Research Radar — 2026-08-23

Summary: We introduce LifeSciBench, a benchmark of 750 expert-authored tasks designed to evaluate whether language models can handle realistic life science research work. The majority of existing life sciences benchmarks have a narrow scope or are purely knowledge-based, and therefore fail to capture the complexity of real-world research, which often involves ambiguities and requires the accurate execution of multiple dependent judgment calls. Additionally, almost all existing benchmarks span at best a small collection of subdomains within the life sciences; there is at present no existing life sciences benchmark with both the requisite breadth and depth required to convincingly measure proficiency in real-world professional research settings. LifeSciBench addresses this gap by spanning seven representative scientific workflows and seven life science domains, with each constituent task paired with a human expert-written rubric. Across five frontier and domain-specialized models, GPT-Rosalind performs best, with a task-weighted mean normalized rubric score of 0.576 and a task-weighted response pass rate of 36.1% (response-level values are first averaged within each task, and the resulting task-level values are then averaged with equal weight). LifeSciBench remains unsaturated, with 171 tasks (22.8%) having no observed passing response from any evaluated model and 261 tasks (34.8%) having a best-model pass rate below 20%. LifeSciBench therefore serves as a high-resolution evaluation of practical scientific reasoning and operational decision-making in the life sciences.

Why it matters: Life-science AI evaluation is directly relevant to biomedical AI and is unusually realistic: expert-authored, multi-step tasks expose unsaturated performance and operational failure modes that knowledge-only benchmarks miss.

Why for Yiru: Fit: biomedical AI; scientific agents.

Summary: Accurate normalization is essential for differential expression analysis of RNA-sequencing data. Popular normalization methods such as the median-of-ratios and trimmed mean of M-values do not leverage information from the experimental design. This may be inefficient in experiments with large-scale systematic expression changes or complex designs. Here, we introduce design-informed size factor estimation (disize), a normalization method that uses information from the experimental design to improve accuracy. disize uses a modified generalized linear mixed model to robustly distinguish between biological signal and sample-specific size factors. We also propose a mechanistically justified data-generating process for RNA-sequencing counts that is derived from previous models of transcription and sequencing. Through simulations based on this data-generating process and validating on true RNA-seq data, we show that disize recovers size factors more accurately than existing methods, particularly in challenging scenarios with low gene expression and a high proportion of differentially expressed genes; this in turn improves downstream analysis. disize provides a robust and accurate approach to normalization, highlighting the significant benefits of integrating experimental design information directly into normalization for transcriptomic datasets. Author summaryIn transcriptomic analysis, normalization adjusts for technical biases arising from library preparation and sequencing. Methods implemented in widely used packages like DESeq2 and edgeR ignore information in the experimental design during normalization. Incorporating information from the experimental design into a normalization method has the potential to yield more accurate results. To do this, we developed a new method, design-informed size factor estimation (disize), that uses a statistical model to jointly account for the biological signal defined by the design and the sample-specific batch effect. By separating the biological variation into its components, disize can more robustly estimate the batch effect. To validate our approach, we constructed a flexible simulation framework relying on a mechanistically justified data-generating process for RNA-seq data. Our benchmarks on both simulated and true RNA-seq data show that disize recovers the true size factors more accurately than existing methods, particularly in challenging scenarios with low counts or a high proportion of differentially expressed genes. This improved normalization yields more reliable downstream results in differential expression analysis.

Why it matters: Design-informed normalization addresses a foundational failure mode in transcriptomics by incorporating experimental design into size-factor estimation. The statistical model and explicit difficult-regime validation are transferable to spatial and single-cell workflows.

Why for Yiru: Fit: spatial omics; transcriptomics; statistical modeling.

Summary: Perturbational transcriptomics is a powerful tool for understanding gene function and drug effects, yet predicting how perturbations manifest across different biological contexts remains a central challenge, limiting translation from model systems to clinically relevant tissues. Despite growing interest in this problem, benchmarking efforts have been hindered by inconsistent evaluation tasks, heterogeneous metrics, and limited assessment across perturbation types and biological systems. Here, we introduce a benchmarking framework for cross-context perturbation-signature prediction (a task we define as signature recontextualization), grounded in explicit definitions of the prediction task, target-data availability, and evaluation metrics centered on signature recovery. The framework evaluates prediction performance across three target-context data regimes: (1) control only, where only control profiles from the target context are measured; (2) low coverage, where a limited subset of perturbations in the target context are measured; and (3) high coverage, where most perturbations in the target context are measured. This design enables systematic assessment of how prediction performance depends on target-context sample size while providing a standardized basis for comparing methods. We evaluate newly developed projection-based (projectCor) and network-based (netProp) methods alongside deep learning-based foundation models (scGPT, STACK) and statistical baselines. The benchmark spans four diverse perturbational datasets: CRISPR knockdowns and drug perturbations in cell lines, plus in vivo chemical perturbations in rat tissues from DrugMatrix, extending evaluation beyond isolated cell-line models to tissue-level responses. Across tasks, projection and network propagation approaches show strong flexibility across perturbation types and biological contexts, and in several cases match or exceed the performance of deep learning and foundation models, suggesting that model complexity does not inherently improve cross-context generalization. We further show that perturbation predictability varies substantially with pathway conservation, transcriptional response strength, and baseline similarity between source and target contexts. All datasets, methods, and evaluation utilities are released as an open-source R package (sigRecon), providing a foundation for reproducible benchmarking and future method development.

Why it matters: This work formalizes cross-context perturbation-signature prediction and compares projection, network propagation, foundation models, and statistical baselines across cell-line and tissue settings. The target-data regimes make transferability a measurable question rather than an assumption.

Why for Yiru: Fit: spatial omics; perturbation biology; foundation models.

Summary: Lacuna, an open-source Python tool for discovering cryptic binding pockets: sites that are absent or too small to detect in a protein's unbound structure and open only during conformational fluctuation. Most binding-site predictors score a single static structure, which is precisely the structure in which a cryptic site is invisible. Lacuna instead generates a conformational ensemble from any input structure, detects pockets independently in every conformer, clusters the detections into persistent sites across the ensemble, and ranks those sites with a model fitted on within-structure pairs. Ensemble generation is pluggable: normal mode analysis by default, with implicit-solvent molecular dynamics, Boltz-2 diffusion sampling, or a user-supplied ensemble as alternatives. On the designated test fold of CryptoBench, Lacuna recovers 55.6% of cryptic sites in its top five predictions, rising to 66.1% with an optional PLM-assisted ranker, and it recovers 73%, 45% and 87% on the PocketMiner set, a curated set of literature apo/holo pairs, and COACH420 respectively. The default backend completes in a median of 2.6 seconds per chain on one CPU core, so ensemble-based pocket finding does not require a simulation budget. Every site carries a continuous crypticity score, and outputs are emitted as docking-ready Boltz YAML constraints, AutoDock Vina boxes and pseudoatom PDB files. Lacuna is MIT licensed and available at https://github.com/mooreneural/lacuna and on PyPI as lacuna-pockets.

Why it matters: Lacuna tackles cryptic-pocket discovery where a static structure is specifically the wrong representation, using conformational ensembles and docking-ready outputs. It is a practical, open method with clear transfer to structure-based drug discovery, though evidence remains benchmark-based.

Why for Yiru: Fit: drug discovery; protein structure; conformational ensembles.

Summary: Shapley-value-based explanations, notably SHAP (SHapley Additive exPlanations), have gained prominence as a principled game-theoretic framework for local explanations and global feature importance. While exact Shapley value computation is exponential in feature count, TreeExplainer exploits the recursive structure of decision trees to achieve polynomial-time computation for tree-based ensembles. In many scientific applications, however, features are naturally organized into a priori groups reflecting domain knowledge, requiring explanations both across and within groups. The Owen value extends the Shapley value through a two-stage allocation rule that incorporates group structure while preserving fairness properties; yet, efficient algorithms for its computation remain limited. In this paper, we propose exact and Monte Carlo algorithms for computing Owen values in tree-based ensembles by combining hierarchy-guided group aggregation with tree-aware dynamic programming. The exact algorithm computes Owen values without sampling under the path-dependent characteristic function, which approximates the conditional expectation, whereas the Monte Carlo algorithm provides a scalable approximation that is unbiased for any prespecified sampling budget and converges almost surely as the sampling budget increases. We also provide global importance measures and visualization tools for structured, multi-resolution explanations. The proposed algorithms and tools are collectively referred to as TreeOwen. Through simulation experiments, we demonstrate the numerical accuracy and substantial computational gains of TreeOwen. We illustrate its practical utility using immunotherapy metagenomic data, showing how microbial genera (groups) and species (features) contribute to patient recovery.

Why it matters: TreeOwen extends interpretable tree-ensemble explanations to biologically meaningful feature groups, with exact and scalable algorithms and an immunotherapy metagenomics example. Group-aware explanations are more transferable than flat feature rankings for multimodal biomedical data.

Why for Yiru: Fit: biomedical AI; interpretability; immunotherapy.