Research Radar — 2026-06-24

Summary: Presents HTS-Oracle X, a multimodal deep learning platform that integrates bidirectional cross-attention fusion of ChemBERTa SMILES embeddings with extended RDKit descriptors for prospective discovery of small molecule immune checkpoint binders. Targeting immune checkpoint protein-protein interactions with small molecules has been limited by the shallow, featureless binding surfaces of co-stimulatory and co-inhibitory receptors and the low hit rates of conventional high-throughput screening. HTS-Oracle X trains on continuous biophysical binding signals rather than binary labels, employs Monte Carlo Dropout uncertainty quantification for uncertainty-adjusted compound selection, and was trained on 45,760 Dianthus TRIC-screened compounds per target under scaffold-aware cross-validation. The platform was applied prospectively and successfully discovered novel small molecule binders targeting immune checkpoint interfaces, demonstrating that AI-guided screening can overcome the limitations of conventional approaches against these challenging targets.

Why it matters: Small molecule immunomodulators that target immune checkpoints could overcome key limitations of antibody-based checkpoint blockade — including poor tumour penetration, lack of oral bioavailability, and high manufacturing costs — but discovering such molecules has proven exceptionally difficult because checkpoint receptor interfaces are large, flat, and featureless. HTS-Oracle X demonstrates that multimodal deep learning combining chemical language models with structural descriptors, trained on continuous binding signals with proper uncertainty quantification, can prospectively discover active compounds against these challenging targets. If this approach generalises to other immune checkpoint targets, it could open a new class of orally available, tumour-penetrating immunomodulators for cancer immunotherapy. The uncertainty quantification component is also critical: it enables compound selection decisions that account for model confidence, reducing the number of false leads that waste downstream screening resources.

Why for Yiru: Small molecule immune checkpoint inhibitors represent a frontier in TME pharmacology. Unlike antibodies, small molecules can penetrate deep into tumour tissue, reach immune synapses within the TME, and potentially be designed with short half-lives for controlled immunomodulation. HTS-Oracle X's multimodal architecture — fusing SMILES embeddings with molecular descriptors — is a design pattern applicable to any TME drug discovery problem where binding surfaces are challenging. From a computational perspective, the uncertainty quantification framework is particularly relevant: TME drug discovery often suffers from high false-positive rates in phenotypic screens, and Monte Carlo Dropout-based uncertainty selection could improve hit-to-lead success rates across TME target classes.

Summary: Presents Drug-Prot, a computational framework that leverages large-scale perturbation proteomics to quantify causal drug effects, drug-drug interactions, and dynamic protein relationships. Understanding drug effects and interactions is essential for developing combination therapies, yet most computational approaches rely on transcriptomic or phenotypic data that miss the proteomic level of drug action. Using data from 63 single drugs and 59 drug combinations applied to 18 breast cancer cell lines at three time points (6, 24, and 48 hours), Drug-Prot estimates drug effects on protein expression and reconstructs directed temporal protein dependency networks. The framework provides publicly available software enabling targeted analyses of user-defined protein sets, substantially reducing the multiple-testing burden, and offers an interactive web application for corrected p-values for single-drug and combination effects with directed temporal network visualisation.

Why it matters: Most drug effect studies measure transcriptomic changes, but proteins are the functional executors of drug action, and transcript-protein correlations are often poor — especially for drugs that affect protein stability, trafficking, or degradation. Drug-Prot fills this gap by providing a rigorous statistical framework for proteomic drug profiling across time, cell lines, and drug combinations. The ability to query user-defined protein sets with proper multiple-testing correction makes the resource practical for hypothesis-driven research. The temporal dimension (three timepoints) is particularly valuable: it enables reconstruction of directed protein networks that distinguish direct targets from downstream effects, providing mechanistic insight beyond static profiling.

Why for Yiru: Proteomic profiling of drug effects is underexplored in TME research compared to transcriptomics, yet many TME-relevant drugs — kinase inhibitors, epigenetic modulators, protein degraders — act primarily at the protein level. Drug-Prot's framework could be applied to TME-relevant drug combinations: for example, profiling how checkpoint blockade combined with targeted therapy alters the TME proteome over time. The directed temporal protein network reconstruction is directly applicable to understanding signalling rewiring in the TME upon drug treatment. From a methodological perspective, Drug-Prot's statistical framework for quantifying combination effects (synergy, additivity, antagonism) at the proteomic level provides a template for analysing perturbation proteomics data from TME model systems.

Summary: Introduces multiRF, a random-forest-based method for cross-modal multi-omics integration that handles complex data types and separates shared and modality-specific structure. Current multi-omics clustering methods typically merge all data types into a single representation, which can blur biology that is strong in one omics layer, or rely on linear structure that misses complex cross-modal relationships. multiRF learns sample similarities across omics layers from multivariate random forests, combines them across data types, and uses the resulting weights to estimate which part of each omics layer is predictive of the others, thereby decomposing each layer into shared and modality-specific components. The method outperforms existing approaches across diverse benchmarks, recovering known biology that is specific to individual omics layers while also extracting integrated signals.

Why it matters: Multi-omics data are increasingly generated in large cohort studies, but integrative analysis remains challenging because different omics layers capture different biological processes — some signals are shared, others are layer-specific. multiRF's explicit separation of shared and modality-specific structure is conceptually important: it recognises that not all biology needs to be integrated, and that layer-specific signals may be as informative as shared signals. The random forest foundation makes the method robust to non-linear relationships, mixed data types, and missing data, all of which are common in real-world multi-omics datasets. As multi-omics cohorts grow, methods that can flexibly integrate diverse data types while preserving interpretability become essential.

Why for Yiru: TME research increasingly generates multi-omics data — genomics, transcriptomics, proteomics, epigenomics — from the same tumour samples. multiRF's decomposition of shared versus modality-specific variation could reveal whether certain TME features (e.g., immune infiltration signatures) are consistently captured across omics layers or are specific to particular molecular measurements. For biomarker discovery, identifying features that are robustly shared across omics layers might yield more reproducible predictors of immunotherapy response than single-omics biomarkers. The method's ability to handle non-linear relationships is particularly relevant for TME data, where molecular measurements often exhibit complex, non-linear associations with clinical outcomes.

Summary: Performs whole-genome nanopore sequencing on normal and tumour prostate tissues to characterise differential methylation, methylation entropy, and allele-specific methylation (ASM) associated with noncoding genetic variants. Long-read nanopore sequencing enables simultaneous detection of germline variation and native DNA base modifications on individual DNA molecules, providing a unique opportunity to investigate allele-specific epigenetic regulation. Genome-wide analysis identified extensive cancer-associated differentially methylated regions, with hypermethylated DMRs significantly enriched near transcription start sites and regulatory regions. Integration with transcriptomic datasets revealed strong inverse relationships between promoter methylation and gene expression. The study further demonstrates that ASM at regulatory noncoding variants can help functionally annotate prostate cancer risk-associated variants that fall outside protein-coding regions.

Why it matters: Most GWAS-identified risk variants for common cancers fall in noncoding regions whose functional significance is difficult to determine. Allele-specific methylation provides a molecular readout of whether a noncoding variant affects local epigenetic regulation in cis, offering a path from statistical association to mechanistic understanding. The use of long-read nanopore sequencing is methodologically important: unlike short-read approaches, it can phase methylation and genetic variation on individual DNA molecules, directly revealing which allele of a heterozygous variant is methylated. This phased information is essential for distinguishing cis-acting regulatory variation from trans-acting effects.

Why for Yiru: Epigenetic regulation in the TME is increasingly recognised as a key determinant of immune cell function and tumour progression. nanoASM's approach to linking noncoding genetic variants with allele-specific methylation could be applied to TME-relevant regulatory variants — for example, identifying whether tumour-associated noncoding variants in immune checkpoint genes affect local methylation and thus expression. The phased methylation analysis is also relevant for understanding how tumour heterogeneity in methylation patterns contributes to immune evasion. Methodologically, the long-read approach could be extended to profile allele-specific methylation in TME immune cells, revealing how germline variation shapes the epigenetic landscape of anti-tumour immunity.

Summary: Introduces DLDN-Bench, a standardised benchmark framework for evaluating deep learning-based de novo peptide sequencing from mass spectrometry data. De novo peptide sequencing is essential for identifying novel peptides without relying on protein sequence databases, a capability critical for discovering tumour neoantigens, post-translational modifications, and peptides from non-model organisms. However, the rapid proliferation of deep learning models has led to heterogeneous evaluation practices and limited comparability across methods. DLDN-Bench provides benchmark datasets derived from human muscle biopsy mass spectrometry data, annotated through consensus across multiple database search engines, and systematically evaluates recent deep learning-based de novo sequencing tools alongside traditional approaches. The framework establishes consistent evaluation metrics and data splits to enable fair comparison and reproducible benchmarking.

Why it matters: De novo peptide sequencing is the only way to identify peptides not present in protein sequence databases, making it essential for discovering tumour-specific neoantigens, peptides from microbial or viral origins, and post-translational modifications that alter peptide mass. As deep learning models for de novo sequencing proliferate, the lack of standardised benchmarking has made it difficult to assess real progress. DLDN-Bench addresses this by providing a rigorous, reproducible evaluation framework with well-characterised ground truth data. The focus on human muscle tissue data provides a clinically relevant benchmark, and the multi-search-engine consensus annotation ensures high-quality ground truth.

Why for Yiru: Personalised cancer vaccines and neoantigen discovery depend on accurate identification of tumour-specific peptides presented by MHC molecules. De novo sequencing from immunopeptidomics data — where peptides are eluted from MHC molecules and identified by mass spectrometry — is a critical computational step in this pipeline. DLDN-Bench provides a framework for evaluating which de novo sequencing tools perform best on the types of peptides relevant to TME immunopeptidomics. Benchmarking on consistently processed, well-annotated data is essential before applying these tools to clinically critical neoantigen discovery tasks.