Deep Generative Models Research Articles

Generating Synthetic Raman Spectra of DMMP and 2-CEES by Mathematical Transforms and Deep Generative Models

To build an automated system detecting toxic chemicals from Raman spectra, we have to obtain sufficient data of toxic chemicals. However, it usually costs high to gather Raman spectra of toxic chemicals in diverse situations. Tackling this problem, we develop methods to generate synthetic Raman spectra of DMMP and 2-CEES without actual experiments. First, we propose certain mathematical transforms to augment few original Raman spectra. Then, we train deep generative models to generate more realistic and diverse data. Analyzing synthetic Raman spectra of toxic chemicals generated by our methods through visualization, we qualitatively verify that the data are sufficiently similar to original data and diverse. For conclusion, we obtain a synthetic dataset of DMMP and 2-CEES with the proposed algorithm.

Knowledge-based In Silico Fragmentation and Annotation of Mass Spectra for Natural Products with MassKG

Liquid chromatography coupled with tandem mass spectrometry (LC-MS/MS) is a potent analytical technique utilized for identifying natural products from complex sources. However, due to the structural diversity, annotating LC-MS/MS data of natural products efficiently remains challenging, hindering the discovery process of novel active structures. Here, we introduce MassKG, an algorithm that combines a knowledge-based fragmentation strategy and a deep learning-based molecule generation model to aid in rapid dereplication and the discovery of novel NP structures. Specifically, MassKG has compiled 407,720 known NP structures and, based on this, generated 266,353 new structures using chemical language models for the discovery of potential novel compounds. Furthermore, MassKG demonstrates exceptional performance in spectra annotation compared to state-of-the-art algorithms. To enhance usability, MassKG has been implemented as a web server for annotating tandem mass spectral data (MS/MS, MS2) with a user-friendly interface, automatic reporting, and fragment tree visualization. Lastly, the interpretive capability of MassKG is comprehensively validated through composition analysis and MS annotation of Panax notoginseng, Ginkgo Biloba, Codonopsis pilosula, and Astragalus membranaceus. MassKG is now accessible at https://xomics.com.cn/masskg.

De novo protein design in the age of artificial intelligence

Proteins with specific functions and characteristics play a crucial role in biomedicine and nanotechnology. De novo protein design enables the customization of sequences to produce proteins with desired structures that do not exist in the nature. In recent years, with the rapid development of artificial intelligence (AI), deep learning-based generative models have increasingly become powerful tools, enabling the design of functional proteins with atomic-level precision. This article provides an overview of the evolution of de novo protein design, with focus on the latest algorithmic models, and then analyzes existing challenges such as low design success rates, insufficient accuracy, and dependence on experimental validation. Furthermore, this article discusses the future trends in protein design, aiming to provide insights for researchers and practitioners in this field.

NetGSR: Towards Efficient and Reliable Network Monitoring with Generative Super Resolution

Network monitoring systems are a key building block in today's networks. They all follow a common framework where measurement data from network elements is aggregated at a central collector for network-wide visibility. When designing network monitoring systems, two key properties have to be taken into account: (1) efficiency, to minimize the communication overhead from network elements to the collector; (2) high-fidelity, to faithfully represent the network status. However, in presence of network dynamics, tracking the right operating point to ensure both high fidelity and efficiency is hard and we observe that prior monitoring approaches trade off one for the other. In this paper, we show that it is possible to satisfy both these properties with NetGSR, a new deep learning based solution we introduce that reconstructs the fine-grained behavior of network status at the collector while requiring low resolution measurement data from network elements. This is achieved through a combination of a new custom-tailored conditional deep generative model (DistilGAN), and a new feedback mechanism (Xaminer) based on model uncertainty estimation and denoising that allows the collector to adjust the sampling rate for measurement data from network elements, at run-time. We extensively evaluate NetGSR using three different network scenarios with corresponding real-world network monitoring datasets as well as two downstream use cases. We show that NetGSR can faithfully reconstruct fine-grained network status with 25x greater measurement efficiency than prior approaches while requiring only few ms of inference time at the collector.

Leveraging a deep learning generative model to enhance recognition of minor asphalt defects

Deep learning-based computer vision systems have become powerful tools for automated and cost-effective pavement distress detection, essential for efficient road maintenance. Current methods focus primarily on developing supervised learning architectures, which are limited by the scarcity of annotated image datasets. The use of data augmentation with synthetic images created by generative models to improve these supervised systems is not widely explored. The few studies that do focus on generative architectures are mostly non-conditional, requiring extra labeling, and typically address only road crack defects while aiming to improve classification models rather than object detection. This study introduces AsphaltGAN, a novel class-conditional Generative Adversarial Network with attention mechanisms, designed to augment datasets with various rare road defects to enhance object detection. An in-depth analysis evaluates the impact of different loss functions and hyperparameter tuning. The optimized AsphaltGAN outperforms state-of-the-art generative architectures on public datasets. Additionally, a new workflow is proposed to improve object detection models using synthetic road images. The augmented datasets significantly improve the object detection metrics of You Only Look Once version 8 by 33.0\\documentclass[12pt]{minimal} \\usepackage{amsmath} \\usepackage{wasysym} \\usepackage{amsfonts} \\usepackage{amssymb} \\usepackage{amsbsy} \\usepackage{mathrsfs} \\usepackage{upgreek} \\setlength{\\oddsidemargin}{-69pt} \\begin{document}$$\\%$$\\end{document}, 3.8\\documentclass[12pt]{minimal} \\usepackage{amsmath} \\usepackage{wasysym} \\usepackage{amsfonts} \\usepackage{amssymb} \\usepackage{amsbsy} \\usepackage{mathrsfs} \\usepackage{upgreek} \\setlength{\\oddsidemargin}{-69pt} \\begin{document}$$\\%$$\\end{document}, 46.3\\documentclass[12pt]{minimal} \\usepackage{amsmath} \\usepackage{wasysym} \\usepackage{amsfonts} \\usepackage{amssymb} \\usepackage{amsbsy} \\usepackage{mathrsfs} \\usepackage{upgreek} \\setlength{\\oddsidemargin}{-69pt} \\begin{document}$$\\%$$\\end{document}, and 51.8\\documentclass[12pt]{minimal} \\usepackage{amsmath} \\usepackage{wasysym} \\usepackage{amsfonts} \\usepackage{amssymb} \\usepackage{amsbsy} \\usepackage{mathrsfs} \\usepackage{upgreek} \\setlength{\\oddsidemargin}{-69pt} \\begin{document}$$\\%$$\\end{document} on the Road Damage Detection 2022 dataset, Crack Dataset, Asphalt Pavement Detection Dataset, and Crack Surface Dataset, respectively.

Deep generative AI models analyzing circulating orphan non-coding RNAs enable detection of early-stage lung cancer

Liquid biopsies have the potential to revolutionize cancer care through non-invasive early detection of tumors. Developing a robust liquid biopsy test requires collecting high-dimensional data from a large number of blood samples across heterogeneous groups of patients. We propose that the generative capability of variational auto-encoders enables learning a robust and generalizable signature of blood-based biomarkers. In this study, we analyze orphan non-coding RNAs (oncRNAs) from serum samples of 1050 individuals diagnosed with non-small cell lung cancer (NSCLC) at various stages, as well as sex-, age-, and BMI-matched controls. We demonstrate that our multi-task generative AI model, Orion, surpasses commonly used methods in both overall performance and generalizability to held-out datasets. Orion achieves an overall sensitivity of 94% (95% CI: 87%–98%) at 87% (95% CI: 81%–93%) specificity for cancer detection across all stages, outperforming the sensitivity of other methods on held-out validation datasets by more than ~ 30%.

MultiDGD: A versatile deep generative model for multi-omics data

Recent technological advancements in single-cell genomics have enabled joint profiling of gene expression and alternative modalities at unprecedented scale. Consequently, the complexity of multi-omics data sets is increasing massively. Existing models for multi-modal data are typically limited in functionality or scalability, making data integration and downstream analysis cumbersome. We present multiDGD, a scalable deep generative model providing a probabilistic framework to learn shared representations of transcriptome and chromatin accessibility. It shows outstanding performance on data reconstruction without feature selection. We demonstrate on several data sets from human and mouse that multiDGD learns well-clustered joint representations. We further find that probabilistic modeling of sample covariates enables post-hoc data integration without the need for fine-tuning. Additionally, we show that multiDGD can detect statistical associations between genes and regulatory regions conditioned on the learned representations. multiDGD is available as an scverse-compatible package on GitHub.

Electrical resistivity tomography inversion combining deep variational autoencoder and stochastic adaptive sampling

Geophysical methods, such as electrical resistivity tomography (ERT), can be used to imaging the near-surface electrical resistivity as field measurements depend on the subsurface porosity, water saturation and fluid salinity. ERT has been widely applied to investigate mineral and groundwater resources, and in archaeological, environmental, and engineering studies. The prediction of subsurface electrical conductivity from ERT data requires solving a geophysical inverse problem. For near-surface characterization studies, this is often accomplished with deterministic inverse methods. These methods linearize the problem around an initial solution, and their smoothness depends on an imposed a priori spatial regularization term. Depending on this parameterization, these methodologies might struggle to capture the natural variability of the subsurface. Moreover, deterministic solutions have limited capabilities for uncertainty assessment. On the contrary, stochastic inverse methods can assess uncertainties by predicting multiple model realizations that fit similarly the recorded ERT data. However, they are often more computationally expensive than deterministic solutions. Deep learning algorithms based on deep generative models have been used to re-parametrize model and data spaces into low-dimensional domains and efficiently solve geophysical inverse problems. However, within this context, uncertainty assessment is challenging. We propose a deep convolutional variational autoencoder (VAE) coupled with stochastic adaptive optimization to perform stochastic ERT inversion. Geostatistical simulations of electrical resistivity are used as training data set of the VAE. After training, the VAE generates electrical resistivity models that reproduce the statistics and spatial continuity patterns of the training data set. Then, the VAE latent space is iteratively perturbed and updated with adaptive stochastic sampling based on the misfit between observed and predicted ERT data. The proposed methodology is illustrated in two-dimensional synthetic and real data sets to illustrate the ability of the proposed method to predict reliable electrical resistivity models while generating multiple possible scenarios for uncertainty assessment.

Unveiling the nonclassicality within quasi-distribution representations through deep learning

Abstract To unequivocally distinguish genuine quantumness from classicality, a widely adopted approach focuses on the negative values of a quasi-distribution representation as compelling evidence of nonclassicality. Prominent examples include the dynamical process nonclassicality characterized by the canonical Hamiltonian ensemble representation (CHER) and the nonclassicality of quantum states
characterized by the Wigner function. However, to construct a multivariate joint quasi-distribution function with negative values from experimental data is typically highly cumbersome. Here we propose a computational approach utilizing a deep generative model, processing three marginals, to construct the bivariate joint quasi-distribution functions. We first apply our model to tackle the
challenging problem of the CHERs, which lacks universal solutions, rendering the problem ground-truth (GT) deficient. To overcome the GT deficiency of the CHER problem, we design optimal synthetic datasets to train our model. While trained with synthetic data, the physics-informed optimization enables our model to capture the detrimental effect of the thermal fluctuations on nonclassicality, which cannot be obtained from any analytical solutions. This underscores the reliability of our approach. This approach also allows us to predict the Wigner functions subject to thermal noises. Our model predicts the Wigner functions with a prominent accuracy by processing three marginals of probability distributions. Our approach also provides a significant reduction of the experimental efforts of
constructing the Wigner functions of quantum states, giving rise to an efficient alternative way to realize the quantum state tomography.

VI-VS: calibrated identification of feature dependencies in single-cell multiomics

Unveiling functional relationships between various molecular cell phenotypes from data using machine learning models is a key promise of multiomics. Existing methods either use flexible but hard-to-interpret models or simpler, misspecified models. VI-VS (Variational Inference for Variable Selection) balances flexibility and interpretability to identify relevant feature relationships in multiomic data. It uses deep generative models to identify conditionally dependent features, with false discovery rate control. VI-VS is available as an open-source Python package, providing a robust solution to identify features more likely representing genuine causal relationships.

A physics-guided deep generative model for predicting melt pool behavior in laser powder bed fusion additive manufacturing

A physics-guided deep generative model for predicting melt pool behavior in laser powder bed fusion additive manufacturing

Deep learning generative model for crystal structure prediction

Recent advances in deep learning generative models (GMs) have created high capabilities in accessing and assessing complex high-dimensional data, allowing superior efficiency in navigating vast material configuration space in search of viable structures. Coupling such capabilities with physically significant data to construct trained models for materials discovery is crucial to moving this emerging field forward. Here, we present a universal GM for crystal structure prediction (CSP) via a conditional crystal diffusion variational autoencoder (Cond-CDVAE) approach, which is tailored to allow user-defined material and physical parameters such as composition and pressure. This model is trained on an expansive dataset containing over 670,000 local minimum structures, including a rich spectrum of high-pressure structures, along with ambient-pressure structures in Materials Project database. We demonstrate that the Cond-CDVAE model can generate physically plausible structures with high fidelity under diverse pressure conditions without necessitating local optimization, accurately predicting 59.3% of the 3547 unseen ambient-pressure experimental structures within 800 structure samplings, with the accuracy rate climbing to 83.2% for structures comprising fewer than 20 atoms per unit cell. These results meet or exceed those achieved via conventional CSP methods based on global optimization. The present findings showcase substantial potential of GMs in the realm of CSP.

DigFrag as a digital fragmentation method used for artificial intelligence-based drug design

Fragment-Based Drug Design (FBDD) plays a pivotal role in the field of drug discovery and development. The construction of high-quality fragment libraries is a critical step in FBDD. Conventional fragmentation approaches often rely on rigid rules and chemical intuition, limiting their adaptability to diverse molecular structures. The rapid development of Artificial Intelligence (AI) technology offers a transformative opportunity to rethink traditional methods. Here, we present DigFrag, a digital fragmentation method that highlights important substructures by focusing locally within the molecular graph. In addition, we feed the fragments segmented by machine intelligence and human expertise into the deep generative model to compare the preference for data from different sources. Experimental results show that the structural diversity of fragments segmented by DigFrag is higher, and more desirable compounds are generated based on these fragments. These results also demonstrate that data generated based on AI methods may be more suitable for AI models. Moreover, a user-friendly platform called MolFrag (https://dpai.ccnu.edu.cn/MolFrag/) is developed based on various fragmentation techniques to support molecular segmentation.

A deep learning approach for rational ligand generation with toxicity control via reactive building blocks.

Deep generative models are gaining attention in the field of de novo drug design. However, the rational design of ligand molecules for novel targets remains challenging, particularly in controlling the properties of the generated molecules. Here, inspired by the DNA-encoded compound library technique, we introduce DeepBlock, a deep learning approach for block-based ligand generation tailored to target protein sequences while enabling precise property control. DeepBlock neatly divides the generation process into two steps: building blocks generation and molecule reconstruction, accomplished by a neural network and a rule-based reconstruction algorithm we proposed, respectively. Furthermore, DeepBlock synergizes the optimization algorithm and deep learning to regulate the properties of the generated molecules. Experiments show that DeepBlock outperforms existing methods in generating ligands with affinity, synthetic accessibility and drug likeness. Moreover, when integrated with simulated annealing or Bayesian optimization using toxicity as the optimization objective, DeepBlock successfully generates ligands with low toxicity while preserving affinity with the target.

Blending is all you need: Data-centric ensemble synthetic data

Deep generative models have gained increasing popularity, particularly in fields such as natural language processing and computer vision. Recently, efforts have been made to extend these advanced algorithms to tabular data. While generative models have shown promising results in creating synthetic data, their high computational demands and the need for careful parameter tuning present significant challenges. This study investigates whether a collective integration of refined synthetic datasets from multiple models can achieve comparable or superior performance to that of a single, large generative model. To this end, we developed a Data-Centric Ensemble Synthetic Data model, leveraging principles of ensemble learning. Our approach involved a data refinement process applied to various synthetic datasets, systematically eliminating noise and ranking, selecting, and combining them to create an augmented, high-quality synthetic dataset. This approach improved both the quantity and quality of the data. Central to this process, we introduced the Ensemble k-Nearest Neighbors with Centroid Displacement (EKCD) algorithm for noise filtering, alongside a density score for ranking and selecting data. Our experiments confirmed the effectiveness of EKCD in removing noisy synthetic samples. Additionally, the ensemble model based on the refined synthetic data substantially enhanced the performance of machine learning models, sometimes even outperforming that of real data.

Self-Supervised Data-Driven Approach Defines Pathological High-Frequency Oscillations in Human.

Interictal high-frequency oscillations (HFOs) are a promising neurophysiological biomarker of the epileptogenic zone (EZ). However, objective criteria for distinguishing pathological from physiological HFOs remain elusive, hindering clinical application. We investigated whether the distinct mechanisms underlying pathological and physiological HFOs are encapsulated in their signal morphology in intracranial EEG (iEEG) recordings and whether this mechanism-driven distinction could be simulated by a deep generative model. In a retrospective cohort of 185 epilepsy patients who underwent iEEG monitoring, we analyzed 686,410 HFOs across 18,265 brain contacts. To learn morphological characteristics, each event was transformed into a time-frequency plot and input into a variational autoencoder. We characterized latent space clusters containing morphologically defined putative pathological HFOs (mpHFOs) using interpretability analysis, including latent space disentanglement and time-domain perturbation. mpHFOs showed strong associations with expert-defined spikes and were predominantly located within the seizure onset zone (SOZ). Discovered novel pathological features included high power in the gamma (30-80 Hz) and ripple (>80 Hz) bands centered on the event. These characteristics were consistent across multiple variables, including institution, electrode type, and patient demographics. Predicting 12-month postoperative seizure outcomes using the resection ratio of mpHFOs outperformed unclassified HFOs (F1=0.72 vs. 0.68) and matched current clinical standards using SOZ resection (F1=0.74). Combining mpHFO data with demographic and SOZ resection status further improved prediction accuracy (F1=0.83). Our data-driven approach yielded a novel, explainable definition of pathological HFOs, which has the potential to further enhance the clinical use of HFOs for EZ delineation.

Joint representation and visualization of derailed cell states with Decipher.

Biological insights often depend on comparing conditions such as disease and health, yet we lack effective computational tools for integrating single-cell genomics data across conditions or characterizing transitions from normal to deviant cell states. Here, we present Decipher, a deep generative model that characterizes derailed cell-state trajectories. Decipher jointly models and visualizes gene expression and cell state from normal and perturbed single-cell RNA-seq data, revealing shared and disrupted dynamics. We demonstrate its superior performance across diverse contexts, including in pancreatitis with oncogene mutation, acute myeloid leukemia, and gastric cancer.

Simulating images of radio galaxies with diffusion models

Context. With increasing amounts of data produced by astronomical surveys, automated analysis methods have become crucial. Synthetic data are required for developing and testing such methods. Current classical approaches to simulations often suffer from insufficient detail or inaccurate representation of source type occurrences. Deep generative modeling has emerged as a novel way of synthesizing realistic image data to overcome those deficiencies. Aims. We implemented a deep generative model trained on observations to generate realistic radio galaxy images with full control over the flux and source morphology. Methods. We used a diffusion model, trained with continuous time steps to reduce sampling time without quality impairments. The two models were trained on two different datasets, respectively. One set was a selection of images obtained from the second data release of the LOFAR Two-Metre Sky Survey (LoTSS). The model was conditioned on peak flux values to preserve signal intensity information after re-scaling image pixel values. The other, smaller set was obtained from the Very Large Array (VLA) survey of Faint Images of the Radio Sky at Twenty-Centimeters (FIRST). In that set, every image was provided with a morphological class label the corresponding model was conditioned on. Conditioned sampling is realized with classifier-free diffusion guidance. We evaluated the quality of generated images by comparing the distributions of different quantities over the real and generated data, including results from the standard source-finding algorithms. The class conditioning was evaluated by training a classifier and comparing its performance on both real and generated data. Results. We have been able to generate realistic images of high quality using 25 sampling steps, which is unprecedented in the field of radio astronomy. The generated images are visually indistinguishable from the training data and the distributions of different image metrics were successfully replicated. The classifier is shown to perform equally well for real and generated images, indicating strong sampling control over morphological source properties.

Event-horizon-scale Imaging of M87* under Different Assumptions via Deep Generative Image Priors

Reconstructing images from the Event Horizon Telescope (EHT) observations of M87*, the supermassive black hole at the center of the galaxy M87, depends on a prior to impose desired image statistics. However, given the impossibility of directly observing black holes, there is no clear choice for a prior. We present a framework for flexibly designing a range of priors, each bringing different biases to the image reconstruction. These priors can be weak (e.g., impose only basic natural-image statistics) or strong (e.g., impose assumptions of black hole structure). Our framework uses Bayesian inference with score-based priors, which are data-driven priors arising from a deep generative model that can learn complicated image distributions. Using our Bayesian imaging approach with sophisticated data-driven priors, we can assess how visual features and uncertainty of reconstructed images change depending on the prior. In addition to simulated data, we image the real EHT M87* data and discuss how recovered features are influenced by the choice of prior.

Unveiling the spatial-temporal dynamics: Diffusion-based learning of conditional distribution for range-dependent ocean sound speed field forecasting.

Environment-aware underwater acoustic detection and communications demand precise forecasting of the sound speed field (SSF) both temporally and spatially. Toward this goal, recent machine learning models, such as recurrent neural networks and Gaussian process regressions, have outperformed classical autoregressive models. However, from the unified theoretical perspective of conditional distribution learning, there is still significant room for improvement, as existing works have not fully learned the conditional distribution of future SSFs given past SSFs. To address these limitations, in this paper, we leverage the use of diffusion models, the foundation of recent successful deep generative models, such as DALL-E 2 and SORA, to learn the conditional distribution even under limited training data, through careful neural architecture and training strategy design. Our experiments, conducted on real-life South China Sea datasets, confirm that our proposed model outperforms the state-of-the-art baselines in forecasting range-dependent SSFs and the associated underwater transmission losses. Additionally, our model provides reliable confidence intervals that quantify the uncertainties of predictions.