Applicability Domain: Towards a More Formal Framework to Express the Applicability of a Model and the Confidence in Individual Predictions

A Graph Neural Network Model with a Transparent Decision-Making Process Defines the Applicability Domain for Environmental Estrogen Screening

The application of deep learning (DL) models for screening environmental estrogens (EEs) for the sound management of chemicals has garnered significant attention. However, the currently available DL model for screening EEs lacks both a transparent decision-making process and effective applicability domain (AD) characterization, making the reliability of its prediction results uncertain and limiting its practical applications. To address this issue, a graph neural network (GNN) model was developed to screen EEs, achieving accuracy rates of 88.9% and 92.5% on the internal and external test sets, respectively. The decision-making process of the GNN model was explored through the network-like similarity graphs (NSGs) based on the model features (FT). We discovered that the accuracy of the predictions is dependent on the feature distribution of compounds in NSGs. An AD characterization method called ADFT was proposed, which excludes predictions falling outside of the model's prediction range, leading to a 15% improvement in the F1 score of the GNN model. The GNN model with the AD method may serve as an efficient tool for screening EEs, identifying 800 potential EEs in the Inventory of Existing Chemical Substances of China. Additionally, this study offers new insights into comprehending the decision-making process of DL models.

Advancing Computational Toxicology by Interpretable Machine Learning

Chemical toxicity evaluations for drugs, consumer products, and environmental chemicals have a critical impact on human health. Traditional animal models to evaluate chemical toxicity are expensive, time-consuming, and often fail to detect toxicants in humans. Computational toxicology is a promising alternative approach that utilizes machine learning (ML) and deep learning (DL) techniques to predict the toxicity potentials of chemicals. Although the applications of ML- and DL-based computational models in chemical toxicity predictions are attractive, many toxicity models are "black boxes" in nature and difficult to interpret by toxicologists, which hampers the chemical risk assessments using these models. The recent progress of interpretable ML (IML) in the computer science field meets this urgent need to unveil the underlying toxicity mechanisms and elucidate the domain knowledge of toxicity models. In this review, we focused on the applications of IML in computational toxicology, including toxicity feature data, model interpretation methods, use of knowledge base frameworks in IML development, and recent applications. The challenges and future directions of IML modeling in toxicology are also discussed. We hope this review can encourage efforts in developing interpretable models with new IML algorithms that can assist new chemical assessments by illustrating toxicity mechanisms in humans.

A benchmark dataset for machine learning in ecotoxicology

The use of machine learning for predicting ecotoxicological outcomes is promising, but underutilized. The curation of data with informative features requires both expertise in machine learning as well as a strong biological and ecotoxicological background, which we consider a barrier of entry for this kind of research. Additionally, model performances can only be compared across studies when the same dataset, cleaning, and splittings were used. Therefore, we provide ADORE, an extensive and well-described dataset on acute aquatic toxicity in three relevant taxonomic groups (fish, crustaceans, and algae). The core dataset describes ecotoxicological experiments and is expanded with phylogenetic and species-specific data on the species as well as chemical properties and molecular representations. Apart from challenging other researchers to try and achieve the best model performances across the whole dataset, we propose specific relevant challenges on subsets of the data and include datasets and splittings corresponding to each of these challenge as well as in-depth characterization and discussion of train-test splitting approaches.

Artificial Intelligence and Machine Learning for Lead-to-Candidate Decision-Making and Beyond

The use of artificial intelligence (AI) and machine learning (ML) in pharmaceutical research and development has to date focused on research: target identification; docking-, fragment-, and motif-based generation of compound libraries; modeling of synthesis feasibility; rank-ordering likely hits according to structural and chemometric similarity to compounds having known activity and affinity to the target(s); optimizing a smaller library for synthesis and high-throughput screening; and combining evidence from screening to support hit-to-lead decisions. Applying AI/ML methods to lead optimization and lead-to-candidate (L2C) decision-making has shown slower progress, especially regarding predicting absorption, distribution, metabolism, excretion, and toxicology properties. The present review surveys reasons why this is so, reports progress that has occurred in recent years, and summarizes some of the issues that remain. Effective AI/ML tools to derisk L2C and later phases of development are important to accelerate the pharmaceutical development process, ameliorate escalating development costs, and achieve greater success rates.

Review of the state of science and evaluation of currently available in silico prediction models for reproductive and developmental toxicity: A case study on pesticides

BACKGROUND

In silico methods for toxicity prediction have increased significantly in recent years due to the 3Rs principle. This also applies to predicting reproductive toxicology, which is one of the most critical factors in pesticide approval. The widely used quantitative structure-activity relationship (QSAR) models use experimental toxicity data to create a model that relates experimentally observed toxicity to molecular structures to predict toxicity. Aim of the study was to evaluate the available prediction models for developmental and reproductive toxicity regarding their strengths and weaknesses in a pesticide database.

METHODS

The reproductive toxicity of 315 pesticides, which have a GHS classification by ECHA, was compared with the prediction of different in silico models: VEGA, OECD (Q)SAR Toolbox, Leadscope Model Applier, and CASE Ultra by MultiCASE.

RESULTS

In all models, a large proportion (up to 77%) of all pesticides were outside the chemical space of the model. Analysis of the prediction of remaining pesticides revealed a balanced accuracy of the models between 0.48 and 0.66.

CONCLUSION

Overall, predictions were only meaningful in rare cases and therefore always require evaluation by an expert. The critical factors were the underlying data and determination of molecular similarity, which offer great potential for improvement.

Studying and mitigating the effects of data drifts on ML model performance at the example of chemical toxicity data

Machine learning models are widely applied to predict molecular properties or the biological activity of small molecules on a specific protein. Models can be integrated in a conformal prediction (CP) framework which adds a calibration step to estimate the confidence of the predictions. CP models present the advantage of ensuring a predefined error rate under the assumption that test and calibration set are exchangeable. In cases where the test data have drifted away from the descriptor space of the training data, or where assay setups have changed, this assumption might not be fulfilled and the models are not guaranteed to be valid. In this study, the performance of internally valid CP models when applied to either newer time-split data or to external data was evaluated. In detail, temporal data drifts were analysed based on twelve datasets from the ChEMBL database. In addition, discrepancies between models trained on publicly-available data and applied to proprietary data for the liver toxicity and MNT in vivo endpoints were investigated. In most cases, a drastic decrease in the validity of the models was observed when applied to the time-split or external (holdout) test sets. To overcome the decrease in model validity, a strategy for updating the calibration set with data more similar to the holdout set was investigated. Updating the calibration set generally improved the validity, restoring it completely to its expected value in many cases. The restored validity is the first requisite for applying the CP models with confidence. However, the increased validity comes at the cost of a decrease in model efficiency, as more predictions are identified as inconclusive. This study presents a strategy to recalibrate CP models to mitigate the effects of data drifts. Updating the calibration sets without having to retrain the model has proven to be a useful approach to restore the validity of most models.

Probabilistic Random Forest improves bioactivity predictions close to the classification threshold by taking into account experimental uncertainty

Measurements of protein–ligand interactions have reproducibility limits due to experimental errors. Any model based on such assays will consequentially have such unavoidable errors influencing their performance which should ideally be factored into modelling and output predictions, such as the actual standard deviation of experimental measurements (σ) or the associated comparability of activity values between the aggregated heterogenous activity units (i.e., K_i versus IC₅₀ values) during dataset assimilation. However, experimental errors are usually a neglected aspect of model generation. In order to improve upon the current state-of-the-art, we herein present a novel approach toward predicting protein–ligand interactions using a Probabilistic Random Forest (PRF) classifier. The PRF algorithm was applied toward in silico protein target prediction across ~ 550 tasks from ChEMBL and PubChem. Predictions were evaluated by taking into account various scenarios of experimental standard deviations in both training and test sets and performance was assessed using fivefold stratified shuffled splits for validation. The largest benefit in incorporating the experimental deviation in PRF was observed for data points close to the binary threshold boundary, when such information was not considered in any way in the original RF algorithm. For example, in cases when σ ranged between 0.4–0.6 log units and when ideal probability estimates between 0.4–0.6, the PRF outperformed RF with a median absolute error margin of ~ 17%. In comparison, the baseline RF outperformed PRF for cases with high confidence to belong to the active class (far from the binary decision threshold), although the RF models gave errors smaller than the experimental uncertainty, which could indicate that they were overtrained and/or over-confident. Finally, the PRF models trained with putative inactives decreased the performance compared to PRF models without putative inactives and this could be because putative inactives were not assigned an experimental pXC₅₀ value, and therefore they were considered inactives with a low uncertainty (which in practice might not be true). In conclusion, PRF can be useful for target prediction models in particular for data where class boundaries overlap with the measurement uncertainty, and where a substantial part of the training data is located close to the classification threshold.

The challenges of generalizability in artificial intelligence for ADME/Tox endpoint and activity prediction

INTRODUCTION

Artificial intelligence (AI) has seen a massive resurgence in recent years with wide successes in computer vision, natural language processing, and games. The similar creation of robust and accurate AI models for ADME/Tox endpoint and activity prediction would be revolutionary to drug discovery pipelines. There have been numerous demonstrations of successful applications, but a key challenge remains: how generalizable are these predictive models?

AREAS COVERED

The authors present a summary of current promising components of AI models in the context of early drug discovery where ADME/Tox endpoint and activity prediction is the main driver of the iterative drug design process. Following that is a review of applicability domains and dataset construction considerations which determine generalizability bottlenecks for AI deployment. Further reviewed is the role of promising learning frameworks - multitask, transfer, and meta learning - which leverage auxiliary data to overcome issues of generalizability.

EXPERT OPINION

The authors conclude that the most promising direction toward integrating reliable and informative AI models into the drug discovery pipeline is a conjunction of learned feature representations, deep learning, and novel learning frameworks. Such a solution would address the sparse and incomplete datasets that are available for key endpoints related to drug discovery.

Artificial intelligence and automation in computer aided synthesis planning

In this perspective we deal with questions pertaining to the development of synthesis planning technologies over the course of recent years.