Chen, C. Y.-C. | bioRxiv Science

bioRxiv2025

scTransMIL Bridges Patient-Level Phenotypes and Single-Cell Transcriptomics for Cancer Screening and Heterogeneity Inference

Single-cell sequencing technology has revolutionized cancer research by revealing unprecedented insights into tumor heterogeneity. However, reliably connecting patient-level cancer phenotypes with single-cell transcriptomic profiles remains challenging, due to technical constraints and labeling ambiguity. This further hinders the precise cancer screening and intensive study of tumor mechanism based on single-cell sequencing. To bridge this gap, we introduce scTransMIL, a scRNA-seq Transformer-based Multi-Instance Learning framework that learns whole-genome context to deliver comprehensive cancer insights at the sample, cell, and gene levels across three biological scales: (1) accurate patient-level cancer phenotype prediction, (2) precise single-cell disease scoring (validated on 4 million single cells), and (3) genome-wide biomarker discovery. Benchmark experiments demonstrate scTransMILs exceptional performance, including robust out-of-distribution generalization and clinically relevant prediction of metastatic tissue-of-origin, a crucial capability for identifying cancers of unknow primary. At single-cell resolution, scTransMIL identified both known and novel biomarkers for tumor B cells that conventional differential expression analysis failed to detect while maintaining consistent concordance with malignant cell annotations. scTransMILs adaptability is exemplified in acute myelocytic leukemia, where minimal fine-tuning with only a few patient sample labels enabled: (i) discovery of novel disease subtypes with distinct clinical outcomes, (ii) reconstruction of differentiation trajectories at the single-cell resolution, and (iii) identification of subtype-specific gene signatures. In summary, by systematically linking cellular and molecular profiles with clinical disease phenotypes, scTransMIL emerges as a transformative tool poised to advance both basic cancer research and precision oncology applications.

bioinformatics↗

bioRxiv2023

De novo generation of antibody CDRH3 with a pre-trained generative large language model

Artificial Intelligence (AI) techniques have made great advances in assisting antibody design. However, antibody design still heavily relies on isolating antigen-specific antibodies from serum, which is a resource-intensive and time-consuming process. To address this issue, we propose a Pre-trained Antibody generative large Language Model (PALM) for the de novo generation of artificial antibodies heavy chain complementarity-determining region 3 (CDRH3) with desired antigen-binding specificity, reducing the reliance on natural antibodies. We also build a high-precision model antigen-antibody binder (A2binder) that pairs antigen epitope sequences with antibody sequences to predict binding specificity and affinity. PALM-generated antibodies exhibit binding ability to SARS-CoV-2 antigens, including the emerging XBB variant, as confirmed through in-silico analysis and in-vitro assays. The in-vitro assays validated that PALM-generated antibodies achieve high binding affinity and potent neutralization capability against both wild-type and XBB spike proteins of SARS-CoV-2. Meanwhile, A2binder demonstrated exceptional predictive performance on binding specificity for various epitopes and variants. Furthermore, by incorporating the attention mechanism into the PALM model, we have improved its interpretability, providing crucial insights into the fundamental principles of antibody design.

bioinformatics↗