Xiang Pan

T-cell receptor (TCR) recognition of antigenic peptides presented by human leukocyte antigen (HLA) molecules underpins adaptive immunity and T cell-based immunotherapy. However, the scarcity of complete interaction data and TCR cross-reactivity challenge robust prediction. Here, we present UpTCR, a unified progressive knowledge-transfer foundation model that learns from incomplete data to predict TCR-antigen-HLA binding. UpTCR progressively transfers knowledge from dimeric and trimeric interactions to tetrameric complexes and uses soft contrastive learning to mitigate false negatives. It outperforms existing methods in predicting TCR binding specificity and antigen-HLA binding affinity, particularly for neoantigens, and reveals pairwise residue-level interactions across the tetramer. UpTCR also transfers effectively to breast cancer cohorts with limited data. Prospective validation against melanoma antigen variants identifies eight immunogenic peptides that elicit T cell responses and one variant associated with immune escape. These findings establish UpTCR as a generalizable and interpretable tool for studying antigen recognition and advancing TCR-based immunotherapies.

Motivation: The scalable identification of bioactive compounds is essential for contemporary drug discovery. This process faces a key trade-off: structural screening offers scalability but lacks biological context, whereas high-content phenotypic profiling provides deep biological insights but is resource-intensive. The primary challenge is to extract robust biological signals from noisy data and encode them into representations that do not require biological data at inference. Results: This study presents DECODE (DEcomposing Cellular Observations of Drug Effects), a framework that bridges this gap by empowering chemical representations with intrinsic biological semantics to enable structure-based in silico biological profiling. DECODE leverages limited paired transcriptomic and morphological data as supervisory signals during training, enabling the extraction of a measurement-invariant biological fingerprint from chemical structures and explicit filtering of modality-specific variation. Across held-out retrieval, scaffold-split and UMAP-clustering virtual-screening benchmarks, DECODE improves functional retrieval and early active-compound prioritization over baselines. Availability and implementation: The codes and datasets of DECODE are available at https://github.com/lian-xiao/DECODE.

This study introduces an efficient Automated Breast Volume Scanner (ABVS) image analysis framework designed to address the limitations of traditional screening methods-namely low diagnostic accuracy, excessive reliance on operator expertise, and the increased workload associated with manual interpretation of three-dimensional imaging. Our primary objective was to develop a single, end-to-end network that simultaneously performs lesion segmentation and benign-malignant classification using a multi-task learning strategy. The proposed model integrates several innovations. A boundary-aware auxiliary task and a semantic guidance module were designed to enhance the correlation between accurate boundary delineation and classification. We also developed a hybrid weighted attention mechanism that combines spatial and channel-wise information, improving the model’s ability to identify lesions against complex tissue backgrounds. Our findings show that this integrated approach significantly improves performance in both segmentation and classification. By leveraging the intrinsic relationship between the two tasks, the model achieves more precise lesion capture and feature extraction. This framework offers a practical and effective automated solution for clinical breast lesion assessment, enhancing diagnostic accuracy and efficiency while reducing the interpretive burden on radiologists.

Rock thin-section image analysis is a fundamental task in geological and mineralogical research. Traditional methods rely on visual inspection by experts using optical microscopes, which are inherently subjective, experience-dependent, and time-consuming. Here, we introduce RoImAI, a vision foundation model specifically designed for rock thin-section microscopy images, enabling rapid and precise rock segmentation, identification, and lithology report generation. A large-scale dataset of rock thin-section microscopy images comprising 30,336 images and approximately two million rock particles from 17 different regions was constructed to develop and validate RoImAI. RoImAI leverages Transformer-based deep learning techniques to achieve high-precision segmentation across multi-center datasets from diverse geological regions. RoImAI employs a hierarchical classification strategy to identify rock particles accurately. Furthermore, RoImAI outperforms human experts in efficiency and accuracy when generating structured lithology reports. The intelligent analytical capabilities and high accuracy of RoImAI strongly enable the automated processing of the rapidly expanding volume of rock thin-section images.