This research article from David Baker team details a novel approach for systematic and accurate prediction of protein-protein interactions (PPIs) in the human proteome. The core innovation lies in the development of omicMSAs, which are deeply enriched multiple sequence alignments derived from petabytes of untapped eukaryotic genomic data to enhance coevolutionary signals, and RoseTTAFold2-PPI (RF2-PPI), a rapid deep learning network. This new network was trained on an extensive dataset, including novel domain-domain interactions (DDIs) distilled from over 200 million predicted protein structures, making it much more efficient and accurate for high-throughput screening than previous methods like AlphaFold2. Using this pipeline, the researchers predicted over 17,849 high-confidence human PPIs, including thousands previously unknown interactions, offering significant new insights into biological mechanisms and disease.

References:

• Zhang J, Humphreys I R, Pei J, et al. Predicting protein-protein interactions in the human proteome[J]. Science, 2025: eadt1630.

This article from Nature Genetics presents a multi-omic study using single-cell and spatial transcriptomics to understand resistance to immuno-chemotherapy in non-small-cell lung cancer (NSCLC). Researchers analyzed tumor samples from patients before and after treatment to identify cellular and molecular factors associated with poor outcomes. A key finding reveals that collagen type XI alpha 1 chain-positive (COL11A1+) cancer-associated fibroblasts (CAFs) and secreted phosphoprotein 1-positive (SPP1+) macrophages accumulate at tumor boundaries, forming a structure that physically blocks T cell infiltration. Additionally, the study characterizes different maturation states of tertiary lymphoid structures (TLSs), noting that activated TLSs are linked to improved prognosis, while a hypoxic microenvironment appears to suppress their beneficial development in non-responders. Overall, the research suggests several potential biomarkers and therapeutic targets aimed at overcoming treatment resistance by modulating the tumor microenvironment.

References:

• Yan Y, Sun D, Hu J, et al. Multi-omic profiling highlights factors associated with resistance to immuno-chemotherapy in non-small-cell lung cancer[J]. Nature Genetics, 2025, 57(1): 126-139.

This article details the development and validation of TORCH, a deep-learning model designed to predict the tumor origin in cancers of unknown primary (CUP) using cytological images from serous effusions (hydrothorax and ascites). The study leveraged a massive dataset of over 57,000 cases to train this artificial intelligence tool, which performed robustly across multiple internal and external testing sets, achieving high Area Under the Receiver Operating Curve (AUROC) values for both cancer diagnosis and origin localization. Notably, TORCH significantly outperformed human pathologists in prediction accuracy and dramatically improved junior pathologists' diagnostic scores when used as an ancillary tool. Furthermore, the research established a link between treatment concordance based on TORCH's predictions and improved overall patient survival, underscoring the model’s potential value in clinical practice for personalized CUP management.

References:

• Tian F, Liu D, Wei N, et al. Prediction of tumor origin in cancers of unknown primary origin with cytology-based deep learning[J]. Nature Medicine, 2024, 30(5): 1309-1319.

This paper introduces TOOLUNIVERSE, a comprehensive, open-source ecosystem designed to facilitate the creation of specialized AI scientists by integrating large language models and reasoning models with a vast array of scientific tools. This platform addresses the current difficulty of building bespoke AI research systems by standardizing tool interaction through an AI-tool interaction protocol and offering over 600 resources, including machine learning models, databases, and APIs. Key components such as the Tool Discover and Tool Optimizer enable the system to automatically generate new tools and refine existing tool specifications, going beyond simple tool orchestration. A case study on hypercholesterolemia demonstrates TOOLUNIVERSE's capacity to drive therapeutic discovery, moving from initial target identification to the validation of optimized drug candidates. Ultimately, the ecosystem aims to democratize the development of sophisticated AI research assistants and accelerate discovery across scientific domains.

References:

• Gao S, Zhu R, Sui P, et al. Democratizing AI scientists using ToolUniverse[J]. arXiv preprint arXiv:2509.23426, 2025.

This paper provides an extensive perspective on the emerging role of Artificial Intelligence (AI) agents in transforming biomedical discovery, envisioning systems capable of autonomous, skeptical learning and reasoning, referred to as "AI scientists." It details the evolution of data-driven models from databases to interactive learning and, finally, to current multi-agent systems, outlining various levels of autonomy for AI agents in research, from simple assistants to full scientists capable of generating novel hypotheses. The text describes the key modular components necessary for these agents, including perception, interaction, memory, and reasoning, and emphasizes the crucial challenges surrounding their implementation, such as ensuring robustness, reliability, and ethical governance. The authors propose that these agents, through collaboration with humans and the integration of tools like large language models (LLMs) and experimental platforms, are set to significantly accelerate and enhance scientific workflows in areas like genetics and chemical biology.

References:

• Gao S, Fang A, Huang Y, et al. Empowering biomedical discovery with AI agents[J]. Cell, 2024, 187(22): 6125-6151.

This paper is an extensive review article focusing on the role of cytokines in cancer and the challenges and progress in therapeutically targeting these pathways. It explains that cytokines are essential immune cell communication proteins whose effects, which can be anti-tumor (like IFN-g) or pro-oncogenic (like innate inflammatory cytokines), depend on the cell type and environment. The review details the mechanisms of cytokine signaling, primarily through the JAK-STAT pathway, and categorizes various cytokines by their function, including the interferons, the IL-2 family, and innate inflammatory cytokines. Finally, it discusses the history and ongoing efforts in developing and engineering cytokine-based therapies and cytokine-blocking agents for cancer treatment, noting that local delivery and stability are key factors for success due to the systemic toxicities and complexities of these networks.

References:

• Kureshi C T, Dougan S K. Cytokines in cancer[J]. Cancer Cell, 2025, 43(1): 15-35.

This paper provides an extensive review of Chimeric Antigen Receptor (CAR) T cell therapy specifically focusing on its application and challenges in treating solid tumors, contrasting its limited success here with its proven efficacy in hematologic malignancies. It identifies major obstacles, such as antigenic heterogeneity and the immunosuppressive tumor microenvironment (TME), which hinder CAR-T cell performance in solid tumors. The text details various emerging strategies being explored to enhance CAR-T cell function, including genetic engineering to improve T cell persistence and cytotoxicity, modifications to CAR structures, and combination therapies like those with immune checkpoint inhibitors. Furthermore, the review summarizes current clinical trial outcomes for CAR-T therapy across several types of solid tumors, such as lung, gastric, and prostate cancers, while discussing innovations in CAR-T cell manufacturing, including non-viral gene editing and allogeneic approaches.

References:

• Du B, Qin J, Lin B, et al. CAR-T therapy in solid tumors[J]. Cancer Cell, 2025, 43(4): 665-679.

The review article provides a comprehensive overview of the integration of artificial intelligence (AI) across the entire drug development pipeline, which is traditionally costly and time-consuming. It details the application of various AI technologies, including large language models (LLMs) and generative AI, in stages such as target identification, drug discovery, preclinical and clinical studies, and post-market surveillance. The text highlights how AI is being used for complex tasks like virtual screening, de novo molecular design, and predicting ADMET (absorption, distribution, metabolism, excretion, and toxicity) properties, aiming to increase efficiency and success rates. Furthermore, the source examines the current challenges facing AI-augmented drug development, such as data scarcity and interpretability issues, before discussing future research directions in this rapidly evolving field.

References:

• Zhang K, Yang X, Wang Y, et al. Artificial intelligence in drug development[J]. Nature medicine, 2025, 31(1): 45-59.

This paper reports on the discovery of a distinct mechanism driving T cell exhaustion (Tex cells) in the context of chronic infections and cancer. The authors utilized proteomic analysis to identify a unique cellular stress response, termed Tex-PSR (proteotoxic stress response), which is characterized by an unusual increase in global protein synthesis and the accumulation of protein aggregates, despite increased protein degradation. Furthermore, the research establishes that this proteotoxic stress can causally drive T cell exhaustion even without persistent T cell receptor stimulation, and links Tex-PSR to sustained AKT signaling. Crucially, targeting associated Tex-PSR chaperone proteins (like BiP and gp96) in preclinical models was shown to improve cancer immunotherapy outcomes, suggesting a new therapeutic target for overcoming resistance to current treatments like immune checkpoint blockade.

References:

• Wang Y, Ma A, Song N J, et al. Proteotoxic stress response drives T cell exhaustion and immune evasion[J]. Nature, 2025: 1-11.

This paper details a study that tracked clonal evolution and drug resistance in high-grade serous ovarian cancer (HGSOC) patients. Researchers developed a method called CloneSeq-SV, which combines single-cell whole-genome sequencing (scWGS) with targeted deep sequencing of structural variants (SVs) in cell-free DNA (cfDNA) to monitor co-existing tumor populations over time. The study of 18 patients revealed that drug resistance often arises from the selective expansion of pre-existing clones present at diagnosis, which frequently exhibit distinctive genomic features like oncogene amplifications. Furthermore, SVs proved to be highly specific markers for tracking tumor DNA in cfDNA, offering an advantageous signal-to-noise ratio compared to single nucleotide variants (SNVs), and demonstrating the potential of this approach to inform future evolution-informed adaptive treatment regimens.

References:

• Williams M J, Vázquez-García I, Tam G, et al. Tracking clonal evolution during treatment in ovarian cancer using cell-free DNA[J]. Nature, 2025: 1-9.