Artificial intelligence in the life sciences最新文献

PROTACs are a promising therapeutic modality that harnesses the cell’s built-in degradation machinery to degrade specific proteins. Despite their potential, developing new PROTACs is challenging and requires significant domain expertise, time, and cost. Meanwhile, machine learning has transformed drug design and development. In this work, we present a strategy for curating open-source PROTAC data and an open-source deep learning tool for predicting the degradation activity of novel PROTAC molecules. The curated dataset incorporates important information such as $pD{C}_{50}$, $D_{max}$, E3 ligase type, POI amino acid sequence, and experimental cell type. Our model architecture leverages learned embeddings from pretrained machine learning models, in particular for encoding protein sequences and cell type information. We assessed the quality of the curated data and the generalization ability of our model architecture against new PROTACs and targets via three tailored studies, which we recommend other researchers to use in evaluating their degradation activity models. In each study, three models predict protein degradation in a majority vote setting, reaching a top test accuracy of 82.6% and 0.848 ROC AUC, and a test accuracy of 61% and 0.615 ROC AUC when generalizing to novel protein targets. Our results are not only comparable to state-of-the-art models for protein degradation prediction, but also part of an open-source implementation which is easily reproducible and less computationally complex than existing approaches.

Targeted protein degradation (TPD) is a rapidly developing drug discovery technique with unique efficacy and target scope stemming from its degradation-based activity. Molecular glue degraders are a promising arm of TPD, as evidenced by the FDA-approved therapeutics within this class, the increasing number of degraders in clinical development, and their predisposition to drug-likeness. Cereblon (CRBN) glue degraders mediate target degradation by generating a neomorphic interface between CRBN and a protein of interest. While promising, the complicated nature of this CRBN-glue-target ternary complex makes the rational design of molecular glue degraders challenging. For other drug modalities, predictive modeling has been established to leverage existing activity data and generate quantitative structure-activity relationships (QSAR). However, the applicability of QSAR strategies for glues remains under-investigated. Herein, machine learning methodologies were developed to predict glue-mediated recruitment of CRBN to target proteins and achieved promising performance. Generated models leveraged more than a hundred internal screening campaigns across thousands of CRBN glues to predict glue-mediated recruitment of targets to CRBN. Our results show that recruitment activity of CRBN glue degraders can be modeled by machine learning, with 89 % of models producing an area under the receiver operating characteristic curve (ROC AUC) > 0.8 and 70 % of models producing a Matthew's correlation coefficient (MCC) > 0.2 for these primary screening data. Importantly, our findings also indicate that the combination of compound and protein descriptors in the so-called proteochemometric models improves performance, with >80 % of the models exhibiting higher ROC AUC and MCC values than per-target models only based on compound information. Hence, our investigations suggest that proteochemometric modeling is a successful approach for molecular glue degraders. The proposed machine learning strategies can aid compound prioritization based on recruitment efficacy and target selectivity, thus have the potential to facilitate the design and discovery of therapeutic CRBN molecular glues.

High throughput screening (HTS) technologies allow the biological testing of hundreds of thousands of compounds per day. Typically, a substantial proportion of the initial hits obtained by HTS are artifacts caused by assay interference. Therefore, global and technology-specific in silico models for identifying and predicting compounds interfering with biological assays have been developed. The global models benefit from training on large screening data sets, while the specialized models benefit from training on assay technology-specific experimental data. In this work, we develop and explore strategies for generating better predictors of technology-specific assay interference by utilizing the large bioactivity data matrices global models are trained on and employing partially new compound labeling approaches to maintain the assay technology awareness of specialized models. We demonstrate the utility of the statistically derived interference labels in machine learning using fluorescence-based assay interference as a representative example. Our random forest and multi-layer perceptron classifiers showed improved performance compared to existing models, achieving Matthews correlation coefficients (MCCs) of up to 0.47 on holdout data and up to 0.45 on an external test set. These results demonstrate that accurate assay-specific interference labels can be derived from large bioactivity data matrices, enabling the development of new machine-learning models without the need for further experimental data.

Having access to sufficient data is essential in order to train accurate machine learning models, but much data is not publicly available. In drug discovery this is particularly evident, as much data is withheld at pharmaceutical companies for various reasons. Federated Learning (FL) aims at training a joint model between multiple parties but without disclosing data between the parties. In this work, we leverage Federated Learning to predict compound Mechanism of Action (MoA) using fluorescence image data from cell painting. Our study evaluates the effectiveness and efficiency of FL, comparing to non-collaborative and data-sharing collaborative learning in diverse scenarios. Specifically, we investigate the impact of data heterogeneity across participants on MoA prediction, an essential concern in real-life applications of FL, and demonstrate the benefits for all involved parties. This work highlights the potential of federated learning in multi-institutional collaborative machine learning for drug discovery and assessment of chemicals, offering a promising avenue to overcome data-sharing constraints.

A variety of environmental and physiological conditions can cause oxidative stress that damage cellular components such as DNA, proteins and lipids. Oxidative stress is implicated in many human diseases including cancer, cardiovascular diseases, neurological diseases, inflammatory diseases, and aging. The nuclear factor erythroid 2–related factor 2 (NRF2) is a transcriptional factor that plays a key role in the cellular antioxidant defense system as it regulates transcription of antioxidant proteins and detoxifying enzymes. There is an urgent need to identify novel compounds that activate NRF2 and enhance antioxidant defense. We collected data from the high-throughput screening of NRF2 activators and identified molecular fragments (structural alerts) associated with the activation of NRF2. We also developed ten classification models using different types of molecular descriptors and machine learning techniques. Two approaches were used to establish the applicability domain of developed models: the structure-based approach and the distance to model approach. The best performing model that used message passing neural network (MPNN) technique showed accuracy of 87 % for the test set of chemicals within the distance to model of 0.3. The integrative approach using a combination of generated structural alerts and MPNN model was used to screen approved drugs collected in the DrugBank to identify potential NRF2 activators. Out of 2393 screened chemicals 138 compounds were predicted as NRF2 activators by both approaches. Analysis of these compounds showed that some drugs were already known activators of NRF2 while others are potentially novel activators.

In chemoinformatics, artificial intelligence (AI) continues to grow a symbiosis with open science (OS). Such a close AI-OS interaction brings substantial practical benefits in research, scientific dissemination, and education, to name a few areas. The AI-OS symbiosis can be further enhanced by combining sufficient substantive expertise, mathematical and statistical knowledge, and coding skills. This Viewpoint discusses the benefits of the smooth and productive interaction between AI, OS, and open data. We also present a short list of misconceptions and pitfalls surrounding AI-OS and propose correct responses and behaviors agreed upon by field experts. In addition, we provide suggestions to continue enhancing the positive contributions of the AI-OS symbiosis towards chemoinformatics.

Biomedical knowledge graphs (KGs) hold valuable information regarding biomedical entities such as genes, diseases, biological processes, and drugs. KGs have been successfully employed in challenging biomedical areas such as the identification of pathophysiology mechanisms or drug repurposing. The creation of high-quality KGs typically requires labor-intensive multi-database integration or substantial human expert curation, both of which take time and contribute to the workload of data processing and annotation. Therefore, the use of automatic systems for KG building and maintenance is a prerequisite for the wide uptake and utilization of KGs. Technologies supporting the automated generation and updating of KGs typically make use of Natural Language Processing (NLP), which is optimized for extracting implicit triples described in relevant biomedical text sources. At the core of this challenge is how to improve the accuracy and coverage of the information extraction module by utilizing different models and tools. The emergence of pre-trained large language models (LLMs), such as ChatGPT which has grown in popularity dramatically, has revolutionized the field of NLP, making them a potential candidate to be used in text-based graph creation as well. So far, no previous work has investigated the power of LLMs on the generation of cause-and-effect networks and KGs encoded in Biological Expression Language (BEL). In this paper, we present initial studies towards one-shot BEL relation extraction using two different versions of the Generative Pre-trained Transformer (GPT) models and evaluate its performance by comparing the extracted results to a highly accurate, manually curated BEL KG curated by domain experts.

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.