Artificial intelligence in the life sciences最新文献

Active machine learning is an established and increasingly popular experimental design technique where the machine learning model can request additional data to improve the model's predictive performance. It is generally assumed that this data is optimal for the machine learning model since it relies on the model's predictions or model architecture and therefore cannot be transferred to other models. Inspired by research in pedagogy, we here introduce the concept of yoked machine learning where a second machine learning model learns from the data selected by another model. We found that in 48% of the benchmarked combinations, yoked learning performed similar or better than active learning. We analyze distinct cases in which yoked learning can improve active learning performance. In particular, we prototype yoked deep learning (YoDeL) where a classic machine learning model provides data to a deep neural network, thereby mitigating challenges of active deep learning such as slow refitting time per learning iteration and poor performance on small datasets. In summary, we expect the new concept of yoked (deep) learning to provide a competitive option to boost the performance of active learning and benefit from distinct capabilities of multiple machine learning models during data acquisition, training, and deployment.

Recent technological advancements have revolutionized research capabilities across the biological sciences by enabling the collection of large data that provides a broader picture of systems from the cellular to ecosystem level at a more refined resolution. The rapid rate of generating these data has exacerbated bottlenecks in study design and data analysis approaches, especially as conventional methods that incorporate traditional statistical tests and assumptions are not suitable or sufficient for highly dimensional data (i.e., more than 1,000 variables). The application of machine learning techniques in large data analysis is one promising solution that is increasingly popular. However, limitations in expertise such that the results from machine learning models can be interpreted to gain meaningful biological insight pose a great challenge. To address this challenge, a user-friendly machine learning workflow that can be applied to a wide variety of data types to reduce these large data to those variables (attributes) most determinant of experimental and/or observed conditions is provided, as well as a general overview of data analysis and machine learning approaches and considerations thereof. The workflow presented here has been beta-tested with great success and is recommended to be incorporated into analysis pipelines of large data as a standardized approach to reduce data dimensionality. Moreover, the workflow is flexible, and the underlying concepts and steps can be modified to best suit user needs, objectives, and study parameters.

The accuracy of machine learning models for protein-ligand binding affinity prediction depends on the quality of the experimental data they are trained on. Most of these models are trained and tested on different subsets of the PDBbind database, which is the main source of protein-ligand complexes with annotated binding affinity in the public domain. However, estimating its experimental uncertainty is not straightforward because just a few protein-ligand complexes have more than one measurement associated. In this work, we analyze bioactivity data from ChEMBL to estimate the experimental uncertainty associated with the three binding affinity measures included in the PDBbind (K_i, K_d, and IC₅₀), as well as the effect of combining them. The experimental uncertainty of combining these three affinity measures was characterized by a mean absolute error of 0.78 logarithmic units, a root mean square error of 1.04 and a Pearson correlation coefficient of 0.76. These estimations were contrasted with the performances obtained by state-of-the-art machine learning models for binding affinity prediction, showing that these models tend to be overoptimistic when evaluated on the core set from PDBbind.

Computational approaches have revolutionized the field of drug discovery, collectively known as Computer-Assisted Drug Design (CADD). Advancements in computing power, data generation, digitalization, and artificial intelligence (AI) techniques have played a crucial role in the rise of CADD. These approaches offer numerous benefits, enabling the analysis and interpretation of vast amounts of data from diverse sources, such as genomics, structural information, and clinical trials data. By integrating and analyzing these multiple data sources, researchers can efficiently identify potential drug targets and develop new drug candidates. Among the AI techniques, machine learning (ML) and deep learning (DL) have shown tremendous promise in drug discovery. ML and DL models can effectively utilize experimental data to accurately predict the efficacy and safety of drug candidates. However, despite these advancements, certain areas in drug discovery face data scarcity, particularly in neglected, rare, and emerging viral diseases. Few-shot learning (FSL) is an emerging approach that addresses the challenge of limited data in drug discovery. FSL enables ML models to learn from a small number of examples of a new task, achieving commendable performance by leveraging knowledge learned from related datasets or prior information. It often involves meta-learning, which trains a model to learn how to learn from few data. This ability to quickly adapt to new tasks with low data circumvents the need for extensive training on large datasets. By enabling efficient learning from a small amount of data, few-shot learning has the potential to accelerate the drug discovery process and enhance the success rate of drug development. In this review, we introduce the concept of few-shot learning and its application in drug discovery. Furthermore, we demonstrate the valuable application of few-shot learning in the identification of new drug targets, accurate prediction of drug efficacy, and the design of novel compounds possessing desired biological properties. This comprehensive review draws upon numerous papers from the literature to provide extensive insights into the effectiveness and potential of few-shot learning in these critical areas of drug discovery and development.

Purpose

The purpose of this study is to develop an accurate deep learning model capable of Inferior Vena Cava (IVC) filter segmentation from CT scans. The study does a comparative assessment of the impact of Residual Networks (ResNets) complemented with reduced convolutional layer depth and also analyzes the impact of using vision transformer architectures without performance degradation.

Materials and Methods

This experimental retrospective study on 84 CT scans consisting of 54618 slices involves design, implementation, and evaluation of segmentation algorithm which can be used to generate a clinical report for the presence of IVC filters on abdominal CT scans performed for any reason. Several variants of patch-based 3D-Convolutional Neural Network (CNN) and the Swin UNet Transformer (Swin-UNETR) are used to retrieve the signature of IVC filters. The Dice Score is used as a metric to compare the performance of the segmentation models.

Results

Model trained on UNet variant using four ResNet layers showed a higher segmentation performance achieving median Dice = 0.92 [Interquartile range(IQR): 0.85, 0.93] compared to the plain UNet model with four layers having median Dice = 0.89 [IQR: 0.83, 0.92]. Segmentation results from ResNet with two layers achieved a median Dice = 0.93 [IQR: 0.87, 0.94] which was higher than the plain UNet model with two layers at median Dice = 0.87 [IQR: 0.77, 0.90]. Models trained using SWIN-based transformers performed significantly better in both training and validation datasets compared to the four CNN variants. The validation median Dice was highest in 4 layer Swin UNETR at 0.88 followed by 2 layer Swin UNETR at 0.85.

Conclusion

Utilization of vision based transformer Swin-UNETR results in segmentation output with both low bias and variance thereby solving a real-world problem within healthcare for advanced Artificial Intelligence (AI) image processing and recognition. The Swin UNETR will reduce the time spent manually tracking IVC filters by centralizing within the electronic health record. Link to GitHub repository.

Imaging plays a fundamental role in the effective diagnosis, staging, management, and monitoring of various cardiac pathologies. Successful radiological analysis relies on accurate image segmentation, a technically arduous process, prone to human-error. To overcome the laborious and time-consuming nature of cardiac image analysis, deep learning approaches have been developed, enabling the accurate, time-efficient, and highly personalised diagnosis, staging and management of cardiac pathologies.

Here, we present a review of over 60 papers, proposing deep learning models for cardiac image segmentation. We summarise the theoretical basis of Convolutional Neural Networks, Fully Convolutional Neural Networks, U-Net, V-Net, No-New-U-Net (nnU-Net), Transformer Networks, DeepLab, Generative Adversarial Networks, Auto Encoders and Recurrent Neural Networks. In addition, we identify pertinent performance-enhancing measures including adaptive convolutional kernels, atrous convolutions, attention gates, and deep supervision modules.

Top-performing models in ventricular, myocardial, atrial and aortic segmentation are explored, highlighting U-Net and nnU-Net-based model architectures achieving state-of-the art segmentation accuracies. Additionally, key gaps in the current research and technology are identified, and areas of future research are suggested, aiming to guide the innovation and clinical adoption of automated cardiac segmentation methods.

The core objective of this research is to describe the behavior of the distribution using the MLE method to estimate its parameters, as well as to determine the optimal Artificial Neural Network method by comparing it to the maximum likelihood estimation method and applying it to real data for breast cancer patients to determine survival, risk, and other survival study functions of the log-logistic distribution. The parameters were defined in the input layer of the artificial neural network developed for the purpose of survival analysis and reliability function, hazard rate function, probability density function, reserved hazard rate function, Mills ratio, Odd function and CHR values were obtained in the output layer. The findings show that risk function increases with the increase in the time of infection and then decreases for a group of breast cancer patients under study, which corresponds to the theoretical properties of this according to the practical conclusions. The examination of survival analysis reveals that practical conclusions correspond to the theoretical properties of log-logistic distribution. Artificial neural networks have proven to be one of the ideal tools that can be used to predict various vital parameters, especially survival of cancer patients, with their high predictive capabilities.

CRISPR/Cas programmable nuclease systems have become ubiquitous in the field of gene editing. With progressing development, applications in in vivo therapeutic gene editing are increasingly within reach, yet limited by possible adverse side effects from unwanted edits. Recent years have thus seen continuous development of off-target prediction algorithms trained on in vitro cleavage assay data gained from immortalised cell lines. It has been shown that in contrast to experimental epigenetic features, computed physically informed features are so far underutilised despite bearing considerably larger correlation with cleavage activity. Here, we implement state-of-the-art deep learning algorithms and feature encodings for off-target prediction with emphasis on physically informed features that capture the biological environment of the cleavage site, hence terming our approach piCRISPR. Features were gained from the large, diverse crisprSQL off-target cleavage dataset. We find that our best-performing models highlight the importance of sequence context and chromatin accessibility for cleavage prediction and compare favourably with literature standard prediction performance. We further show that our novel, environmentally sensitive features are crucial to accurate prediction on sequence-identical locus pairs, making them highly relevant for clinical guide design. The source code and trained models can be found ready to use at github.com/florianst/picrispr.