Hongchao Ji

Mass spectrometry (MS) is central to molecular discovery, yet the interpretation of tandem mass spectra (MS/MS) remains limited by database dependence and an incomplete structural resolution. Here, we explore whether large language models (LLMs) can directly translate MS/MS spectra to chemically meaningful molecular descriptions. We introduce MS2LLM, a framework that represents spectra and molecular structures as natural languages and learns their correspondence through instruction tuning. MS2LLM generates hierarchical molecular descriptions, including functional groups, substructures, and chemical classes, directly from the spectral input. Across multiple datasets, it outperforms general-purpose LLMs and conventional spectral learning methods in both descriptive accuracy and chemical classification. Importantly, the model produces interpretable outputs that capture structural semantics rather than exact structures, offering a complementary paradigm for structural inference of unknowns.

Accurate prediction of gas chromatographic Kováts retention indices across diverse stationary phases is crucial for reliable chemical identification. Existing high-performance models often rely on separate architectures for derivatized and underivatized analytes and are frequently restricted to specific column chemistries. In this study, we propose a unified and derivatization-agnostic RI prediction framework capable of supporting three common stationary phases, which are semi-standard non-polar, standard non-polar, and standard polar. Our approach employs a hybrid architecture that integrates an AttentiveFP-based graph neural network with expert-engineered global descriptors, including Morgan fingerprints and a comprehensive suite of physicochemical properties. To address inherent data scarcity in polar and standard phases, we implemented a phase transfer learning strategy where the model is pre-trained on a large-scale semi-standard non-polar dataset of approximately 120,000 entries and subsequently fine-tuned on standard non-polar and standard polar datasets. Evaluation via 10-fold cross-validation demonstrates that our model significantly outperforms current state-of-the-art predictors. On the independent external test set of 500 molecules, our model achieved a mean absolute error of 18.9 RI units for the standard non-polar phase and 51.6 RI units for the standard polar phase, representing a substantial improvement over existing benchmarks. Notably, our unified model maintains high accuracy across both derivatized and underivatized compounds without requiring explicit structural labeling, demonstrating robust generalization across diverse chemical spaces. This work establishes that high-precision and multi-phase retention index prediction is achievable through a single end-to-end architecture, laying a robust algorithmic foundation for broader applications in automated library searching and untargeted metabolomics.

Mass spectrometry (MS) is pivotal for molecular identification in the chemical and biological sciences, yet its practical application is often constrained by sparse spectral libraries, limited software interoperability, and institutional data silos. To address these challenges, we introduce MassLeague, a deployable hybrid annotation software for accurate and collaborative EI-MS compound identification. MassLeague is designed as an integrated software platform featuring two major capabilities: (1) a hybrid annotation engine that combines in silico spectral matching with molecular fingerprint prediction, and (2) a distributed software architecture that supports federated model training and HNSW-based private spectral searching without requiring raw data exchange. In addition to improving annotation accuracy, MassLeague provides an extensible graphical interface, customizable spectral library management, and flexible deployment across local or multi-institutional environments. Our results demonstrate that MassLeague achieves a Top-1 accuracy of 53.74%, outperforming state-of-the-art methods such as FastEI and DeepEI. By integrating analytical performance, software usability, and data privacy into a unified platform, MassLeague provides a practical solution for collaborative EI-MS annotation in metabolomics, toxicology, and drug discovery.

High-throughput proteomics enables detailed molecular phenotyping but poses challenges for predictive modeling and interpretation due to high dimensionality, sparsity, and nonlinear interactions. Biologically informed neural networks (BINNs) address these challenges by embedding pathway knowledge into network architectures, providing interpretable models of complex omics data. We present BINN-FiLM, which extends conventional BINNs by integrating feature-wise linear modulation, enabling task-specific scaling and shifting of pathway-level activations while preserving a fixed Reactome-based hierarchy. Unlike BINNs that share all pathway parameters across classes, BINN-FiLM captures class-dependent biological rewiring, making it particularly effective for multiclass classification tasks. Multiclass problems are reformulated as multitask binary learning, allowing pathway activations to adapt to disease-specific contexts. We evaluated BINN-FiLM on three multiclass proteomic data sets, and BINN-FiLM consistently outperformed conventional BINNs and standard machine learning models. FiLM modulation parameters revealed task-specific pathway activity, and SHAP-based interpretation identified key proteins and pathways driving disease discrimination.

Target deconvolution of bioactive natural products (NPs) is frequently hampered by the inability of conventional thermal shift assays to distinguish direct ligand binding from indirect proteomic stabilization. Here, we developed an integrated target discovery framework that combines matrix-augmented pooling strategy-isothermal shift assay (MAPS-iTSA) thermal proteomics with a structure-aware graph neural network (HoloGNN). This strategy identified ATP-citrate lyase (ACLY) as a high-confidence target of Demethylzeylasteral (DEM), a triterpenoid from Tripterygium wilfordii. Orthogonal biochemical assays confirmed direct binding of DEM to ACLY (surface plasmon resonance, KD = 9.86 μM) and potent enzymatic inhibition (IC ₅₀ = 3.91 μM). To relate this interaction to the pharmacology of the source herb, thermal profiling of Tripterygium wilfordii extract combined with ACLY-based affinity-ultrafiltration mass spectrometry identified DEM as an ACLY-binding constituent of the extract. Given the established role of ACLY-mediated lipid metabolic reprogramming in psoriasis, we next evaluated the pharmacological significance of this interaction. DEM suppressed keratinocyte proliferation in vitro and, in vivo, alleviated imiquimod-induced psoriasiform dermatitis and reduced inflammatory cytokine expression. Single-cell RNA sequencing further revealed reversal of ACLY-SREBP-associated lipogenic reprogramming in keratinocytes. Untargeted lipidomics of lesional skin showed concordant reversal of the psoriatic lipid profile, including de novo synthesized fatty acids and acylcarnitines. Collectively, these findings establish ACLY as a direct and functionally relevant target of DEM and provide a robust AI-chemoproteomic paradigm for mechanism-guided NP drug discovery.

Retention time (RT) is a key parameter in liquid chromatography–mass spectrometry (LC-MS) workflows, supporting compound identification, feature alignment, and quality control. However, traditional RT prediction models are built for specific chromatographic conditions, resulting in fragmented knowledge and limited scalability. We introduce Uni-RT, a unified multitask learning framework that simultaneously learns from heterogeneous data sets to capture both shared molecular retention patterns and condition-specific differences. By leveraging data across multiple chromatographic setups, Uni-RT achieves higher accuracy and robustness than pooled or condition-specific models while greatly simplifying model deployment. Evaluation on 28 reversed-phase liquid chromatography (RPLC) and hydrophilic interaction liquid chromatography (HILIC) data sets demonstrates that multitask learning provides a powerful and generalizable solution for integrating RT prediction into diverse applications.

The escalating crisis of antimicrobial resistance poses a devastating and immediate threat to human life. Antimicrobial peptides (AMPs) are a promising antibiotic substitute to combat antimicrobial resistance. Compared with the traditional wet-lab screening approaches, computational models have largely improved the efficiency of predicting antimicrobial peptide. However, most computational models overlook or underutilize the evolution and structural information of peptides, which is crucial for understanding the peptide functions. Here, we proposed a sophisticated deep learning model to predict AMPs, Antimicrobial Peptide Bilinear Attention Network (AMPBAN), which incorporates peptide evolution features from ESM3 protein language model, structure features from ESMFold predicted with equivariant graph neural network (EGNN), and the joint information from sequence and structure learned via Bilinear Attention Network. AMPBAN consistently demonstrated superior accuracy and generalization compared to nine state-of-the-art AMP prediction models across multiple independent benchmarks. Furthermore, an ablation study confirms that our multimodal fusion strategy significantly refines the integration of sequence and structural signals, yielding superior predictive balance over single-modality models. This framework provides a robust tool for the accelerated discovery of novel AMPs and the advancement of next-generation antimicrobial drug development. The datasets, source code and models are available at https://github.com/baiwenhuim/ampban .

DeepPHSI is expected to be applied to the quality control and online detection of P. cablin , and it is a transferable, universal and potential method for hyperspectral analysis.