Deep Model Research Articles

Channel reconstruction through improvised deep learning architecture for high-speed networks

Efficient acquisition of channel state information (CSI) is quite complicated process but immensely essential to exploit probable benefits of massive multiple input multiple output (MIMO) systems. Therefore, a deep learningbased model is utilized to estimate channel feedback in a massive MIMO system. The proposed improvised deep learning-based channel estimation (IDLCE) model enhances channel reconstruction efficiency by using multiple convolutional layers and residual blocks. The proposed IDLCE model utilizes encoder network to compress CSI matrices where decoder network is used to downlink reconstruct CSI matrices. Here, an additional quantization block is incorporated to improve feedback reconstruction accuracy by reducing channel errors. A COST 2,100 model is adopted to analyse performance efficiency for both indoor and outdoor scenarios. Further, deep learning-based model is used to train thousands of parameter and correlation coefficients much faster and to minimize computational complexity. The proposed IDLCE model evaluate performance in terms of normalized mean square error (NMSE), correlation efficiency and reconstruction accuracy and compared against varied state-of-art-channel estimation techniques. Excellent performance results are obtained with large improvement in channel reconstruction accuracy

Weakly supervised deep learning-based prognostic model for AI-assisted prospective clinical investigation of cervical cancer

Weakly supervised deep learning-based prognostic model for AI-assisted prospective clinical investigation of cervical cancer

Phishing website detection using novel integration of BERT and XLNet with deep learning sequential models

Phishing website detection using novel integration of BERT and XLNet with deep learning sequential models

Modeling the occupancy rate and nights spent at tourism accommodations of Spain using autoregressive deep learning networks

Purpose The purpose of this paper is to autoregressively model the net occupancy rate of beds and bedrooms in hotels and similar accommodations and the nights spent at these accommodations of Spain for the period of 1990–2023 using monthly data. Design/methodology/approach The monthly occupancy rate of hotels and the total number of hotel nights data of Spain for the 1990M01–2023M09 range is considered. An autoregressive deep learning network is developed for the modeling of both metrics. Moreover, the results of the proposed autoregressive deep learning method are compared to those of a classical artificial neural network. Findings The actual occupancy rate, total night data and the deep learning model results are compared showing the accuracy of the developed model. Moreover, the R2, mean absolute error, root mean square error and mean absolute percentage error of the models are calculated further demonstrating the high performance of the developed model. The R2 values higher than 0.9 are achieved for both occupancy rate and total number of hotel nights data. Practical implications The modeling results given in this paper demonstrate that the previous values of the net occupancy rate and the total number of nights can be used as inputs of a deep learning network model by which accurate forecasts can be made for the future values of the occupancy rate and the total number of hotel nights. This modeling approach possesses importance from the practical viewpoint considering that the accurate planning and forecast of the net occupancy rate and the total number of nights affect the tourism income. Originality/value This study differs from existing literature by attempting to model the occupancy rate and the total number of hotel nights data autoregressively using deep learning networks.

CLAP. I. Resolving miscalibration for deep learning-based galaxy photometric redshift estimation

Obtaining well-calibrated photometric redshift probability densities for galaxies without a spectroscopic measurement remains a challenge. Deep learning discriminative models, typically fed with multi-band galaxy images, can produce outputs that mimic probability densities and achieve state-of-the-art accuracy. However, several previous studies have found that such models may be affected by miscalibration, an issue that would result in discrepancies between the model outputs and the actual distributions of true redshifts. Our work develops a novel method called the C ontrastive L earning and A daptive KNN for P hotometric Redshift (CLAP) that resolves this issue. It leverages supervised contrastive learning (SCL) and $k$-nearest neighbours (KNN) to construct and calibrate raw probability density estimates, and implements a refitting procedure to resume end-to-end discriminative models ready to produce final estimates for large-scale imaging data, bypassing the intensive computation required for KNN. The harmonic mean is adopted to combine an ensemble of estimates from multiple realisations for improving accuracy. Our experiments demonstrate that CLAP takes advantage of both deep learning and KNN, outperforming benchmark methods on the calibration of probability density estimates and retaining high accuracy and computational efficiency. With reference to CLAP, a deeper investigation on miscalibration for conventional deep learning is presented. We point out that miscalibration is particularly sensitive to the method-induced excessive correlations among data instances in addition to the unaccounted-for epistemic uncertainties. Reducing the uncertainties may not guarantee the removal of miscalibration due to the presence of such excessive correlations, yet this is a problem for conventional methods rather than CLAP. These discussions underscore the robustness of CLAP for obtaining photometric redshift probability densities required by astrophysical and cosmological applications. This is the first paper in our series on CLAP.

An interpretable deep learning model for detecting BRCA pathogenic variants of breast cancer from hematoxylin and eosin-stained pathological images

Background Determining the status of breast cancer susceptibility genes (BRCA) is crucial for guiding breast cancer treatment. Nevertheless, the need for BRCA genetic testing among breast cancer patients remains unmet due to high costs and limited resources. This study aimed to develop a Bi-directional Self-Attention Multiple Instance Learning (BiAMIL) algorithm to detect BRCA status from hematoxylin and eosin (H&E) pathological images. Methods A total of 319 histopathological slides from 254 breast cancer patients were included, comprising two dependent cohorts. Following image pre-processing, 633,484 tumor tiles from the training dataset were employed to train the self-developed deep-learning model. The performance of the network was evaluated in the internal and external test sets. Results BiAMIL achieved AUC values of 0.819 (95% CI [0.673–0.965]) in the internal test set, and 0.817 (95% CI [0.712–0.923]) in the external test set. To explore the relationship between BRCA status and interpretable morphological features in pathological images, we utilized Class Activation Mapping (CAM) technique and cluster analysis to investigate the connections between BRCA gene mutation status and tissue and cell features. Significantly, we observed that tumor-infiltrating lymphocytes and the morphological characteristics of tumor cells appeared to be potential features associated with BRCA status. Conclusions An interpretable deep neural network model based on the attention mechanism was developed to predict the BRCA status in breast cancer. Keywords: Breast cancer, BRCA, deep learning, self-attention, interpretability.

A comprehensive comparison of deep learning-based compound-target interaction prediction models to unveil guiding design principles

The evaluation of compound-target interactions (CTIs) is at the heart of drug discovery efforts. Given the substantial time and monetary costs of classical experimental screening, significant efforts have been dedicated to develop deep learning-based models that can accurately predict CTIs. A comprehensive comparison of these models on a large, curated CTI dataset is, however, still lacking. Here, we perform an in-depth comparison of 12 state-of-the-art deep learning architectures that use different protein and compound representations. The models were selected for their reported performance and architectures. To reliably compare model performance, we curated over 300 thousand binding and non-binding CTIs and established several gold-standard datasets of varying size and information. Based on our findings, DeepConv-DTI consistently outperforms other models in CTI prediction performance across the majority of datasets. It achieves an MCC of 0.6 or higher for most of the datasets and is one of the fastest models in training and inference. These results indicate that utilizing convolutional-based windows as in DeepConv-DTI to traverse trainable embeddings is a highly effective approach for capturing informative protein features. We also observed that physicochemical embeddings of targets increased model performance. We therefore modified DeepConv-DTI to include normalized physicochemical properties, which resulted in the overall best performing model Phys-DeepConv-DTI. This work highlights how the systematic evaluation of input features of compounds and targets, as well as their corresponding neural network architectures, can serve as a roadmap for the future development of improved CTI models.Scientific contributionThis work features comprehensive CTI datasets to allow for the objective comparison and benchmarking of CTI prediction algorithms. Based on this dataset, we gained insights into which embeddings of compounds and targets and which deep learning-based algorithms perform best, providing a blueprint for the future development of CTI algorithms. Using the insights gained from this screen, we provide a novel CTI algorithm with state-of-the-art performance.

Multi-Step-Ahead Time Series Forecasting using Deep Learning and Fuzzy Time Series-based Error Correction Method

Recently time series forecasting has become one of the prime application areas of climatology, economics and industries. Many research works are conducted to forecast the time series more accurately. But few of them are concentrated on predicting the time series over an extended future horizon, and there is also a scope to improve their forecasting accuracy. This work proposes a multi-step-ahead foresting method to produce a stable and accurate forecasting result for the extended future horizon. Firstly, a deep learning-based forecasting model is proposed to predict the time series. Secondly, a fuzzy time series-based error correction model is implemented to enhance the prediction performance of the deep learning model. Here to optimize all the fuzzy time series (FTS) parameters in an integrated way, an integrated butterfly optimization (FTS-IBO) algorithm is proposed. In this study, two different types of real-world multivariate time series datasets are used to analyze the forecasting performance of the proposed model. The performance of the proposed FTS-IBO algorithm is compared with the traditional butterfly optimization (FTS-BO) algorithm. The experimental results show that the FTS-IBO technique is superior to the FTS-BO technique. The forecasting performance of the proposed model has also compared the other state-of-the-art models, and the simulation results exhibit that the proposed model produces a more accurate prediction performance for multi-step-ahead time series forecasting problems compared to other models.

Deep learning-based high-frequency jump test for detecting stock market manipulation: evidence from China’s securities market

PurposeThis study aims to tackle the critical issue of detecting stock market manipulation, which undermines the integrity and stability of financial markets globally. Even enhanced with machine learning, traditional statistical methods often struggle to analyze high-frequency trading data effectively due to inherent noise and the limited availability of publicly known manipulation cases. This leads to poor model generalization and a tendency toward over-fitting. Focusing on China's securities market, our study introduces an innovative approach that employs deep learning-based high-frequency jump tests to overcome these challenges and to develop a more effective method for identifying manipulative activities.Design/methodology/approachWe employed the “Jump Variation – Time-of-Day” (JV-TOD) non-parametric technique for jump tests on high-frequency data, coupled with the synthetic minority over-sampling technique (SMOTE) algorithm for re-balancing sample data. Our approach trains a deep neural network (DNN) on refined data to enhance its ability to identify manipulation patterns accurately.FindingsOur results show that the deep neural network model, calibrated with high-frequency price jump data, identifies manipulation behavior more specifically and accurately than traditional models. The model achieved an accuracy rate of 94.64%, an F1-score of 95.26% and a recall rate of 95.88%, significantly outperforming traditional models. These results demonstrate the effectiveness of our approach in mitigating over-fitting and improving the robustness of market manipulation detection.Practical implicationsThe proposed model provides regulatory entities and financial institutions with a more efficient tool to monitor and counteract market manipulation, thereby improving market fairness and investor protection.Originality/valueBy integrating the JV-TOD jump test with deep learning, this study proposed a new approach to market manipulation detection. The innovation is in its capacity to detect subtle manipulation signals that traditional methods typically overlook. Our model, which is trained on jump test data enhanced by the SMOTE algorithm, excels at learning complex manipulation patterns. This enhances both detection accuracy and robustness. In contrast to existing methods that are challenged by the noisy and intricate nature of high-frequency data, our approach shows enhanced performance in identifying nuanced market manipulations, offering a more effective and reliable method for detecting market manipulation.

Modeling takeover time for non-driving related tasks using hybrid deep neural networks in conditionally automated vehicles

To enhance the safety of the transition from autonomous to manual driving control, one of the key factors is quantifying the safe takeover transition, namely the takeover time. Previous studies have focused on factors influencing takeover time and predicting takeover time using traditional methods, but there is a lack of research utilizing deep learning networks and driver physiological factors for takeover time prediction. Toward this end, we propose a hybrid deep learning network model dedicated to predicting drivers’ takeover time with a dataset obtained from a bench experiment involving 46 drivers. Firstly, to address the potential information loss commonly associated with traditional manual feature extraction techniques, a convolutional neural network (CNN) is developed to automatically extract physiological spatial representations from drivers. Additionally, a double-layer bidirectional long short-term memory network (DBiLSTM) is incorporated to model the physiological time representation dimension, enabling a more refined representation of physiological signals. Then, to tackle the complexities introduced by individual driver variations, a self-attention mechanism is enabling the model to automatically prioritize critical features, thereby enhancing the accuracy and reliability of its predictions. The results demonstrate that the proposed approach surpasses traditional machine learning network models in predictive accuracy for takeover time prediction, thereby exhibiting novelty in the field of vehicle takeover control research. This provides a more reliable technical foundation for anticipating the durations of vehicle takeovers.

Predicting Obesity via Deep Learning: The Role of Chrononutrition

Abstract Background Obesity is a significant public health challenge, prompting the need for assessing effective tools for its monitoring and prevention. Chrononutrition investigates the role of timing, regularity and frequency of food intake on body metabolism and health. Here we aim to check if data on the distribution of calorie intake in the day may improve BMI prediction as compared to using only data on total day intake. Methods We apply deep learning on chrononutritional data on calorie intake over 6 daily timeslots from a nationally representative cross-sectional survey sample of the Italian population (INRAN-SCAI 2005/2006) comprising 2313 Italian adults 18-64 ys old with 3 day diet diaries. In particular, we implemented three deep neural network models, varying network configurations and feature selection. The 1st includes 18 (6 for each day diary) chrononutritional variables and the 3 day total intakes, beside age and sex; the 2nd includes age, sex and the 18 chrononutritional variables only; the 3rd uses age, sex and the 3 day totals only. We applied Early Stopping to limit overfitting. Optimization was conducted using the Adam algorithm, minimizing the mean squared error (RMSE), while mean absolute error (MAE) was the primary performance indicator. Results The results highlight that Model 3 outperformed the other models with MAE=2.514 and RMSE=3.476, indicating that for the purpose of predicting obesity, information on the time-of-day distribution of calorie intake is not increasing predictive ability of BMI on top of age, sex and total calorie intake in the day. Conclusions Additional use of day-time intake information worsened prediction of BMI via deep learning using this cross sectional Italian survey. However the error made is still sizeable and warrants the inclusion of other variables; our results don’t rule out the role of chrononutrition on metabolic health which may act through other biological mechanisms related to specific nutrients. Key messages • Chrononutrition data don’t improve BMI prediction over total calorie intake data in a cross-setional representative sample of the Italian population. • Deep learning approach to BMI prediction shows the importance of targeted selection of features in public health nutrition.

A Novel Natural Language Processing Model Transfer Strategy Tailored for Deep Learning Platforms

As artificial intelligence technology continues to evolve, numerous advancements enable smoother communication and collaboration between humans and computers. Natural language processing (NLP), as a key direction in the field of computer and artificial intelligence, has been attracting much attention. It focuses on the language used by people in daily life, devoting itself to improving the ability of language generation and understanding, thus making communication between people and computers more fluent and natural, effectively breaking through the problem of human-computer interaction. With the advancement of deep learning technology, the migration of NLP models onto deep learning platforms has emerged as a key trend. Model transfer, a crucial aspect of deep learning, holds significant value in enhancing the performance and efficiency of NLP models. This paper begins by outlining the fundamentals of deep learning platforms and NLP model transfer, followed by a comprehensive examination of current research progress and challenges in this field. Subsequently, we introduce a novel NLP model transfer strategy tailored for deep learning platforms and validate its effectiveness through rigorous experiments. In conclusion, the paper highlights our noteworthy advancements, pointing toward promising future developments.

Retraction Note: Deep belief network-based probabilistic generative model for detection of robotic manipulator failure execution

Retraction Note: Deep belief network-based probabilistic generative model for detection of robotic manipulator failure execution

Detecting GNSS Spoofing and Re-localization on UAV based on Imagery Matching

Abstract Due to rapid advancements in GNSS, computer, and microelectronics technologies, there is a growing popularity and widespread promotion of high-performance, cost-effective intelligent UAVs across various applications. Recently, GNSS spoofing attacks have emerged as a significant obstacle hindering the long-term development of UAVs. UAVs rely heavily on unprotected GNSS signals for navigation, making them highly vulnerable to spoofing. This paper presents an algorithm that employs image matching for detecting GNSS spoofing and re-localization on UAVs using a deep learning methodology. This method functions autonomously solely reliance on camera and does not require alterations to the antenna configuration. Utilizing a camera-equipped UAV, we evaluate the likeness between real-time aerial photographs and satellite imagery by leveraging the position information provided by UAV. By identifying disparities between images taken by a spoofing affected drone and authentic ones, the spoofing can be identified using deep neural network models. Upon detecting spoofing, this paper presents a vision-based re-localization method for UAVs. Experimental results demonstrate an approximately 88.4% accuracy of our model in detecting GNSS spoofing attacks within 100 milliseconds and 88% success rate of re-localization with the accuracy of less than 15 meters. Our algorithm exclusively depends on publicly accessible satellite imagery, offering an intelligent and efficient approach for detecting UAV GNSS spoofing and re-localization in GNSS denied environment.

Detecting Unrecognized Dementia Using Deep Learning Methods with Korean National Claims Data

Abstract Background Dementia is one of the leading causes of death in elderly people. In Korea, the government supports dementia screening for people aged 60 and older, but it is not a population-based organized screening program. It is difficult to achieve early diagnosis and early treatment. The objective of this study is to detect unrecognized dementia using Korean National Claims Data for early diagnosis. Methods This study used the National Health Insurance Claims data in Korea. The case group is the new dementia visit as an unrecognized group, while the control group was randomly selected from the general population without dementia. The predictors included health utilization, socioeconomic and demographic, procedure codes, diagnostic codes and health screenings data. The health utilization variables included length of stay, the number of outpatient visits, health expenditures and other encounter information. All medical utilization data was aggregated monthly. In the case group, 12 months of data from 6-month or 2-years before the onset of dementia were used. In the control group, 12 months of data from randomly selected time point were included. Transformer, embedding methods, and Time2Vec were used as the deep learning methods. This study included 453,306 incidents of dementia, representing almost all cases reported from 2010 to 2019 in Korea. The control group consisted of 669,873. Results After training the deep learning prediction model, the AUC for 6-month dementia prediction on the test dataset was 0.87, and for 2-year dementia prediction, it was 0.76. The most important variables were health utilization variables, especially health expenditures. The health utilization trajectories of the case group were significantly different from those of the control group. Conclusions Based on the claims data, we could detect unrecognized dementia case. We can extend the model to other chronic disease such as cancer and stroke, which are significant burden of disease. Key messages • Screening in healthcare facilities or using lab test results would be more accurate, but it would be difficult to apply to a large population due to the cost-effectiveness and timeliness constraints. • On the other hand, the detection model using the existing claims data would be helpful to to develop cost-effective mass screening program. It can cover the weakness of current screening program.

Left atrial appendage volume estimated by artificial intelligence predicts atrial fibrillation recurrence after cryobaloon ablation

Abstract Background/Introduction The volume of the left atrial appendage (LAA), as measured in contrast-enhanced cardiac computed tomography (CCTA), has been previously reported as a predictor of long-term recurrence of atrial fibrillation (AF) following cryoballoon ablation procedures. Unfortunately manual assessment of LAA volume, is tedious and time-consuming, making it impractical for clinical practice. The recent advent of artificial intelligence (AI)-based methods offers a potential solution to this limitation. Purpose This study aims to evaluate the utility of LAA volume, as automatically measured by a state-of-the-art AI model for the fully automatic segmentation of LAA in CCTA, in predicting AF recurrence following cryoballoon ablation procedures. Methods We retrospectively analyzed a cohort of patients who underwent cryoballoon ablation for pulmonary vein isolation at our facility between 2010 and 2016. AF recurrence was determined through in-person visits and Holter monitoring at 3, 6, and 12 months post-ablation, followed by annual visits. The first three months post-ablation were considered a blanking period. An open-source deep neural network AI model [1] was utilized to perform automatic segmentation and calculation of LAA and left atrium (LA) volumes in. Receiver-operating characteristic curves were used to obtain optimal cut-off values for increased LAA and LA volumes. We then assessed the association of increased LAA and LA volume, and the LAA to LA volume ratio with recurrence-free survival using the Cox regression model. Results The cohort consisted of 212 patients (median age 62, inter-quartile range [IQR]: 55–65), with 25% having persistent AF before the procedure. Subjects were observed for a median period of 12 months (IQR: 12–36), with 70 of them (33%) experiencing a documented recurrent AF episode. The optimal thresholds for predicting AF recurrence were established as 9.8ml for LAA volume, 124ml for LA volume, and 0.086 for the LAA to LA volume ratio, respectively. An increased LAA volume demonstrated the strongest association with AF recurrence (hazard ratio [HR]: 2.6, 95% confidence interval [CI]: 1.55–4.2), followed by an increased LA volume (HR: 2.1, 95% CI: 1.3–3.4), persistence of AF (HR: 2, 95% CI: 1.2–3.3), and increased LAA to LA volume ratio (HR: 1.9, 95% CI: 1.1–3). Kaplan-Meier plots illustrating AF-free survival, are presented in Figure 1. After adjusting for age, persistent AF, left ventricular ejection fraction, and body surface area, all AI measurements remained significantly associated with a higher AF recurrence (Table 1). Conclusions Increased LAA volume, as assessed automatically by AI in CCTA, is a useful predictor of AF recurrence after catheter ablation for atrial fibrillation. The use of freely available, AI-assisted analysis of CCTA could be instrumental in assessing LAA volume, among other predictors, for long-term ablation success.Figure 1.

ADGRU: Adaptive DenseNet with gated recurrent unit for automatic diagnosis of periodontal bone loss and stage periodontitis with tooth segmentation mechanism.

Periodontics and gingivitis are two of the most widely prevalent illnesses that affect people nowadays. The sixth most common disease in the world is periodontitis, and detecting periodontal bone loss is essential in the earlier condition and is crucial for the development of the proper diagnosis. Early bone loss detection can be assisted by using computer-assisted radiography examination. Understanding disease progression helps to select the most effective treatment action. An effective deep model is suggested to detect periodontal bone loss at an earlier stage for preventing the progression of Periodontics bone loss. This work is intimated by collecting images from online resources. Further, the images gathered from the dataset are preceded by the tooth segmentation which is done using DenseUNet + + . Further, the segmented images are given to the Adaptive DenseNet with Gated Recurrent Unit (AD-GRU) for detecting periodontal bone loss and this diagnosis is used for the periodontitis stage, where the ADGRU performance is augmented by optimizing the attributes using the Refined Red Kite Optimization Algorithm (RRKOA). The offered approach attained an accuracy of 94.45% which is higher than the88.63%, 90.58%, 89.54%, and 92.96% attained by the LSTM, DenseNet, GRU, DenseNet-GRU. The findings of the simulation proved the designed framework outperformed the traditional model with high accuracy. The developed effectual deep model-based periodontal bone loss and stage periodontitis diagnosis structure is used in healthcare applications.

Feasibility Analysis for Predicting Indian Ocean Bigeye Tuna (Thunnus obesus) Fishing Grounds Based on Temporal Characteristics of FY-3 Microwave Radiation Imager Data

Efficient and accurate fishery forecasting is of great significance in ensuring the efficiency of fishery operations. This paper proposes a fishery forecasting method using a brightness temperature (TB) time series spatial feature extraction and fusion model. Using Indian Ocean bigeye tuna fishery data from 2009 to 2021 as a reference, this paper discusses the feasibility of fishery forecasting using FY-3 Microwave Radiation Imager (MWRI) Level 1 TB data. For this paper, we designed a deep learning network model for radiometer TB time series feature extraction (TimeTB-FishNet) based on the Convolutional Neural Network (CNN), Gated Recurrent Unit (GRU), and Attention mechanism. After expanding the dimensions of TB features, the model uses them together with spatiotemporal feature factors (year, month, longitude, and latitude) as features. By adding the GRU and Attention to the CNN, the CNN-GRU-Attention model architecture is established and can extract deep time series spatial features from the data to achieve the best results. In the model validation experiments, the TimeTB-FishNet model performed optimally, with a coefficient of determination (R2) of 0.6643. In the generalization experiments, the R2 also reached 0.6261, with a root mean square error (RMSE) of 46.6031 kg/1000 hook. When the sea surface height (SSH) was introduced, the R2 further reached 0.6463, with a lower RMSE of 45.1318 kg/1000 hook. The experimental results show that the proposed method and model are feasible and effective. The proposed model can directly use enhanced radiometer TB data without relying on lagging ocean environmental product data, performing deep temporal and spatial feature extraction for fishery forecasting. This method can provide a reference for the fishing of bigeye tuna in the Indian Ocean.

Generation Method for HVAC Systems Design Schemes in Office Buildings Based on Deep Graph Generative Models

The design process of heating, ventilation, and air conditioning (HVAC) systems is complex and time consuming due to the need to follow design codes. Since the design standards are not fixed, the final outcome often depends on the designer’s experience. The development of building information modeling (BIM) technology has made information throughout the building lifecycle more integrated. BIM-based forward design is now widely used, providing a data foundation for combining HVAC system design with machine learning. This paper proposes an unsupervised learning method based on deep graph generative models to uncover hidden design patterns and optimization strategies from the design results. We trained and validated four deep graph generative models—GAE, GNF, GAN, and diffusion—using HVAC system terminal pipeline layout data. Accuracy and precision metrics were used to compare the generated designs with automated forward design solutions, assessing the models’ ability to capture both local variations and broader changes in design logic. A graph-neural-network-based evaluation method was employed to measure the models’ capacity to detect changes. The results indicate that all four models achieved prediction accuracies exceeding 90% and precision rates above 75%. The models effectively captured both local modifications made by designers and global design changes, showing greater sensitivity to global layout adjustments than to local updates. When comparing the results generated by deep graph generative models and the actual design, it is obvious that the accuracy of the predictions varies significantly due to the complexity of the test buildings.

Predicting transcriptional responses to novel chemical perturbations using deep generative model for drug discovery

Understanding transcriptional responses to chemical perturbations is central to drug discovery, but exhaustive experimental screening of disease-compound combinations is unfeasible. To overcome this limitation, here we introduce PRnet, a perturbation-conditioned deep generative model that predicts transcriptional responses to novel chemical perturbations that have never experimentally perturbed at bulk and single-cell levels. Evaluations indicate that PRnet outperforms alternative methods in predicting responses across novel compounds, pathways, and cell lines. PRnet enables gene-level response interpretation and in-silico drug screening for diseases based on gene signatures. PRnet further identifies and experimentally validates novel compound candidates against small cell lung cancer and colorectal cancer. Lastly, PRnet generates a large-scale integration atlas of perturbation profiles, covering 88 cell lines, 52 tissues, and various compound libraries. PRnet provides a robust and scalable candidate recommendation workflow and successfully recommends drug candidates for 233 diseases. Overall, PRnet is an effective and valuable tool for gene-based therapeutics screening.