Nature Machine Intelligence最新文献

Automated detection of tumour lesions on positron emission tomography–computed tomography (PET/CT) image data is a clinically relevant but highly challenging task. Progress in this field has been hampered in the past owing to the lack of publicly available annotated data and limited availability of platforms for inter-institutional collaboration. Here we describe the results of the autoPET challenge, a biomedical image analysis challenge aimed to motivate research in the field of automated PET/CT image analysis. The challenge task was the automated segmentation of metabolically active tumour lesions on whole-body ¹⁸F-fluorodeoxyglucose PET/CT. Challenge participants had access to a large publicly available annotated PET/CT dataset for algorithm training. All algorithms submitted to the final challenge phase were based on deep learning methods, mostly using three-dimensional U-Net architectures. Submitted algorithms were evaluated on a private test set composed of 150 PET/CT studies from two institutions. An ensemble model of the highest-ranking algorithms achieved favourable performance compared with individual algorithms. Algorithm performance was dependent on the quality and quantity of data and on algorithm design choices, such as tailored post-processing of predicted segmentations. Future iterations of this challenge will focus on generalization and clinical translation.

Gene expression involves transcription and translation. Despite large datasets and increasingly powerful methods devoted to calculating genetic variants’ effects on transcription, discrepancy between messenger RNA and protein levels hinders the systematic interpretation of the regulatory effects of disease-associated variants. Accurate models of the sequence determinants of translation are needed to close this gap and to interpret disease-associated variants that act on translation. Here we present Translatomer, a multimodal transformer framework that predicts cell-type-specific translation from messenger RNA expression and gene sequence. We train the Translatomer on 33 tissues and cell lines, and show that the inclusion of sequence improves the prediction of ribosome profiling signal, indicating that the Translatomer captures sequence-dependent translational regulatory information. The Translatomer achieves accuracies of 0.72 to 0.80 for the de novo prediction of cell-type-specific ribosome profiling. We develop an in silico mutagenesis tool to estimate mutational effects on translation and demonstrate that variants associated with translation regulation are evolutionarily constrained, both in the human population and across species. In particular, we identify cell-type-specific translational regulatory mechanisms independent of the expression quantitative trait loci for 3,041 non-coding and synonymous variants associated with complex diseases, including Alzheimer’s disease, schizophrenia and congenital heart disease. The Translatomer accurately models the genetic underpinnings of translation, bridging the gap between messenger RNA and protein levels as well as providing valuable mechanistic insights for uninterpreted disease variants.

Understanding the mechanisms of T cell antigen recognition that underpin adaptive immune responses is critical for developing vaccines, immunotherapies and treatments against autoimmune diseases. Despite extensive research efforts, accurate prediction of T cell receptor (TCR)–antigen binding pairs remains a great challenge due to the vast diversity and cross-reactivity of TCRs. Here we propose a deep-learning-based framework termed epitope-anchored contrastive transfer learning (EPACT) tailored to paired human CD8⁺ TCRs. Harnessing the pretrained representations and co-embeddings of peptide–major histocompatibility complex (pMHC) and TCR, EPACT demonstrated generalizability in predicting binding specificity for unseen epitopes and distinct TCR repertoires. Contrastive learning enabled highly precise predictions for immunodominant epitopes and interpretable analysis of epitope-specific T cells. We applied EPACT to SARS-CoV-2-responsive T cells, and the predicted binding strength aligned well with the surge in spike-specific immune responses after vaccination. We further fine-tuned EPACT on structural data to decipher the residue-level interactions involved in TCR–antigen recognition. EPACT was capable of quantifying interchain distance matrices and identifying contact residues, corroborating the presence of TCR cross-reactivity across multiple tumour-associated antigens. Together, EPACT can serve as a useful artificial intelligence approach with important potential in practical applications and contribute towards the development of TCR-based immunotherapies.

Optimizing a candidate molecule’s physiochemical and functional properties has been a critical task in drug and material design. Although the non-trivial task of balancing multiple (potentially conflicting) optimization objectives is considered ideal for artificial intelligence, several technical challenges such as the scarcity of multiproperty-labelled training data have hindered the development of a satisfactory AI solution for a long time. Prompt-MolOpt is a tool for molecular optimization; it makes use of prompt-based embeddings, as used in large language models, to improve the transformer’s ability to optimize molecules for specific property adjustments. Notably, Prompt-MolOpt excels in working with limited multiproperty data (even under the zero-shot setting) by effectively generalizing causal relationships learned from single-property datasets. In comparative evaluations against established models such as JTNN, hierG2G and Modof, Prompt-MolOpt achieves over a 15% relative improvement in multiproperty optimization success rates compared with the leading Modof model. Furthermore, a variant of Prompt-MolOpt, named Prompt-MolOpt^P, can preserve the pharmacophores or any user-specified fragments under the structural transformation, further broadening its application scope. By constructing tailored optimization datasets, with the protocol introduced in this work, Prompt-MolOpt steers molecular optimization towards domain-relevant chemical spaces, enhancing the quality of the optimized molecules. Real-world tests, such as those involving blood–brain barrier permeability optimization, underscore its practical relevance. Prompt-MolOpt offers a versatile approach for multiproperty and multi-site molecular optimizations, suggesting its potential utility in chemistry research and drug and material discovery.

Neural networks suffer from spectral bias and have difficulty representing the high-frequency components of a function, whereas relaxation methods can resolve high frequencies efficiently but stall at moderate to low frequencies. We exploit the weaknesses of the two approaches by combining them synergistically to develop a fast numerical solver of partial differential equations (PDEs) at scale. Specifically, we propose HINTS, a hybrid, iterative, numerical and transferable solver by integrating a Deep Operator Network (DeepONet) with standard relaxation methods, leading to parallel efficiency and algorithmic scalability for a wide class of PDEs, not tractable with existing monolithic solvers. HINTS balances the convergence behaviour across the spectrum of eigenmodes by utilizing the spectral bias of DeepONet, resulting in a uniform convergence rate and hence exceptional performance of the hybrid solver overall. Moreover, HINTS applies to large-scale, multidimensional systems; it is flexible with regards to discretizations, computational domain and boundary conditions; and it can also be used to precondition Krylov methods.