15 May 2026 in Nature Communications
Post-translational modifications (PTMs) form a complex combinatorial “code” that orchestrates protein function and cellular signaling. However, deciphering this code by predicting PTM sites and linking sites to their regulatory enzymes remains a fundamental challenge. Here, we present COMPASS-PTM, a mechanism-aware, coarse-to-fine learning framework that unifies residue-level multi-label PTM prediction with enzyme-substrate assignment by jointly modeling PTM patterns and their catalytic regulators. COMPASS-PTM builds upon protein language models, integrating physicochemical descriptors and a crosstalk-aware prompting mechanism to learn biologically coherent patterns of cooperative and antagonistic modifications, while addressing the dual long-tail distribution inherent in PTM data. Across multiple proteome-scale benchmarks, COMPASS-PTM improves over the strongest evaluated baselines, with a 122% relative improvement in F1-score for multi-label site prediction and a 54% gain in zero-shot enzyme assignment. Furthermore, the model demonstrates interpretable generalization, recovering canonical kinase motifs and mechanistically linking missense variants to both local PTM disruptions and global rewiring of enzyme-substrate networks. By coupling statistical learning with explicit biochemical knowledge, COMPASS-PTM unifies site-level and enzyme-level prediction into a single framework that learns the grammar underlying protein regulation and signaling.
8 February 2024 in bioRxiv
Cells are the fundamental structural and functional units of life. Studying the definition and composition of different cell types can help us understand the complex mechanisms underlying biological diversity and functionality. The increasing volume of extensive single-cell omics data makes it possible to provide detailed characterisations of cell types. Recently, there has been a rise in deep learning-based approaches that generate cell type labels solely through mapping query data to reference data. However, these approaches lack multi-scale descriptions and interpretations of identified cell types. Here, we propose Cell Decoder, a biological prior knowledge informed model to achieve multi-scale representation of cells. We implemented automated machine learning and post-hoc analysis techniques to decode cell identity. We have shown that Cell Decoder compares favourably to existing methods, offering multi-view interpretability for decoding cell identity and data integration. Furthermore, we have showcased its applicability in uncovering novel cell types and states in both human bone and mouse embryonic contexts, thereby revealing the multi-scale heterogeneity inherent in cell identities.
14 September 2026 in Nature Communications
Single-cell sequencing offers profound insights into tumor complexity. However, reliably linking individual cellular profiles to overall sample-level phenotypes remains computationally challenging, particularly when cell-specific annotations are absent. This disconnect limits the technology’s utility in studying underlying tumor mechanisms. Here we show how scTransMIL, an artificial intelligence framework, bridges this gap. By conceptualizing a biological sample as a collection of individual cells, scTransMIL employs a transformer-based multi-instance learning approach to directly connect single-cell transcriptomic profiles with overarching sample-level labels. We demonstrate that scTransMIL accurately predicts sample-level cancer phenotypes, including the tissue-of-origin for metastatic tumors. At cellular resolution, it robustly identifies tumor-associated cell populations and maps biological progression trajectories using only minimal sample-level annotations. Furthermore, the model’s attention mechanisms facilitate full-transcriptome biomarker discovery. By systematically integrating molecular and cellular scales with broader phenotypic states, our approach provides a computational tool to dissect tumor heterogeneity and advance fundamental cancer biology.