QSAR-derived affinity fingerprints (part 1): fingerprint construction and modeling performance for similarity searching, bioactivity classification and scaffold hopping

The Goldilocks paradigm: comparing classical machine learning, large language models, and few-shot learning for drug discovery applications

Recent advances in machine learning (ML) have led to newer model architectures including transformers (large language models, LLMs) showing state of the art results in text generation and image analysis as well as few-shot learning (FSLC) models which offer predictive power with extremely small datasets. These new architectures may offer promise, yet the 'no-free lunch' theorem suggests that no single model algorithm can outperform at all possible tasks. Here, we explore the capabilities of classical (SVR), FSLC, and transformer models (MolBART) over a range of dataset tasks and show a 'goldilocks zone' for each model type, in which dataset size and feature distribution (i.e. dataset "diversity") determines the optimal algorithm strategy. When datasets are small ( < 50 molecules), FSLC tend to outperform both classical ML and transformers. When datasets are small-to-medium sized (50-240 molecules) and diverse, transformers outperform both classical models and few-shot learning. Finally, when datasets are of larger and of sufficient size, classical models then perform the best, suggesting that the optimal model to choose likely depends on the dataset available, its size and diversity. These findings may help to answer the perennial question of which ML algorithm is to be used when faced with a new dataset.

Chemoinformatic approaches for navigating large chemical spaces

INTRODUCTION

Large chemical spaces (CSs) include traditional large compound collections, combinatorial libraries covering billions to trillions of molecules, DNA-encoded chemical libraries comprising complete combinatorial CSs in a single mixture, and virtual CSs explored by generative models. The diverse nature of these types of CSs require different chemoinformatic approaches for navigation.

AREAS COVERED

An overview of different types of large CSs is provided. Molecular representations and similarity metrics suitable for large CS exploration are discussed. A summary of navigation of CSs in generative models is provided. Methods for characterizing and comparing CSs are discussed.

EXPERT OPINION

The size of large CSs might restrict navigation to specialized algorithms and limit it to considering neighborhoods of structurally similar molecules. Efficient navigation of large CSs not only requires methods that scale with size but also requires smart approaches that focus on better but not necessarily larger molecule selections. Deep generative models aim to provide such approaches by implicitly learning features relevant for targeted biological properties. It is unclear whether these models can fulfill this ideal as validation is difficult as long as the covered CSs remain mainly virtual without experimental verification.

MATEO: intermolecular α-amidoalkylation theoretical enantioselectivity optimization. Online tool for selection and design of chiral catalysts and products

The enantioselective Brønsted acid-catalyzed α-amidoalkylation reaction is a useful procedure is for the production of new drugs and natural products. In this context, Chiral Phosphoric Acid (CPA) catalysts are versatile catalysts for this type of reactions. The selection and design of new CPA catalysts for different enantioselective reactions has a dual interest because new CPA catalysts (tools) and chiral drugs or materials (products) can be obtained. However, this process is difficult and time consuming if approached from an experimental trial and error perspective. In this work, an Heuristic Perturbation-Theory and Machine Learning (HPTML) algorithm was used to seek a predictive model for CPA catalysts performance in terms of enantioselectivity in α-amidoalkylation reactions with R2 = 0.96 overall for training and validation series. It involved a Monte Carlo sampling of > 100,000 pairs of query and reference reactions. In addition, the computational and experimental investigation of a new set of intermolecular α-amidoalkylation reactions using BINOL-derived N-triflylphosphoramides as CPA catalysts is reported as a case of study. The model was implemented in a web server called MATEO: InterMolecular Amidoalkylation Theoretical Enantioselectivity Optimization, available online at: https://cptmltool.rnasa-imedir.com/CPTMLTools-Web/mateo . This new user-friendly online computational tool would enable sustainable optimization of reaction conditions that could lead to the design of new CPA catalysts along with new organic synthesis products.

Applications and Potential of In Silico Approaches for Psychedelic Chemistry

Molecular-level investigations of the Central Nervous System have been revolutionized by the development of computational methods, computing power, and capacity advances. These techniques have enabled researchers to analyze large amounts of data from various sources, including genomics, in vivo, and in vitro drug tests. In this review, we explore how computational methods and informatics have contributed to our understanding of mental health disorders and the development of novel drugs for neurological diseases, with a special focus on the emerging field of psychedelics. In addition, the use of state-of-the-art computational methods to predict the potential of drug compounds and bioinformatic tools to integrate disparate data sources to create predictive models is also discussed. Furthermore, the challenges associated with these methods, such as the need for large datasets and the diversity of in vitro data, are explored. Overall, this review highlights the immense potential of computational methods and informatics in Central Nervous System research and underscores the need for continued development and refinement of these techniques and more inclusion of Quantitative Structure-Activity Relationships (QSARs).

Prediction of KRASG12C inhibitors using conjoint fingerprint and machine learning-based QSAR models

Kirsten rat sarcoma virus G12C (KRAS^G12C) is the major protein mutation associated with non-small cell lung cancer (NSCLC) severity. Inhibiting KRAS^G12C is therefore one of the key therapeutic strategies for NSCLC patients. In this paper, a cost-effective data driven drug design employing machine learning-based quantitative structure-activity relationship (QSAR) analysis was built for predicting ligand affinities against KRAS^G12C protein. A curated and non-redundant dataset of 1033 compounds with KRAS^G12C inhibitory activity (pIC₅₀) was used to build and test the models. The PubChem fingerprint, Substructure fingerprint, Substructure fingerprint count, and the conjoint fingerprint-a combination of PubChem fingerprint and Substructure fingerprint count-were used to train the models. Using comprehensive validation methods and various machine learning algorithms, the results clearly showed that the XGBoost regression (XGBoost) achieved the highest performance in term of goodness of fit, predictivity, generalizability and model robustness (R² = 0.81, Q²_CV = 0.60, Q²_Ext = 0.62, R² - Q²_Ext = 0.19, R²_Y-Random = 0.31 ± 0.03, Q²_Y-Random = -0.09 ± 0.04). The top 13 molecular fingerprints that correlated with the predicted pIC₅₀ values were SubFPC274 (aromatic atoms), SubFPC307 (number of chiral-centers), PubChemFP37 (≥1 Chlorine), SubFPC18 (Number of alkylarylethers), SubFPC1 (number of primary carbons), SubFPC300 (number of 1,3-tautomerizables), PubChemFP621 (N-C:C:C:N structure), PubChemFP23 (≥1 Fluorine), SubFPC2 (number of secondary carbons), SubFPC295 (number of C-ONS bonds), PubChemFP199 (≥4 6-membered rings), PubChemFP180 (≥1 nitrogen-containing 6-membered ring), and SubFPC180 (number of tertiary amine). These molecular fingerprints were virtualized and validated using molecular docking experiments. In conclusion, this conjoint fingerprint and XGBoost-QSAR model demonstrated to be useful as a high-throughput screening tool for KRAS^G12C inhibitor identification and drug design.

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.

Concepts and applications of chemical fingerprint for hit and lead screening

Molecular fingerprints are used to represent chemical (structural, physicochemical, etc.) properties of large-scale chemical sets in a low computational cost way. They have a prominent role in transforming chemical data sets into consistent input formats (bit strings or numeric values) suitable for in silico approaches. In this review, we summarize and classify common and state-of-the-art fingerprints into eight different types (dictionary based, circular, topological, pharmacophore, protein-ligand interaction, shape based, reinforced, and multi). We also highlight applications of fingerprints in early drug research and development (R&D). Thus, this review provides a guide for the selection of appropriate fingerprints of compounds (or ligand-protein complexes) for use in drug R&D.

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.

Prediction of the Neurotoxic Potential of Chemicals Based on Modelling of Molecular Initiating Events Upstream of the Adverse Outcome Pathways of (Developmental) Neurotoxicity

Developmental and adult/ageing neurotoxicity is an area needing alternative methods for chemical risk assessment. The formulation of a strategy to screen large numbers of chemicals is highly relevant due to potential exposure to compounds that may have long-term adverse health consequences on the nervous system, leading to neurodegeneration. Adverse Outcome Pathways (AOPs) provide information on relevant molecular initiating events (MIEs) and key events (KEs) that could inform the development of computational alternatives for these complex effects. We propose a screening method integrating multiple Quantitative Structure–Activity Relationship (QSAR) models. The MIEs of existing AOP networks of developmental and adult/ageing neurotoxicity were modelled to predict neurotoxicity. Random Forests were used to model each MIE. Predictions returned by single models were integrated and evaluated for their capability to predict neurotoxicity. Specifically, MIE predictions were used within various types of classifiers and compared with other reference standards (chemical descriptors and structural fingerprints) to benchmark their predictive capability. Overall, classifiers based on MIE predictions returned predictive performances comparable to those based on chemical descriptors and structural fingerprints. The integrated computational approach described here will be beneficial for large-scale screening and prioritisation of chemicals as a function of their potential to cause long-term neurotoxic effects.

Revealing Unknown Controlled Substances and New Psychoactive Substances Using High-Resolution LC-MS/MS Machine Learning Models and the Hybrid Similarity Search Algorithm

High-resolution LC-MS/MS tandem mass spectra-based machine learning models are constructed to address the analytical challenge of identifying unknown controlled substances and new psychoactive substances (NPS's). Using a training set comprised of 770 LC-MS/MS barcode spectra (with binary entries 0 or 1) obtained generally by high-resolution mass spectrometers, three classification machine learning models were generated and evaluated. The three models are artificial neural network (ANN), support vector machine (SVM), and k-nearest neighbor (k-NN) models. In these models, controlled substances and NPS's were classified into 13 subgroups (benzylpiperazine, opiate, benzodiazepine, amphetamine, cocaine, methcathinone, classical cannabinoid, fentanyl, 2C series, indazole carbonyl compound, indole carbonyl compound, phencyclidine, and others). Using 193 LC-MS/MS barcode spectra as an external test set, accuracy of the ANN, SVM, and k-NN models were evaluated as 72.5%, 90.0%, and 94.3%, respectively. Also, the hybrid similarity search (HSS) algorithm was evaluated to examine whether this algorithm can successfully identify unknown controlled substances and NPS's whose data are unavailable in the database. When only 24 representative LC-MS/MS spectra of controlled substances and NPS's were selectively included in the database, it was found that HSS can successfully identify compounds with high reliability. The machine learning models and HSS algorithms are incorporated into our home-coded AI-SNPS (artificial intelligence screener for narcotic drugs and psychotropic substances) standalone software that is equipped with a graphic user interface. The use of this software allows unknown controlled substances and NPS's to be identified in a convenient manner.

Effect of Dataset Size and Train/Test Split Ratios in QSAR/QSPR Multiclass Classification

Applied datasets can vary from a few hundred to thousands of samples in typical quantitative structure-activity/property (QSAR/QSPR) relationships and classification. However, the size of the datasets and the train/test split ratios can greatly affect the outcome of the models, and thus the classification performance itself. We compared several combinations of dataset sizes and split ratios with five different machine learning algorithms to find the differences or similarities and to select the best parameter settings in nonbinary (multiclass) classification. It is also known that the models are ranked differently according to the performance merit(s) used. Here, 25 performance parameters were calculated for each model, then factorial ANOVA was applied to compare the results. The results clearly show the differences not just between the applied machine learning algorithms but also between the dataset sizes and to a lesser extent the train/test split ratios. The XGBoost algorithm could outperform the others, even in multiclass modeling. The performance parameters reacted differently to the change of the sample set size; some of them were much more sensitive to this factor than the others. Moreover, significant differences could be detected between train/test split ratios as well, exerting a great effect on the test validation of our models.

Profiling and analysis of chemical compounds using pointwise mutual information

Pointwise mutual information (PMI) is a measure of association used in information theory. In this paper, PMI is used to characterize several publicly available databases (DrugBank, ChEMBL, PubChem and ZINC) in terms of association strength between compound structural features resulting in database PMI interrelation profiles. As structural features, substructure fragments obtained by coding individual compounds as MACCS, PubChemKey and ECFP fingerprints are used. The analysis of publicly available databases reveals, in accord with other studies, unusual properties of DrugBank compounds which further confirms the validity of PMI profiling approach. Z-standardized relative feature tightness (ZRFT), a PMI-derived measure that quantifies how well the given compound's feature combinations fit these in a particular compound set, is applied for the analysis of compound synthetic accessibility (SA), as well as for the classification of compounds as easy (ES) and hard (HS) to synthesize. ZRFT value distributions are compared with these of SYBA and SAScore. The analysis of ZRFT values of structurally complex compounds in the SAVI database reveals oligopeptide structures that are mispredicted by SAScore as HS, while correctly predicted by ZRFT and SYBA as ES. Compared to SAScore, SYBA and random forest, ZRFT predictions are less accurate, though by a narrow margin (AccZRFT = 94.5%, AccSYBA = 98.8%, AccSAScore = 99.0%, AccRF = 97.3%). However, ZRFT ability to distinguish between ES and HS compounds is surprisingly high considering that while SYBA, SAScore and random forest are dedicated SA models, ZRFT is a generic measurement that merely quantifies the strength of interrelations between structural feature pairs. The results presented in the current work indicate that structural feature co-occurrence, quantified by PMI or ZRFT, contains a significant amount of information relevant to physico-chemical properties of organic compounds.

Data-driven machine learning model for the prediction of oxygen vacancy formation energy of metal oxide materials

Metal oxides are widely used in the fields of chemistry, physics and materials science. Oxygen vacancy formation energy is a key parameter to describe the chemical, mechanical, and thermodynamic properties of metal oxides. How to acquire quickly and accurately oxygen vacancy formation energy remains a challenge for both experimental and theoretical researchers. Herein, we propose a machine learning model for the prediction of oxygen vacancy formation energy via data-driven analysis and the definition of simple descriptors. Starting with the database containing oxygen vacancy formation energies for 1750 metal oxides with enough structural diversity, new descriptors that effectively avoid the defects of molecular fingerprints, molecular graphic descriptors and site descriptors are defined. The descriptors have obvious physical meanings and wide practicability. Multiple linear regression analysis is then used to screen important features for machine learning model development, and two strongly associated features are obtained. The selected descriptors are used as input for the training of 21 machine learning models to select and develop the most accurate machine learning model. Finally, it is shown that the least squares support vector regression method exhibits the best performance for accurate prediction of the targeted oxygen vacancy formation energy through systematic error analysis, and the prediction accuracy is also verified by the external dataset. Our work establishes a novel and simple computational approach for accurate prediction of the oxygen vacancy formation energy of metal oxides and highlights the availability of data-driven analysis for metal oxide material research.

From Big Data to Artificial Intelligence: chemoinformatics meets new challenges

The increasing volume of biomedical data in chemistry and life sciences requires development of new methods and approaches for their analysis. Artificial Intelligence and machine learning, especially neural networks, are increasingly used in the chemical industry, in particular with respect to Big Data. This editorial highlights the main results presented during the special session of the International Conference on Neural Networks organized by "Big Data in Chemistry" project and draws perspectives on the future progress of the field.

Bioactivity Comparison across Multiple Machine Learning Algorithms Using over 5000 Datasets for Drug Discovery

Machine learning methods are attracting considerable attention from the pharmaceutical industry for use in drug discovery and applications beyond. In recent studies, we and others have applied multiple machine learning algorithms and modeling metrics and, in some cases, compared molecular descriptors to build models for individual targets or properties on a relatively small scale. Several research groups have used large numbers of datasets from public databases such as ChEMBL in order to evaluate machine learning methods of interest to them. The largest of these types of studies used on the order of 1400 datasets. We have now extracted well over 5000 datasets from CHEMBL for use with the ECFP6 fingerprint and in comparison of our proprietary software Assay Central with random forest, k-nearest neighbors, support vector classification, naïve Bayesian, AdaBoosted decision trees, and deep neural networks (three layers). Model performance was assessed using an array of fivefold cross-validation metrics including area-under-the-curve, F1 score, Cohen's kappa, and Matthews correlation coefficient. Based on ranked normalized scores for the metrics or datasets, all methods appeared comparable, while the distance from the top indicated that Assay Central and support vector classification were comparable. Unlike prior studies which have placed considerable emphasis on deep neural networks (deep learning), no advantage was seen in this case. If anything, Assay Central may have been at a slight advantage as the activity cutoff for each of the over 5000 datasets representing over 570,000 unique compounds was based on Assay Central performance, although support vector classification seems to be a strong competitor. We also applied Assay Central to perform prospective predictions for the toxicity targets PXR and hERG to further validate these models. This work appears to be the largest scale comparison of these machine learning algorithms to date. Future studies will likely evaluate additional databases, descriptors, and machine learning algorithms and further refine the methods for evaluating and comparing such models.

Harnessing In Silico, In Vitro, and In Vivo Data to Understand the Toxicity Landscape of Polycyclic Aromatic Compounds (PACs)

Polycyclic aromatic compounds (PACs) are compounds with a minimum of two six-atom aromatic fused rings. PACs arise from incomplete combustion or thermal decomposition of organic matter and are ubiquitous in the environment. Within PACs, carcinogenicity is generally regarded to be the most important public health concern. However, toxicity in other systems (reproductive and developmental toxicity, immunotoxicity) has also been reported. Despite the large number of PACs identified in the environment, research attention to understand exposure and health effects of PACs has focused on a relatively limited subset, namely polycyclic aromatic hydrocarbons (PAHs), the PACs with only carbon and hydrogen atoms. To triage the rest of the vast number of PACs for more resource-intensive testing, we developed a data-driven approach to contextualize hazard characterization of PACs, by leveraging the available data from various data streams (in silico toxicity, in vitro activity, structural fingerprints, and in vivo data availability). The PACs were clustered on the basis of their in silico toxicity profiles containing predictions from 8 different categories (carcinogenicity, cardiotoxicity, developmental toxicity, genotoxicity, hepatotoxicity, neurotoxicity, reproductive toxicity, and urinary toxicity). We found that PACs with the same parent structure (e.g., fluorene) could have diverse in silico toxicity profiles. In contrast, PACs with similar substituted groups (e.g., alkylated-PAHs) or heterocyclics (e.g., N-PACs) with varying ring sizes could have similar in silico toxicity profiles, suggesting that these groups are better candidates for toxicity read-across analysis. The clusters/regions associated with certain in silico toxicity, in vitro activity, and structural fingerprints were identified. We found that genotoxicity/carcinogenicity (in silico toxicity) and xenobiotic homeostasis and stress response (in vitro activity), respectively, dominate the toxicity/activity variation seen in the PACs. The "hot spots" with enriched toxicity/activity in conjunction with availability of in vivo carcinogenicity data revealed regions of either data-poor (hydroxylated-PAHs) or data-rich (unsubstituted, parent PAHs) PACs. These regions offer potential targets for prioritization of further in vivo assessment and for chemical read-across efforts. The analysis results are searchable through an interactive web application (https://ntp.niehs.nih.gov/go/pacs_tableau), allowing for alternative hypothesis generation.

QSAR-derived affinity fingerprints (part 2): modeling performance for potency prediction

Affinity fingerprints report the activity of small molecules across a set of assays, and thus permit to gather information about the bioactivities of structurally dissimilar compounds, where models based on chemical structure alone are often limited, and model complex biological endpoints, such as human toxicity and in vitro cancer cell line sensitivity. Here, we propose to model in vitro compound activity using computationally predicted bioactivity profiles as compound descriptors. To this aim, we apply and validate a framework for the calculation of QSAR-derived affinity fingerprints (QAFFP) using a set of 1360 QSAR models generated using K_i, K_d, IC₅₀ and EC₅₀ data from ChEMBL database. QAFFP thus represent a method to encode and relate compounds on the basis of their similarity in bioactivity space. To benchmark the predictive power of QAFFP we assembled IC₅₀ data from ChEMBL database for 18 diverse cancer cell lines widely used in preclinical drug discovery, and 25 diverse protein target data sets. This study complements part 1 where the performance of QAFFP in similarity searching, scaffold hopping, and bioactivity classification is evaluated. Despite being inherently noisy, we show that using QAFFP as descriptors leads to errors in prediction on the test set in the ~ 0.65-0.95 pIC₅₀ units range, which are comparable to the estimated uncertainty of bioactivity data in ChEMBL (0.76-1.00 pIC₅₀ units). We find that the predictive power of QAFFP is slightly worse than that of Morgan2 fingerprints and 1D and 2D physicochemical descriptors, with an effect size in the 0.02-0.08 pIC₅₀ units range. Including QSAR models with low predictive power in the generation of QAFFP does not lead to improved predictive power. Given that the QSAR models we used to compute the QAFFP were selected on the basis of data availability alone, we anticipate better modeling results for QAFFP generated using more diverse and biologically meaningful targets. Data sets and Python code are publicly available at https://github.com/isidroc/QAFFP_regression .