bioRxiv Daily Picks

[bioRxiv] Sung et al. present FateDriver, a weakly supervised dual-view graph neural network that jointly represents cells and genes from single-cell transcriptomes, treats experimentally delivered factors as sparse gene-level labels, and ranks remaining genes for a specified source-to-target transition under a positive-unlabeled formulation, according to the preprint. The authors said FateDriver ranked first or second among six methods in recovering externally curated reference genes across two OSKM datasets and one GMT dataset, exceeded a supervision-matched non-graph model despite weaker co-expression with the delivered factors, and achieved the highest recovery of protein targets of pluripotency-inducing chemical cocktails in both OSKM datasets. Holding either the source or the endpoint fixed yielded distinct candidate programs, including a DDX58-interferon axis associated with cells lacking endogenous pluripotency markers, they said. Applied to candidate source-target pairs, the framework prioritized distinct candidate origins for two in vivo partially reprogrammed intestinal populations.

[bioRxiv] A bioRxiv preprint by Lee et al. combines single-molecule long-read footprinting with deep learning to study how protein-DNA interactions are coordinated within different types of cis-regulatory elements in mouse embryonic stem cells. The authors first quantify how often pairs of positions are simultaneously accessible on individual molecules, then train ChAI-Net, a neural network that predicts these locus-specific chromatin co-accessibility maps from the underlying DNA sequence. Through model interpretation and in silico experiments, the authors report that ChAI-Net learns unique combinations of sequence motifs underlying co-accessibility within enhancers, promoters and CTCF-bound sites; many of these motifs are missed when modeling average accessibility, suggesting a specific role in coordinated binding. According to the authors, the model's regulatory grammar also highlights the importance of transcription factor motif orientation, spacing, positioning and co-occurrence in driving co-accessibility.

[bioRxiv] A preprint posted to bioRxiv describes an AI system for automated de novo structure elucidation that pairs a molecular generative model with a scientific reasoning agent. The generative component, ChefNMR2, is an atomic flow-matching model that predicts small-molecule structures from molecular formulae and one-dimensional 1H and 13C NMR spectra; ChefNMR2-Agent adds large language models to reason over and refine candidate structures. According to the authors, ChefNMR2 substantially outperforms existing tools across diverse benchmarks, offers calibrated confidence estimates and runs two to three orders of magnitude faster than frontier LLMs. The team said it used ChefNMR2-Agent to elucidate the structures of four novel natural products, matching human-chemist accuracy in a fraction of the time. The preprint has not been peer-reviewed.

[bioRxiv] Ho, Xing and Ma present EMBER, a gene-centric foundation model that learns context-dependent functional representations of genes, according to a bioRxiv preprint. EMBER combines interaction-guided contrastive learning with masked expression reconstruction during pretraining on a single-cell corpus curated for diversity and balanced sampling across cell types and experimental batches. The model decouples gene representation learning from task-specific prediction: the pretrained encoder remains frozen while lightweight probes combine contextualized gene embeddings to predict single-gene perturbation responses, digenic interactions and multi-target drug responses. The authors said EMBER improved cross-context perturbation-response prediction and genetic-interaction ranking relative to evaluated baselines, and prioritized SCFD1, a mediator of autophagosome-lysosome fusion, as a candidate dependency during mTOR inhibition. The findings are from a preprint that has not been peer-reviewed.

[bioRxiv] Riffle, Zelter and MacCoss present the Findings Workflow, a free, open-source Claude Code plugin designed to retain the flexibility of conversational LLM-assisted analysis while placing it within a structured, verifiable and reproducible environment, according to the preprint. The authors said the workflow uses staged, gated progression to ensure the science, metadata and data are understood and tested data loaders are built before downstream analysis begins. Specialized subagents perform statistical analysis, programming, visualization, research, reporting and validation, with independent review incorporated throughout, they said. The plugin also provides vetted analysis templates, conventions and methodological guidance. Scientists remain actively involved, exploring data and directing analysis through natural-language conversation, while the workflow supplies structure, verification and persistent state, according to the authors. Findings are recorded automatically together with the data, code, methods, figures and provenance needed to reproduce them. The authors demonstrate the workflow with a worked example reproducing central findings from a published plasma extracellular vesicle proteomics dataset examining Alzheimer's disease (Mag-Net). The preprint has not been peer-reviewed.

[bioRxiv] A bioRxiv preprint by Stamenov et al. benchmarks graph neural networks, frozen pretrained models (ChemBERTa, MolCLR) and fingerprint baselines across six Therapeutics Data Commons ADMET endpoints under scaffold-split evaluation, giving every model family a matched 50-trial, validation-selected tuning budget and reporting the median over ten seeds. According to the authors, under this fair comparison protocol no model family dominates, and differences between families fall within the seed-to-seed range on most endpoints. The GNN is clearly best only on Caco-2 permeability (R² about 0.48), and a controlled ablation shows this advantage comes entirely from auxiliary physicochemical descriptors rather than the graph representation; Morgan fingerprints are competitive on hERG (AUC about 0.77) and give the best median on both clearance endpoints, while MolCLR is best on Tox21. The authors also report that for hepatocyte clearance and half-life all models yield near-zero or negative R², that none of seven hyperparameter-optimization strategies consistently outperforms random search at this budget, and that selecting the optimizer using test-set performance or reporting only single best-trial runs can substantially inflate a favored model's apparent advantage.

[bioRxiv] A preprint using longitudinal metagenomic data from 106 individuals in HMP2/iHMP inferred individual-specific microbial association networks and used them to train and interpret graph neural network (GNN) models, comparing them with abundance-based models such as Random Forest (RF). According to the authors, RF achieved higher predictive performance (AUPRC ~0.59 vs 0.446), but interpreting the GNNs showed that their predictions relied on a densely connected module of co-abundant Bacteroidales species; unlike the taxa prioritized by RF, these species were not differentially abundant between conditions. The authors said that in an independent cohort this module carried most of the predictive signal, and removing its associations reduced performance to near chance. They concluded that coordinated association structure, rather than abundance shifts alone, is a feature of the IBD gut microbiome, and that interpreting GNNs can uncover such patterns to complement abundance-based analyses.

[bioRxiv] A bioRxiv preprint by Khan, DeGiorgio and co-authors introduces SHADES, a deep learning framework that uses information on sequence differences and their frequencies at multiple spatial resolutions to capture evolutionary signatures across genomic scales. According to the authors, SHADES can distinguish among neutrality, selective sweeps, balancing selection and adaptive introgression, and outperformed several contemporary methods evaluated for classification accuracy and power. The framework remained robust to a number of confounding factors, with only modest changes in classification performance, the authors said. Applied to empirical haplotypic variation from European humans, SHADES recovered previously reported signals for selective sweeps, balancing selection and adaptive introgression while also identifying novel genes broadly involved in immune defense, inflammatory regulation and epithelial barrier biology. The work is a preprint and has not been peer-reviewed.

[bioRxiv] A preprint study evaluated zero-shot embeddings from Geneformer, scGPT and UCE against PCA and scVI in developmental atlases of Dictyostelium discoideum and Caenorhabditis elegans, organisms separated from humans by 1.2 billion and 600 million years. According to the study, without retraining the models recovered developmental order and preserved organism-specific cell identities. UCE, which encodes genes by protein sequence, could use 97% of measured genes, whereas Geneformer and scGPT used under one-fifth after ortholog mapping. The authors said UCE performed most consistently, often exceeding species-native baselines; restricted to ortholog-mappable genes, it still scored above vocabulary-based models on five of six annotation tasks, some narrowly. The preprint concludes that foundation models can yield informative cell embeddings across more than a billion years of evolution, and that protein-sequence gene representations make these embeddings more complete and often more accurate.

[bioRxiv] A bioRxiv preprint by Uckelmann et al. presents LPT (little protein tiger), a framework that fixes the orchestration of LLM stages, each running a bounded agentic loop with custom system prompts and tool access, to enable de novo protein binder design from a single natural language prompt. The stages can query a curated literature database containing about 14,000 papers digested into key findings. According to the paper, an LLM with database access answered 41 of 48 test questions drawn from database key findings correctly, while a model without access answered none. In three end-to-end design campaigns initiated from a single prompt, LPT autonomously identified design intent, selected targets, structures and hotspot residues, and designed candidate binders to disrupt CALCRL/RAMP1, PD-1/PD-L1 and PVR-TIGIT interactions. Cross-validation with multiple folding models indicated a high in silico success rate at minimal token consumption and low computational cost. The findings have not been peer reviewed.

[bioRxiv] A bioRxiv preprint by Bose et al. describes a multi-layered graph convolutional network (GCN) developed to identify tumor microenvironment (TME)-guided subtypes of pancreatic ductal adenocarcinoma (PDAC). The authors first built a TME reference atlas from single-cell RNA-seq data covering immune and non-immune cell types, then used in silico cytometry to estimate cell-type proportions in bulk tumor RNA-seq samples from Cancer Genome Atlas PDAC patients. They integrated patient-specific cellular proportions with multi-omics data, including gene expression, DNA methylation, copy number variation and clinical data, in the GCN model, and predicted PDAC subtypes by unsupervised clustering of the GCN-based embedding matrices. According to the authors, the method identified seven PDAC subtypes, which they grouped into two ecological tumor families; the subtypes showed distinct immune infiltration, stromal activity and malignant epithelial cell abundance, differential expression analysis revealed subtype-specific molecular programs, and survival analysis showed significant differences in overall survival across ecological families and subtypes. The authors said the approach provides a biologically informed framework for patient stratification with potential implications for precision oncology; the findings are preprint results and have not been peer reviewed.

[bioRxiv] Yang et al. present Creator 1.0, a 2-billion-parameter plant genome foundation model pretrained on 1,000 plant genomes using phylogeny-guided stagewise pretraining (PGSP), according to the preprint. Training comprised 83 consecutive stages and produced 52 cumulative checkpoints associated with distinct plant orders. According to the authors, on GENEB the stage 83 checkpoint ranked first among 46 models across 10 plant-related tasks with a mean Matthews correlation coefficient (MCC) of 0.565, and fourth across all 100 tasks. Evaluation of the 52 checkpoints on six plant lncRNA tasks identified different optimal stages under full-data, 10-shot and 1-shot settings, corresponding to stages 71, 72 and 29, respectively. The authors said checkpoint performance varied across tasks, without consistent advantages for corresponding plant orders. The preprint has not been peer-reviewed.

[bioRxiv] A bioRxiv preprint by Echavarria Galindo et al. introduces DAPE-PD (Diffusion Adaptation of Pretrained Evo for Promoter Design), which adapts the autoregressive Evo model into a bidirectional masked diffusion model without foundation-scale retraining. The method combines pretrained genomic representations with iterative unmasking to generate regulatory DNA sequences. After genus-specific fine-tuning on genome-derived intergenic sequences, DAPE-PD generated sequences that retained genus-associated compositional features despite low similarity to the fine-tuning datasets, according to the authors. Experimental validation identified functional de novo promoters, and in silico prioritization recovered active designs from small experimental candidate sets across three non-model gut commensals. The authors said the framework connects genomic pretraining with diffusion-based generation, extending genomic foundation models from sequence understanding toward functional cis-regulatory DNA engineering in living cells.

[bioRxiv] A bioRxiv preprint by Amburn et al. introduces AltereGO, a framework that combines molecular function predictions from three sequence representations, ESM2-3B, ProtT5 and Ankh, and a structure representation, ProstT5-3Di, through weighted score-level fusion. The authors evaluated it on 495,055 proteins spanning 8,446 Gene Ontology molecular function terms and on a low-homology test set of 2,000 proteins with experimental or curated annotations. According to the paper, ESM2-3B was the strongest standalone representation, while four-way fusion achieved the strongest overall classification and ranking performance and outperformed the learned embedding-level fusion architectures tested. In matched comparisons with eight released prediction methods, AltereGO showed significantly improved AUPRC ranking in every comparison and significantly better Fmax and Smin in six; the remaining two showed no significant differences in those metrics. Structure-derived predictions contributed smaller, targeted gains in catalytic, redox and cofactor-associated functions.

[bioRxiv] A bioRxiv preprint by Cheng, Shireman and co-authors presents SMART, a multi-agent large language model framework for "discovery transfer" in single-cell RNA-sequencing studies, meaning testing whether a published finding holds in another cancer type or patient population. According to the authors, given a source manuscript and a target scRNA-seq dataset, SMART extracts major clinical and mechanistic findings, generates and executes analyses for the testable ones, and reports which findings transfer along with suggested follow-up. Applied to 121 studies across six cancer types, SMART identified findings that transfer across cancers, including a lung cancer prognostic signature the authors say they confirmed to be associated with survival in liver cancer, pancreatic cancer and pleural mesothelioma; SMART also generated new mechanistic hypotheses. The preprint has not been peer-reviewed.

[bioRxiv] A preprint presents CellCompass, an agentic AI system that couples a scientific agent with an extensible dynamical engine for modeling state transitions, population change, and fate and molecular readout. According to the paper, the system reuses existing formulations when sufficient and designs, implements and refines new ones when experimental measurements introduce additional constraints. The authors developed DynBench100, a benchmark of 100 simulated single-cell time-course scenarios with hidden dynamical ground truth, and reported that CellCompass achieved the highest aggregate performance among tested general-purpose and biomedical agents across six dynamical tasks. In four real single-cell studies, incorporating metabolic RNA labeling in drug-response data reduced pulse-control identifiability error by 31.1%; incorporating clonal barcodes in hematopoietic lineage tracing improved clone-fate agreement by 54% over the strongest dynamical baseline; and in time-resolved CITE-seq data the system reduced future surface-protein distribution error by 13.2% for RNA-initialized trajectories. The findings are from a preprint and have not been peer reviewed.

[bioRxiv] A preprint study benchmarked large language models and agentic coding systems, including ChatGPT 5.5, Claude Sonnet 4.6, Gemini 3.1 and Llama4-Scout, on automated analysis and visualization workflows for transcriptomic data, focusing on human responses to COVID-19 infection. According to authors Iyer, Subedi, Butler and Krishnakumar, the protocol varied prompt specificity and tested workflow components such as dimensionality reduction, clustering, outlier detection, reproducibility and code generation, recording quantitative metrics including clustering scores and outlier counts alongside qualitative aspects such as code quality. The authors said prompt specificity improved code structure and methodological correctness, but quantitative performance depended heavily on preprocessing choices and execution environment. Some models produced strong code and reproducible outputs, yet alignment with scientific targets such as label-based clustering was inconsistent. The authors said the findings highlight both the promise and limitations of current LLMs for automating scientific data analysis and visualization, and suggest directions for future benchmarking and workflow standardization.

[bioRxiv] Hu, Li, Jiang and co-authors present ProtCell, a foundation model for single-cell proteomics developed using approximately 380 million cells from 98 datasets for pretraining and evaluation, according to the bioRxiv preprint. The model combines protein-identity and continuous-abundance embeddings in a transformer encoder and learns cell- and protein-level representations through masked abundance modeling. According to the authors, ProtCell first connects fragmented profiles through transferable annotation and batch correction while preserving biological organization, then extends limited profiles through abundance reconstruction and inference of an unmeasured cytotoxic protein program enriched in natural killer cells and memory CD8+ T cells. The authors further say the learned protein dependencies guide selection of compact panels that retain information for profile reconstruction and biological classification. The preprint has not been peer-reviewed.

[bioRxiv] A preprint study examined how protein language models learn from suboptimal datasets and tested augmentation approaches to improve predictive performance. Stump et al. said that narrowing the distribution of protein-protein complexes in a limited dataset allows an order-of-magnitude decrease in training dataset size if a model is only expected to generalize to in-distribution data. In active learning, both in- and out-of-distribution Matthews Correlation Coefficient (MCC) showed a small but consistent increase after augmenting the training set with complexes produced by ESM-3, a generative protein language model, according to the authors. The work has not been peer-reviewed.

[bioRxiv] A bioRxiv preprint by da Silva de Araujo and co-authors introduces dataset interrogation, a conversational framework in which a researcher uses a large language model (LLM) agent to translate emerging questions into executable tests, check answers against controls, and keep code, inputs and decisions in a dated record. Applied retrospectively to four LC-MS/MS datasets, the framework exposed problems that changed the interpretation of three in-house studies: a search database lacking one of two symbionts, a swapped pair of tandem mass tag (TMT) channels, statistics inflated by non-host proteins, and false deamidation assignments caused by selecting the 13C isotope peak as monoisotopic; the same question asked of every dataset revealed failed or partial cysteine alkylation in all three, according to the authors. A public TMT dataset was reanalyzed from raw files in 24 minutes, a feasibility measure without a human comparison, the authors said; among spectra identified by both workflows, 99.2% received the same peptide assignment as in Proteome Discoverer, and protein fold changes correlated (r = 0.93). In seeded tests, a purpose-built channel assignment check detected all 216 between-strain channel swaps and none of 90 within-strain swaps, according to the authors. The authors said the results establish feasibility in a limited retrospective setting with one researcher, and that broader evaluation across users, models and datasets is required to determine general performance.

[bioRxiv] Grigson, Neri, Roux and colleagues propose a viral host prediction framework, VHAMSTeR, which can classify viral sequences into five host categories: prokaryotes, animals, plants, fungi and protists. The paper states that the method combines fine-tuned genomic language model embeddings with explicit biological features such as genomic structure and protein homology, and significantly outperforms existing host prediction methods on both DNA viruses (MCC>0.91) and RNA viruses (MCC>0.69). The authors applied it to the global metagenomic catalog, resolving host associations in the RNA virome and identifying 175 cryptic RNA phages; analysis of marine viral genomes showed enrichment of protist viruses in polar oceans; comparison of human gut viromes along an urbanization gradient in Africa showed that symbiotic protist viruses detected in rural cohorts decrease in industrialized populations. The code has been made public.

Sun, Hudelmaier, Gilligan and colleagues report that they used a genomic language model fine-tuned on prokaryotic intergenic sequences to learn representations of noncoding elements from DNA sequences, and organized 376 million intergenic regions from 1.3 terapascal-base-pair prokaryotic genomes and metagenomes into 5.14 million clusters. The paper reports that the method nominated 20,149 clusters that may form structural RNAs and that carry candidate protein partners; periodicity in token entropy revealed internal structure, recovering CRISPR, TIGR and VIPR arrays and their corresponding effector proteins from sequence alone. Targeted clustering of sequences adjacent to bridge recombinases identified bridge RNA variants whose guide recognition sequences are longer than those of any characterized homolog. The study also had a team of AI agents mine databases according to researcher-defined criteria, discovering differentiated riboswitches, ribozymes and other structural RNAs. The authors state that the database has been made public.

Tiwari, Mahajan, Bellad and others have released a preprint study proposing a benchmarking framework called EL-PERTURB, which stress-tests 20 perturbation prediction models, covering six classes of architectures, and evaluates them on essential-gene and genome-wide Perturb-seq screening data. The authors state that existing benchmarks mostly rely on random train-test splits, which mask the systematic failure modes of models under real out-of-distribution (OOD) conditions; EL-PERTURB adopts a stratified OOD splitting strategy that places the training set and the test set in functionally distinct regions of transcriptomic space. The study reports that model performance drops substantially under OOD conditions, and that architectural complexity—including single-cell foundation model embeddings—offers no reliable advantage over well-tuned simple baselines. The authors state that the framework provides a rigorous benchmark covering different levels of generalization difficulty for perturbation-effect prediction, and can guide the construction of models that can truly replace wet-lab screening.

[bioRxiv] Hunewald, Grant, Boehm and others have posted a preprint study proposing a pipeline that requires no gating hierarchy, manually set marker thresholds or task-specific training data: median marker intensities from mass cytometry clustering (such as FlowSOM and GigaSOM) are scaled by marker and merged into dozens of metaclusters, and then the continuous marker profiles are handed to a large language model in a single pass, which returns cell type labels and confidence scores. The paper says that across four human and mouse datasets covering blood, bone marrow and intestine, a single selected model achieved 85-98% agreement with the reference-annotated metaclusters at the major lineage level (Claude Sonnet 4.6: 85-98%; the best local model: 83-96%), and Cohen's kappa of 0.67-0.83 after correcting for chance agreement. The authors say local open-weight models were comparable to cloud models on an independent benchmark, but reliability varied by model, and Claude Opus 5 was not more reliable than Sonnet 4.6. Two experts who blindly annotated the same heatmaps agreed at about 99% at the major lineage level but only about 73% on T cell maturation status, and the best model fell within the range of expert disagreement.

[bioRxiv] Rana, Hossan, Rafi et al. published a preprint study combining pan-cancer multi-omics analysis with machine-learning virtual screening for targeting protein arginine methyltransferase 5 (PRMT5). The paper states that bioinformatics analysis showed PRMT5 has tumor-specific prognostic effects in GBM, LGG and KIRC; a deep neural network screened from approximately 1.5 million PubChem compounds, achieving a ROC-AUC of 0.9857, and preferentially selected three candidates: CID 44345182, CID 2794188 and CID 44345334. Following molecular docking, ADMET filtering, toxicity prediction and 100-nanosecond molecular dynamics simulations, CID 44345182 was identified as a potential hit, with a docking score of -9.8 kcal/mol, compared with -9.2 kcal/mol for the reference ligand. The authors state that experimental validation is still needed.