only where at least 75% of a program's genes were detected, and tested every claim against expression-matched nulls, lineage proxies and cell-level constraints. Because clearance consumes ATP, supply and demand were scored on one scale under one decision rule. Results. After adjustment for five leukocyte lineages the CAPACITY difference survived in one of three cohorts (+0.473, 95% CI +0.350 to +0.588), fell to +0.014 in the second and was unidentifiable in the third; the unadjusted differences (+0.613 and +0.374) exceeded expression- and detection-matched nulls, so the signal is real, but its compositional share differs by cohort. In matched PBMCs the point estimate was negative and not distinguishable from zero. In the only sorted human parenchyma series, non-microglial myeloid cells made up 0.371 of CD45-positive cells against 0.199 in controls, while every microglial layer fell below its expression-matched floor: it moves the composition boundary without testing capacity. In rodent brain it rose with injury: day 3 in microglia, day 7 in macrophages. Energy supply cleared both gates in eleven of eighteen contrasts and demand in eight, concentrated in entering cells (oxidative phosphorylation +1.512 in infiltrating macrophages at day 3). The two scores ranked samples almost identically ({rho} 0.59 to 0.95), and the one significant gap vanished under the same adjustment. Conclusions. Efferocytic transcripts are a measurement bounded by compartment, composition and time, not a flux assay or a biomarker. Energy supply rises with the work of clearance in the cells that carry it, and can rise in cells that do not engulf, but the two scores move closely enough that no transcript design can show one limiting the other. https://www.ebi.ac.uk/pride/archive/projects/PXD012151 https://www.ebi.ac.uk/pride/archive/projects/PXD016538 https://www.ebi.ac.uk/pride/archive/projects/PXD022850 https://www.ebi.ac.uk/pride/archive/projects/PXD025077 https://www.ebi.ac.uk/pride/archive/projects/PXD055588 https://www.ncbi.nlm.nih.gov/geo/query/acc.cgi?acc=GSE105132 https://www.ncbi.nlm.nih.gov/geo/query/acc.cgi?acc=GSE137482 https://www.ncbi.nlm.nih.gov/geo/query/acc.cgi?acc=GSE137595 https://www.ncbi.nlm.nih.gov/geo/query/acc.cgi?acc=GSE141972 https://www.ncbi.nlm.nih.gov/geo/query/acc.cgi?acc=GSE154396 https://www.ncbi.nlm.nih.gov/geo/query/acc.cgi?acc=GSE162955 https://www.ncbi.nlm.nih.gov/geo/query/acc.cgi?acc=GSE163654 https://www.ncbi.nlm.nih.gov/geo/query/acc.cgi?acc=GSE16561 https://www.ncbi.nlm.nih.gov/geo/query/acc.cgi?acc=GSE171169 https://www.ncbi.nlm.nih.gov/geo/query/acc.cgi?acc=GSE174574 https://www.ncbi.nlm.nih.gov/geo/query/acc.cgi?acc=GSE202518 https://www.ncbi.nlm.nih.gov/geo/query/acc.cgi?acc=GSE22255 https://www.ncbi.nlm.nih.gov/geo/query/acc.cgi?acc=GSE225948 https://www.ncbi.nlm.nih.gov/geo/query/acc.cgi?acc=GSE227651 https://www.ncbi.nlm.nih.gov/geo/query/acc.cgi?acc=GSE23160 https://www.ncbi.nlm.nih.gov/geo/query/acc.cgi?acc=GSE233815 https://www.ncbi.nlm.nih.gov/geo/query/acc.cgi?acc=GSE23651 https://www.ncbi.nlm.nih.gov/geo/query/acc.cgi?acc=GSE250245 https://www.ncbi.nlm.nih.gov/geo/query/acc.cgi?acc=GSE267023 https://www.ncbi.nlm.nih.gov/geo/query/acc.cgi?acc=GSE318800 https://www.ncbi.nlm.nih.gov/geo/query/acc.cgi?acc=GSE319238 https://www.ncbi.nlm.nih.gov/geo/query/acc.cgi?acc=GSE331114 https://www.ncbi.nlm.nih.gov/geo/query/acc.cgi?acc=GSE332910 https://www.ncbi.nlm.nih.gov/geo/query/acc.cgi?acc=GSE337578 https://www.ncbi.nlm.nih.gov/geo/query/acc.cgi?acc=GSE345807 https://www.ncbi.nlm.nih.gov/geo/query/acc.cgi?acc=GSE37587 https://www.ncbi.nlm.nih.gov/geo/query/acc.cgi?acc=GSE58294 https://www.ncbi.nlm.nih.gov/geo/query/acc.cgi?acc=GSE86941
]]></description>
<dc:creator><![CDATA[ Ding, S., Liu, Y., Zeng, Y., Huang, Z., Yuan, Z. ]]></dc:creator>
<dc:date>2026-11-16</dc:date>
<dc:identifier>doi:10.64898/2026.11.13.766327</dc:identifier>
<dc:title><![CDATA[Bounded by compartment, composition and time: efferocytic transcriptional readouts after ischaemic stroke]]></dc:title>
<dc:publisher>Cold Spring Harbor Laboratory</dc:publisher>
<prism:publicationDate>2026-11-16</prism:publicationDate>
<prism:section></prism:section>
</item>
<item rdf:about="https://www.biorxiv.org/content/10.64898/2026.11.13.766344v1?rss=1">
<title>
<![CDATA[
Know Thy Neighbour: Benchmarking Protein Language Model Embedding-Based Annotation Reveals Anisotropy Substantially Impacts Retrieval 
]]>
</title>
<link>
https://www.biorxiv.org/content/10.64898/2026.11.13.766344v1?rss=1
</link>
<description><![CDATA[
Functional annotation using protein language models (`pLMs') usually proceeds by transferring the function of a query's nearest neighbour in a reference database. A major drawback is that embedding-based annotation transfer has no null model, i.e. no accepted way to set a cosine-similarity threshold to control false positives. This contrasts with traditional bioinformatics methods, including sequence- and structure-based alignment tools, which rank hits with E-values derived from explicit null models. Here, we benchmark final-layer, mean-pooled embeddings from twelve models against MMseqs2 and Foldseek for remote-homology retrieval on curated datasets based on Swiss-Prot and SCOPe, fixing per-model thresholds at a 1% false-positive rate and probing the score with a ladder of synthetic decoys. On Swiss-Prot, pLMs without retrieval training recover at most 15% of remote homologs at this operating point, against 85% for MMseqs2, and for the ESM-2 and ESM-C families shuffled decoys have higher cosine similarities to their nearest database hits than real remote homologs do. These two pLM families exhibit extreme anisotropy, using fewer than 15 effective dimensions out of 640-2560, and low effective dimensionality is strongly associated with poor performance across models. Applying ZCA whitening, fit on the reference database, to the embeddings sharply raises threshold-free Top-1 superfamily retrieval, e.g. for ESM-2 650M from 59% to 88% on SCOPe, and generally improves remote homology sensitivity at a fixed false-positive rate. As correcting embedding anisotropy requires only the reference database and is compute cheap, it should be considered and tested for any pLM embedding-based annotation task.
]]></description>
<dc:creator><![CDATA[ Bouras, G., Edwards, R. A. ]]></dc:creator>
<dc:date>2026-11-16</dc:date>
<dc:identifier>doi:10.64898/2026.11.13.766344</dc:identifier>
<dc:title><![CDATA[Know Thy Neighbour: Benchmarking Protein Language Model Embedding-Based Annotation Reveals Anisotropy Substantially Impacts Retrieval]]></dc:title>
<dc:publisher>Cold Spring Harbor Laboratory</dc:publisher>
<prism:publicationDate>2026-11-16</prism:publicationDate>
<prism:section></prism:section>
</item>
</rdf:RDF>
