Research Radar — 2026-05-15

Summary: TxPert predicts unseen single-gene transcriptomic perturbation effects by integrating multiple knowledge graphs, achieving accuracy approaching split-half experimental reproducibility. The framework generalizes across cell types and perturbation classes without requiring paired pre/post perturbation training data.

Why it matters: Achieving near-experimental accuracy for unseen perturbations is a major advance for in silico drug screening and CRISPR experimental design. Knowledge graph integration provides a principled way to incorporate prior biological knowledge into deep learning.

Why for Yiru: Transcriptomic perturbation prediction with knowledge graph priors directly relevant to modeling drug responses and genetic perturbations in cancer and immune cells — core to computational oncology research.

Summary: Demonstrates that standard evaluation metrics for single-cell perturbation prediction (correlation and cosine similarity of differential expression) are systematically inflated by statistical bias from reusing the same control population. A simple control-splitting procedure removes this bias, substantially reducing previously reported performance — even non-informative methods appear competitive under biased metrics.

Why it matters: This finding demands re-evaluation of published perturbation prediction benchmarks and raises the bar for method development. The field risks optimizing for artifacts rather than biological signal.

Why for Yiru: Methodological rigor in computational biology is essential. Boss's work on perturbation modeling must account for evaluation bias to avoid overoptimistic conclusions.

Summary: Fine-tunes ESM2, a 650M-parameter protein language model, for phosphorylation, acetylation, and ubiquitination-site prediction using parameter-efficient fine-tuning with focal loss to handle extreme class imbalance. The task-specialized models reveal sequence grammar governing PTM site selection, identifying motifs that distinguish modified from chemically eligible but unmodified residues at proteome scale.

Why it matters: PTM prediction at proteome scale with PLMs bridges sequence and function. Understanding what makes a residue modified vs. merely chemically eligible is a fundamental biochemical question with implications for signaling biology and drug targeting.

Why for Yiru: Protein language models applied to functional site prediction connect Boss's interests in AI methods with biological mechanism discovery. PTM regulation is central to immune signaling and cancer biology.

Summary: Comprehensive review of explainable artificial intelligence (XAI) methods tailored for protein language models. Covers attention analysis, probing classifiers, concept-based explanations, and structural interpretability approaches, with practical guidance on when each method is appropriate and what biological questions they can address.

Why it matters: As PLMs become central to computational biology, understanding what they learn is critical for trust and scientific discovery. This review provides a roadmap for extracting mechanistic insight rather than just predictive performance.

Why for Yiru: Interpretability of deep learning models in biology is a recurring challenge. The XAI methods surveyed here apply broadly to any biological sequence or structure model Boss might develop or use.

Summary: Introduces the CLIC (Cross-modality Link Confidence) score, a quantitative measure of empirical concordance between gene expression and nearby chromatin accessibility derived from diverse single-cell multiome datasets. CLIC enables filtering of low-confidence associations that compromise integration accuracy when combining separately profiled scRNA-seq and scATAC-seq data.

Why it matters: Most single-cell data remains unimodal. Better integration of separately profiled modalities is essential for maximizing the value of existing datasets. CLIC provides a principled confidence metric for a problem that's typically treated as uniformly reliable.

Why for Yiru: Multi-omics integration is central to spatial and single-cell analysis. A confidence framework for cross-modal associations directly improves the interpretability of integrated TME atlases.

Summary: Systematic comparison of different masking approaches for rare variant association tests across 54 traits in UK Biobank identifies optimal strategies that increase study power and replicability. Provides practical guidance for baseline masking, variant annotation, and covariate adjustment in gene-level burden analyses.

Why it matters: Rare variant association testing is a cornerstone of statistical genetics but methodological choices dramatically affect power. Empirically grounded best practices from this scale of analysis are highly actionable.

Why for Yiru: Statistical genetics methodology is relevant to understanding the genetic architecture of immune-related traits and cancer susceptibility, which contextualize Boss's translational research.