Artificial intelligence in the life sciences最新文献

A rising amount of research demonstrates that artificial intelligence and machine learning approaches can provide an essential basis for the drug design and discovery process. Deep learning algorithms are being developed in response to recent advances in computer technology as part of the creation of therapeutically relevant medications for the treatment of a variety of ailments. In this review, we focus on the most recent advances in the areas of drug design and discovery research employing generative deep learning methodologies such as generative adversarial network (GAN) frameworks. To begin, we examine drug design and discovery studies that use several GAN methodologies to evaluate one key application, such as molecular de novo design in drug design and discovery. Furthermore, we discuss many GAN models for dimension reduction of single-cell data at the preclinical stage of the drug development pipeline. We also show various experiments in de novo peptide and protein creation utilizing GAN frameworks. Furthermore, we discuss the limits of past drug design and discovery research employing GAN models. Finally, we give a discussion on future research prospects and obstacles.

The study of protein-peptide interactions is an active research field from an experimental and computational perspective, with the latest presenting challenges to model and simulate the peptides' intrinsic flexibility. Predicting affinities towards protein systems of interest, such as proteases, is crucial to understand the specificity of the interactions and support the discovery of novel substrates. Here we provide a set of computational protocols to run structural and dynamical analysis of protein-peptide complexes from a binding perspective. The protocols are based on state-of-the-art methods, but the code is open and can be customized depending on the user needs. These include a fragment-growing peptide docking protocol to predict bound conformations of flexible peptides, a protocol to extract descriptors from protein-peptide molecular dynamics trajectories, and a workflow to build and test machine learning regression models. As a toy example, we applied the protocols to a serine protease structure with a set of known peptide substrates and random sequences to illustrate the use of the code, which is publicly available at: https://github.com/rochoa85/Protocols-Peptide-Binding

Anyone involved in designing or finding molecules in the life sciences over the past few years has witnessed a dramatic change in how we now work due to the COVID-19 pandemic. Computational technologies like artificial intelligence (AI) seemed to become ubiquitous in 2020 and have been increasingly applied as scientists worked from home and were separated from the laboratory and their colleagues. This shift may be more permanent as the future of molecule design across different industries will increasingly require machine learning models for design and optimization of molecules as they become “designed by AI”. AI and machine learning has essentially become a commodity within the pharmaceutical industry. This perspective will briefly describe our personal opinions of how machine learning has evolved and is being applied to model different molecule properties that crosses industries in their utility and ultimately suggests the potential for tight integration of AI into equipment and automated experimental pipelines. It will also describe how many groups have implemented generative models covering different architectures, for de novo design of molecules. We also highlight some of the companies at the forefront of using AI to demonstrate how machine learning has impacted and influenced our work. Finally, we will peer into the future and suggest some of the areas that represent the most interesting technologies that may shape the future of molecule design, highlighting how we can help increase the efficiency of the design-make-test cycle which is currently a major focus across industries.

While Relative Binding Free Energy (RBFE) calculations have become a mainstay in lead optimization programs, the computational expense of performing these calculations has limited their broader application. Active learning (AL), a machine learning method used to direct a search iteratively, has explored larger chemical libraries using RBFE calculations. While AL has been successfully applied, there has not been a systematic study of the impact of parameter settings on the performance of AL. To address this gap, we have generated an exhaustive dataset of RBFE calculations on 10,000 congeneric molecules. We used this dataset to explore the impact of several AL design choices, including the number of molecules sampled at each iteration, the method used to select an initial sample, the method used to build a machine learning model, and the acquisition function that defines the balance between exploration and exploitation in the search. Our studies demonstrated that the performance of AL is largely insensitive to the specific machine learning method and acquisition functions used. In our studies, the most significant factor impacting performance was the number of molecules sampled at each iteration where selecting too few molecules hurts performance. Under the best conditions, we were able to identify 75% of the 100 top scoring molecules by sampling only 6% of the dataset. We hope that the dataset of 10K molecules will provide the basis for future studies exploring additional AL strategies. The source code and supporting data for the work are available at https://github.com/google-research/google-research/tree/master/al_for_fep.

The Bioconcentration Factor (BCF) is an important parameter in the environmental risk assessment of chemicals, relevant for industrial and academic research as well as required in many regulatory contexts. It represents the potential of a substance to accumulate in organic tissues or whole animals and is most frequently measured in fish. However, animal welfare reasons, throughput limitations, and costs push the need for alternative methods that allow accurate and reliable estimations of BCF in silico. We present a new deep learning model to predict BCF values from chemical structures, that outperforms currently available models ($R^2$ of 0.68 and RMSE of 0.59 log units on an external test set; $R^2$ of 0.70 and RMSE of 0.74 log units in a demanding cluster split validation). The model is based on molecular representations encoded as CDDD descriptors and exploits a large in-house dataset with measured logD values as an auxiliary task.

Additionally, we developed a post-hoc explainability method based on SMILES character substitutions to accompany our predictions with atom-level interpretations. These sensitivity scores highlight the most influential moieties in the molecule and can help to understand the predictions better and design new molecules.

The interpretation of quantitative structure–activity or structure–property relationships is important in the field of chemoinformatics. Although multivariate linear regression models are typically interpretable, they do not generally have high predictive abilities. Symbolic regression (SR) combined with genetic programming (GP) is a well-established technique for generating the mathematical expressions that describe the relationships within a dataset. However, SR sometimes produces complicated expressions that are hard for humans to interpret. This paper proposes a method for generating simpler expressions by incorporating three filters into GP-based SR. The filters are further combined with nonlinear least-squares optimization to give filter-introduced GP (FIGP), which improves the predictive ability of SR models while retaining simple expressions. As a proof-of-concept, the quantitative estimate of drug-likeness and the synthetic accessibility score are predicted based on the chemical structures of compounds. Overall, FIGP generates less-complicated expressions than previous SR methods. In terms of predictive ability, FIGP is better than GP, but is outperformed by a support vector machine with a radial basis function kernel. Furthermore, quantitative structure–activity relationship models are constructed for three matching molecular series with biological targets. In the case of one target, the activity prediction models given by FIGP exhibit better predictive ability than multivariate linear regression and support vector regression with the radial basis function kernel, whereas for the remaining cases, FIGP is slightly less accurate than multivariate linear regression.

Recent methodological advances in deep learning (DL) architectures have not only improved the performance of predictive models but also enhanced their interpretability potential, thus considerably increasing their transparency. In the context of medicinal chemistry, the potential to not only accurately predict molecular properties, but also chemically interpret them, would be strongly preferred. Previously, we developed accurate multi-task convolutional neural network (CNN) and graph convolutional neural network (GCNN) models to predict a set of diverse intrinsic metabolic clearance parameters from image- and graph-based molecular representations, respectively. Herein, we introduce several model interpretability frameworks to answer whether the model explanations obtained from CNN and GCNN multi-task clearance models could be applied to predict chemical transformations associated with experimentally confirmed metabolic products. We show a strong correlation between the CNN pixel intensities and corresponding clearance predictions, as well as their robustness to different molecular orientations. Using actual case examples, we demonstrate that both CNN and GCNN interpretations frequently complement each other, suggesting their high potential for combined use in guiding medicinal chemistry design.

Computational linear T-cell epitope prediction tools allow cost and labor reduction in downstream in vitro testing, but the quality of currently available methods is compromised by the scarcity of experimental data and extensive HLA polymorphism. However, it is possible to improve prediction quality by forgoing HLA-dependency that allows treating all immunogenic sequences as a single group. This reduces the problem to a much simpler two-classes classification of determining whether a peptide is immunogenic or not. Here, we use a deep convolutional neural network capable of predicting linear T-cell epitope regions in primary structures trained using all peptides deposited in the IEDB website. We also investigate the possibility of using peptides derived from known human proteins as non-immunogenic counterexamples. We compared our model with a state-of-the-art tool and analyze the benefits of using larger databases. Our results corroborate the usefulness of HLA-free methods for practical applications that require the identification of immunogenic sequences. Deepitope is an open source project that can be found at https://github.com/raphaeltrevizani/deepitope.