Research Radar — 2026-08-25
Methods & AI
Computational
Model Validation Protocols for Machine Learning in Small Molecule Drug Discovery
bioRxiv (Bioinformatics) Published 2026-08-24 Preprint DOI:
drug discovery molecular property prediction machine learning model validation distribution shift
Summary: A cross-industry consortium presents five recommendations for validating molecular property-prediction models under realistic distribution shifts. ADME case studies across two model algorithms identify extrapolation, interpolation, representation, and evaluation failure modes, showing that common protocols can substantially overestimate performance.
Why it matters: The framework moves drug-discovery ML evaluation beyond aggregate benchmark scores toward understanding where and why models fail, including failures caused by molecular representations.
Why for Yiru: This is a transferable validation template for biomedical AI and drug discovery, especially when assessing generalization beyond retrospective molecular benchmarks.
Virtual-cell models compress unseen intervention geometry through a target-specific generalization bottleneck
bioRxiv (Bioinformatics) Published 2026-08-24 Preprint DOI:
virtual cells cellular perturbation generalization intervention geometry digital twins
Summary: The study identifies Intervention Geometry Compression: models can reconstruct molecular states while losing the relationships that distinguish unseen interventions. Diagnostic projections localize missing geometry to residual response directions, and time-resolved analyses show that correct trajectory entry and same-target empirical anchoring improve intervention-identity transfer.
Why it matters: It reveals a failure of virtual-cell models that state-level similarity can conceal and proposes empirical anchoring of intervention identity as a concrete design principle.
Why for Yiru: The result is directly relevant to biomedical AI and digital-twin modeling of perturbations, where preserving intervention identity matters as much as reconstructing expression states.
A mechanism-annotated benchmark reveals limited fidelity to drug-response signatures in single-cell perturbation models
bioRxiv (Bioinformatics) Published 2026-08-24 Preprint DOI:
single-cell perturbation drug response mechanism-aware benchmarking foundation models drug discovery
Summary: scDrugPerturb-Bench links 181 datasets, 423 annotated response cases, 717 key genes, and 2.5 million cells to evaluate direction, effect size, gene-set coherence, mechanism specificity, and pathway polarity. Across 12 models, expression-similarity metrics were weakly aligned with the Mechanism Fidelity Score, while source context and hard-negative tests exposed shortcut-driven plausible responses.
Why it matters: The benchmark shows that expression reconstruction is an insufficient proxy for preserving drug-response mechanisms and that mechanism-aware selection can improve early drug retrieval.
Why for Yiru: This provides a reusable evaluation design for single-cell perturbation and computational immunology models that claim to infer treatment mechanisms.
RADF: Reference-Anchored Dynamic Flow for Spatial Perturbation Profile Completion
bioRxiv (Bioinformatics) Published 2026-08-24 Preprint DOI:
spatial omics spatial perturbation functional genomics generative modeling response heterogeneity
Summary: RADF completes unmeasured spatial perturbation profiles by using the reported population as an empirical response distribution. A Sinkhorn-balanced decoder creates a population-valued reference anchor, while bounded dynamic relational flow adapts relations to expression state and query geometry; the method reduces macro E-distance by 70.6% versus an existing spatial method.
Why it matters: Reference-supported population modeling offers a way to preserve observed response heterogeneity while learning location-specific variation in destructive, capacity-limited spatial screens.
Why for Yiru: This is a direct spatial-omics method fit for modeling perturbation responses across tissue locations without discarding empirical population structure.
sc-pcQTL: hurdle-based co-expression modeling for multi-gene QTL mapping in single-cell RNA-seq data
bioRxiv (Bioinformatics) Published 2026-08-23 Preprint DOI:
single-cell RNA-seq QTL mapping co-expression genetic regulation computational immunology
Summary: sc-pcQTL combines hurdle modeling, sliding-window clustering, principal components, and cis-pcQTL mapping for sparse single-cell counts. In 1.24 million PBMCs from 982 donors across 10 cell types, it identified 2,485 local co-expression clusters and 2,040 clusters with significant cis-pcQTL associations; fine-mapping and FinnGen colocalization yielded 394 QTL-GWAS signal groups.
Why it matters: The method captures coordinated multi-gene regulation that gene-by-gene eQTL analyses miss and connects cell-type-specific cluster phenotypes to disease loci.
Why for Yiru: It offers a scalable statistical approach for single-cell genetic regulation and computational immunology, with unusually strong donor-scale and GWAS-linked validation.
Time-resolved operator archetypes characterize dynamical sensitivity during cell-state transitions
bioRxiv (Bioinformatics) Published 2026-08-23 Preprint DOI:
cell-state transitions single-cell dynamics Jacobians trajectory inference perturbation biology
Summary: scJDO treats time-indexed Jacobians as analytical objects, projects them into a shared subspace, and decomposes them into interpretable operator archetypes with temporal activation profiles. Using a neural drift field learned from cell-state geometry, it distinguishes productive from diverted iPSC reprogramming trajectories and validates the framework in synthetic, hematopoietic, time-course, and CRISPRi Perturb-seq settings.
Why it matters: The operator-level representation makes changing dynamical sensitivity measurable while explicitly addressing what snapshot data can and cannot recover.
Why for Yiru: This provides a useful dynamical lens for cell-state modeling and perturbation analysis, including settings where splicing data are unavailable.
Cross-disciplinary watchlist
Other Fields
Continuous tissue fields organize immune composition in pancreatic cancer
bioRxiv (Cancer Biology) Published 2026-08-24 Preprint DOI:
spatial transcriptomics pancreatic cancer tumor microenvironment immune composition continuous tissue fields
Summary: Across three Visium cohorts and an independent single-cell imaging dataset, myCAFs formed millimeter-scale fields along which immune composition changed continuously from cytotoxic T cells and mast cells toward SPP1 macrophages, monocytes, and neutrophils. Single-cell spatial data also showed a continuous basal-to-classical tumor identity axis associated with distance from myCAF interfaces.
Why it matters: The findings challenge discrete-neighborhood descriptions by showing continuous stromal and epithelial organization at both tissue and cell-contact scales.
Why for Yiru: This is an excellent spatial-omics and oncology example for modeling tumor ecosystems as continuous fields rather than forcing biologically meaningful gradients into compartments.
Spatial immune ecosystems govern therapeutic response in HER2-low breast cancer
bioRxiv (Cancer Biology) Published 2026-08-24 Preprint DOI:
spatial immunology HER2-low breast cancer dendritic cells tumor microenvironment therapeutic resistance
Summary: Single-cell spatial transcriptomics identifies remodeled dendritic-cell states and spatial immune architectures in HER2-low breast tumors. Resistant tumors show segregation of epithelium from effector immune populations, myeloid-rich niches, and unfavorable homeostatic cDC2 signatures, whereas sensitive tumors retain immune-intermixed niches; TCGA BRCA provides independent clinical validation.
Why it matters: The study links treatment resistance to coordinated dendritic-cell state remodeling and immune segregation, identifying a spatially organized myeloid niche as a potential biomarker and vulnerability.
Why for Yiru: It directly connects spatial immune architecture to therapeutic response and is highly relevant to computational immunology and tumor-microenvironment modeling.
Intracranial Targeting of Cholesterol Processing Reveals a Therapeutic Vulnerability that Reprograms Glioblastoma and Promotes Antitumor Immunity
bioRxiv (Cancer Biology) Published 2026-08-19 Preprint DOI:
glioblastoma cholesterol metabolism single-cell RNA-seq tumor immunity drug repurposing
Summary: Combined clemastine and bexarotene treatment targets cholesterol biosynthesis, transport, and homeostasis in patient-derived glioma models, inducing endoplasmic-reticulum stress, autophagy, and apoptosis. Orthotopic models show reduced progression and longer survival with local delivery, while single-cell RNA-seq reveals regeneration, plasticity, and immune-microenvironment remodeling.
Why it matters: The work integrates a metabolic vulnerability, patient-derived and orthotopic models, single-cell state changes, and antitumor immunity into a testable therapeutic strategy.
Why for Yiru: It is a strong oncology and drug-discovery case study for connecting metabolic intervention with tumor-state and immune-microenvironment modeling.
Drug screen and machine learning predict neuroprotective agents in a preclinical human model of childhood dementia
Nature Communications Published 2026-08-20 Research article DOI:
biomedical AI drug screening childhood dementia patient-derived models multimodal phenotyping
Summary: Patient-derived neural models of childhood dementia reproduce key disease features, and a screening platform integrates machine learning with imaging, transcriptomics, and electrophysiology to identify repurposed compounds that mitigate adverse effects.
Why it matters: The multimodal workflow demonstrates how patient-derived human models and complementary readouts can support translational drug repurposing beyond a single assay.
Why for Yiru: This is a useful biomedical-AI and drug-discovery example of combining imaging, transcriptomics, electrophysiology, and human disease models.