Nature Machine Intelligence最新文献

Accurate in silico determination of CD8⁺ T cell epitopes would greatly enhance T cell-based vaccine development, but current prediction models are not reliably successful. Here, motivated by recent successes applying machine learning to complex biology, we curated a dataset of 651,237 unique human leukocyte antigen class I (HLA-I) ligands and developed MUNIS, a deep learning model that identifies peptides presented by HLA-I alleles. MUNIS shows improved performance compared with existing models in predicting peptide presentation and CD8⁺ T cell epitope immunodominance hierarchies. Moreover, application of MUNIS to proteins from Epstein–Barr virus led to successful identification of both established and novel HLA-I epitopes which were experimentally validated by in vitro HLA-I-peptide stability and T cell immunogenicity assays. MUNIS performs comparably to an experimental stability assay in terms of immunogenicity prediction, suggesting that deep learning can reduce experimental burden and accelerate identification of CD8⁺ T cell epitopes for rapid T cell vaccine development.

The immune checkpoint inhibitors have demonstrated promising clinical efficacy across various tumour types, yet the percentage of patients who benefit from them remains low. The bindings between tumour antigens and human leukocyte antigen class I/T cell receptor molecules determine the antigen presentation and T cell activation, thereby playing an important role in the immunotherapy response. In this paper, we propose UnifyImmun, a unified cross-attention transformer model designed to simultaneously predict the bindings of peptides to both receptors, providing more comprehensive evaluation of antigen immunogenicity. We devise a two-phase strategy using virtual adversarial training that enables these two tasks to reinforce each other mutually, by compelling the encoders to extract more expressive features. Our method demonstrates superior performance in predicting both peptide-HLA and peptide-TCR binding on multiple independent and external test sets. Notably, on a large-scale COVID-19 peptide-TCR binding test set without any seen peptide in the training set, our method outperforms the current state-of-the-art methods by more than 10%. The predicted binding scores significantly correlate with the immunotherapy response and clinical outcomes on two clinical cohorts. Furthermore, the cross-attention scores and integrated gradients reveal the amino acid sites critical for peptide binding to receptors. In essence, our approach marks an essential step towards comprehensive evaluation of antigen immunogenicity.

Large language models (LLMs) have become increasingly capable, but their development often requires substantial computational resources. Although model merging has emerged as a cost-effective promising approach for creating new models by combining existing ones, it currently relies on human intuition and domain knowledge, limiting its potential. Here we propose an evolutionary approach that overcomes this limitation by automatically discovering effective combinations of diverse open-source models, harnessing their collective intelligence without requiring extensive additional training data or compute. Our approach operates in both parameter space and data flow space, allowing optimization beyond just the weights of the individual models. This approach even facilitates cross-domain merging, generating models such as a Japanese LLM with math reasoning capabilities. Surprisingly, our Japanese math LLM achieved state-of-the-art performance on a variety of established Japanese LLM benchmarks, even surpassing models with substantially more parameters, despite not being explicitly trained for such tasks. Furthermore, a culturally aware Japanese vision–language model generated through our approach demonstrates its effectiveness in describing Japanese culture-specific content, outperforming previous Japanese vision–language models. This work not only contributes new state-of-the-art models back to the open-source community but also introduces a new paradigm for automated model composition, paving the way for exploring alternative, efficient approaches to foundation model development.

High-throughput omics profiling advancements have greatly enhanced cancer patient stratification. However, incomplete data in multi-omics integration present a substantial challenge, as traditional methods like sample exclusion or imputation often compromise biological diversity and dependencies. Furthermore, the critical task of accurately classifying new patients with partial omics data into existing subtypes is commonly overlooked. To address these issues, we introduce Integrate Any Omics (IntegrAO), an unsupervised framework for integrating incomplete multi-omics data and classifying new samples. IntegrAO first combines partially overlapping patient graphs from diverse omics sources and utilizes graph neural networks to produce unified patient embeddings. Our systematic evaluation across five cancer cohorts involving six omics modalities demonstrates IntegrAO’s robustness to missing data and its accuracy in classifying new samples with partial profiles. An acute myeloid leukaemia case study further validates its capability to uncover biological and clinical heterogeneities in incomplete datasets. IntegrAO’s ability to handle heterogeneous and incomplete data makes it an essential tool for precision oncology, offering a holistic approach to patient characterization.

As artificial intelligence systems, particularly large language models (LLMs), become increasingly integrated into decision-making processes, the ability to trust their outputs is crucial. To earn human trust, LLMs must be well calibrated such that they can accurately assess and communicate the likelihood of their predictions being correct. Whereas recent work has focused on LLMs’ internal confidence, less is understood about how effectively they convey uncertainty to users. Here we explore the calibration gap, which refers to the difference between human confidence in LLM-generated answers and the models’ actual confidence, and the discrimination gap, which reflects how well humans and models can distinguish between correct and incorrect answers. Our experiments with multiple-choice and short-answer questions reveal that users tend to overestimate the accuracy of LLM responses when provided with default explanations. Moreover, longer explanations increased user confidence, even when the extra length did not improve answer accuracy. By adjusting LLM explanations to better reflect the models’ internal confidence, both the calibration gap and the discrimination gap narrowed, significantly improving user perception of LLM accuracy. These findings underscore the importance of accurate uncertainty communication and highlight the effect of explanation length in influencing user trust in artificial-intelligence-assisted decision-making environments.

The increasing frequency of emerging viral infections necessitates a rapid human response, highlighting the cost-effectiveness of computational methods. However, existing computational approaches are limited by their input forms or incomplete functionalities, preventing a unified prediction of diverse virus variation drivers and hindering in-depth applications. To address this issue, we propose a unified evolution-driven framework for predicting virus variation drivers, named Evolution-driven Virus Variation Driver prediction (E2VD), which is guided by virus evolutionary traits. With evolution-inspired design, E2VD comprehensively and significantly outperforms state-of-the-art methods across various virus mutational driver prediction tasks. Moreover, E2VD effectively captures the fundamental patterns of virus evolution. It not only distinguishes different types of mutations but also accurately identifies rare beneficial mutations that are critical for viruses to survive, while maintaining generalization capabilities across different lineages of SARS-CoV-2 and different types of viruses. Importantly, with predicted biological drivers, E2VD perceives virus evolutionary trends in which potential high-risk mutation sites are accurately recommended. Overall, E2VD represents a unified, structure-free and interpretable approach for analysing and predicting viral evolutionary fitness, providing an ideal alternative to costly wet-lab measurements to accelerate responses to emerging viral infections.

Deep learning has significantly advanced molecular modelling and design, enabling an efficient understanding and discovery of novel molecules. In particular, large language models introduce a fresh research paradigm to tackle scientific problems from a natural language processing perspective. Large language models significantly enhance our understanding and generation of molecules, often surpassing existing methods with their capabilities to decode and synthesize complex molecular patterns. However, two key issues remain: how to quantify the match between model and data modalities and how to identify the knowledge-learning preferences of models. To address these challenges, we propose a multimodal benchmark, named ChEBI-20-MM, and perform 1,263 experiments to assess the model’s compatibility with data modalities and knowledge acquisition. Through the modal transition probability matrix, we provide insights into the most suitable modalities for tasks. Furthermore, we introduce a statistically interpretable approach to discover context-specific knowledge mapping by localized feature filtering. Our analysis offers an exploration of the learning mechanism and paves the way for advancing large language models in molecular science.

Foundation models have demonstrated remarkable performance across various tasks, primarily due to their abilities to comprehend instructions and access extensive, high-quality data. These capabilities showcase the effectiveness of current foundation models and suggest a promising trajectory. Owing to multiple constraints, such as the extreme scarcity or inaccessibility of raw data used to train foundation models and the high cost of training large-scale foundation models from scratch, the use of pre-existing foundation models or application programming interfaces for downstream tasks has become a new research trend, which we call Learn from Model (LFM). LFM involves extracting and leveraging prior knowledge from foundation models through fine-tuning, editing and fusion methods and applying it to downstream tasks. We emphasize that maximizing the use of parametric knowledge in data-scarce scenarios is critical to LFM. Analysing the LFM paradigm can guide the selection of the most appropriate technology in a given scenario to minimize parameter storage and computational costs while improving the performance of foundation models on new tasks. This Review provides a comprehensive overview of current methods based on foundation models from the perspective of LFM.

Omics studies produce a large number of measurements, enabling the development, validation and interpretation of systems-level biological models. Large cohorts are required to power these complex models; yet, the cohort size remains limited due to clinical and budgetary constraints. We introduce clinical and omics multimodal analysis enhanced with transfer learning (COMET), a machine learning framework that incorporates large, observational electronic health record databases and transfer learning to improve the analysis of small datasets from omics studies. By pretraining on electronic health record data and adaptively blending both early and late fusion strategies, COMET overcomes the limitations of existing multimodal machine learning methods. Using two independent datasets, we showed that COMET improved the predictive modelling performance and biological discovery compared with the analysis of omics data with traditional methods. By incorporating electronic health record data into omics analyses, COMET enables more precise patient classifications, beyond the simplistic binary reduction to cases and controls. This framework can be broadly applied to the analysis of multimodal omics studies and reveals more powerful biological insights from limited cohort sizes.