Research Radar — 2026-05-19

Summary: Introduces MIDAS (Mining Immunotherapy Drug tArgetS), a multimodal graph neural network system for immuno-oncology target discovery that integrates gene interactions, multi-omic patient profiles, immune cell biology, antigen processing, disease associations, and genetic perturbation phenotypes. MIDAS generalizes to time-sliced data, outcompetes state-of-the-art baselines including OpenTargets, ranks approved targets above those in clinical development, and recovers immunotherapy-response-associated genes in unseen patients. Functional perturbation of oncostatin M–oncostatin M receptor signaling, a proposed MIDAS target, in TRACERx melanoma patient-derived explants reduced dysfunctional CD8+ T cells and CCL4 levels — consistent with oncostatin M modulating the TME toward immunosuppressive phenotypes.

Why it matters: Current immunotherapy benefits only a minority of patients, and novel target discovery remains slow and expensive. MIDAS represents a systematic computational framework that not only predicts targets but validates them in clinically relevant patient-derived explants — bridging the gap between computational prediction and translational immunology with direct therapeutic implications.

Why for Yiru: Immuno-oncology target discovery and TME modulation are directly aligned with Boss's research interests. The patient-derived explant validation approach and the identification of oncostatin M as a TME modulator connect to interests in spatial TME biology and immunotherapy resistance mechanisms.

Summary: Presents SpecGP, a transformer-based deep learning model for predicting the structural spectra of intact N-glycopeptides at different collision energies. Glycopeptide spectra are inherently complex and high-dimensional due to the combinatorial diversity of glycan structures and peptide backbones. SpecGP addresses this challenge by learning energy-adaptable spectral representations, enabling high-throughput spectral library construction for glycoproteomics without requiring exhaustive experimental acquisition.

Why it matters: Glycoproteomics is a critical but technically challenging dimension of proteomics — glycans regulate protein function, immune recognition, and disease progression. Accurate in silico spectral prediction dramatically reduces the experimental burden of glycoproteomics and enables systematic glycoproteome profiling at scale.

Why for Yiru: Post-translational modifications including glycosylation are important regulators of immune recognition and cell signaling in the TME. Methods that expand the analytical toolkit for glycoproteomics support comprehensive molecular characterization of tumor and immune cell states.

Summary: Addresses a critical problem in AI-driven protein design: generative models produce plausible-looking protein binders that are actually non-functional hallucinations. Proposes a zero-shot, ensemble-driven statistical physics scoring framework that rescues true binders from AI-generated pools by evaluating conformational ensemble properties rather than single-structure metrics. Demonstrates that ensemble-based scoring separates genuine binders from hallucinated ones where conventional single-structure metrics fail.

Why it matters: AI hallucination in protein design wastes enormous experimental resources — researchers spend months testing AI-designed proteins that never had a chance of working. A computational filter that distinguishes real binders from hallucinations without requiring experimental data could dramatically improve the efficiency of AI-guided protein engineering pipelines.

Why for Yiru: Protein design and binder engineering are relevant to developing therapeutic proteins, biosensors, and research tools for TME biology. The hallucination problem mirrors broader challenges in AI for biology where generative models produce plausible but non-functional outputs.

Summary: Demonstrates that Checkpoint kinase 1 (Chk1) inhibitor combination treatment reprograms tumour-associated macrophages (TAMs) from an immune-suppressive M2-like state to an inflammatory, anti-tumour M1-like state. The study characterizes the molecular mechanism by which Chk1 inhibition reshapes the TAM transcriptional and functional landscape, and shows that TAM reprogramming contributes to the anti-tumour efficacy of Chk1 inhibitor combinations in preclinical models, adding an immune component to what was previously considered a purely tumour-cell-intrinsic therapeutic strategy.

Why it matters: Chk1 inhibitors are in clinical development primarily for their tumour-cell-intrinsic effects on DNA damage response. The discovery that they also reprogram TAMs toward anti-tumour phenotypes adds an unexpected immunological dimension and suggests combination strategies with immunotherapy that leverage both tumour-cell and immune-cell effects.

Why for Yiru: TAM reprogramming is a major therapeutic axis in the TME. Understanding how existing clinical-stage drugs like Chk1 inhibitors reshape macrophage states could inform rational combination immunotherapy design and spatial analysis of treatment response.

Summary: Explores the use of ESM3, a state-of-the-art protein language model, for guided generation of knotted protein topologies — a class of protein folds previously considered extremely challenging for computational design. Provides topological insights into how deep learning models navigate the complex fold space of knotted proteins and demonstrates successful generation of diverse knotted architectures with validated structural properties.

Why it matters: Knotted proteins represent a frontier in protein design — their complex topologies confer unique stability and functional properties but have been largely inaccessible to rational design. Demonstrating that generative models can navigate this space expands the repertoire of designable protein architectures for therapeutic and industrial applications.

Why for Yiru: Protein design and engineering are relevant to developing novel biologics and understanding structure-function relationships in molecular recognition. The topological dimension of protein design connects to broader interests in biomolecular structure and function.

Summary: Systematically evaluates open-source large language models (LLMs) for their ability to orchestrate multi-step agentic analysis workflows in a typical biomedical lab setting. Tests models on real-world bioinformatics tasks including data preprocessing, statistical analysis, visualization generation, and result interpretation. Identifies strengths and failure modes of open LLMs compared to proprietary alternatives for autonomous biomedical data analysis.

Why it matters: As AI agents become more integrated into research workflows, understanding which models can reliably orchestrate biomedical analyses is critical. This evaluation provides practical guidance for labs considering deploying open LLMs for automated analysis pipelines and highlights where human oversight remains essential.

Why for Yiru: AI agents for biomedical analysis are directly relevant to automating and scaling computational research workflows. Understanding the capabilities and limitations of open LLMs informs decisions about tool adoption in Boss's research environment.