Chemoinformatics and artificial intelligence colloquium: progress and challenges in developing bioactive compounds | Journal of Cheminformatics | Full Text

One of the most important issues in chemoinformatics is how to compare molecules. There are two equally important aspects to this issue: (1) how to represent the information in a molecular structure in a computationally appropriate form and (2) how to determine the structural relationship of one molecule to another using this information. In the first instance, a common approach in widespread use today is the development of ‘vectorized’ representations of molecular structure such as that exemplified by Extended Connectivity Fingerprints (ECFP) [37] or MACCS key fingerprints [38], that represent the structural features of molecules as binary vectors whose components are based on the presence or absence of specific substructural features. In addition, SMILES sequences and molecular graphs are being used as features for the most recent neural networks architectures. Many of these and closely related methods provide a basis for developing all manner of AI models. An important caveat regarding these approaches is that they deal almost exclusively with 2D molecular structures. Three-dimensional structural features, such as multiple conformations, are rarely treated for a variety of reasons.

Ligand-based drug-design opportunities

In addition to in vitro and in vivo methods, in silico methods can enhance serendipity and help to rationalize phenomena that experimental methods alone cannot explain. For example, serendipity in drug design can lead to unexpected but potentially positive results, as exemplified by the discovery of Lyrica (pregabalin) [52]. An excellent opportunity for ligand-based methods to enhance compound comparisons is through the addition or augmentation [15] of chemical and physicochemical property data, of in vitro, in vivo, and ‘omics’ biological data, and of preclinical, clinical, and post-marketing pharmacovigilance data. The added information would support the development of a comprehensive similarity searching capability that would likely, in specific instances, be able to identify chemical mimetics capable of reverting disease signatures. For example, drug-design procedures might be developed for reversing (or preventing) molecular pathway alterations or for predicting toxicity or safety issues for marketed drugs [53].

Two new applications, Extended Similarity Indices [23, 24] and the structure–activity relationships Matrix (SARM) approach and its deep learning extension (DeepSARM) [25], were presented at the Colloquium by Quintana (Talk 12) and Bajorath (Talk 13), respectively. These applications support multiple procedures such as analog series identification (fragmentation?), analysis of de novo drug-design signatures, similarity searching, and visualization of SAR and chemical spaces.

Structure-based drug-design opportunities

Over the past few decades, SBDD has attained a significant degree of maturity. This is especially true with regard to structure-based virtual screening, which has made remarkable progress despite its intrinsic limitations [54, 55]. In recent years, DL has been used in attempts to further improve the performance of SBDD methods. Perhaps the most well-known example of this is the usage of DL for protein structure prediction. De novo structure prediction with Alphafold [10] RoseTTAfold [50], or other programs [51, 56] has yielded many protein models of near-experimental accuracy which has further expanded the opportunities and the applicability domain of homology modeling. Protein models are now increasingly used for prediction of many biophysical properties [57].

Other uses of AI in SBDD include, but are not limited to, potential energy functions that are similar to quantum-chemical descriptions (ANAKIN-ME) [9]. For example, DFT-like interaction potentials at the computational cost of a geometrical optimization with molecular mechanics; force field development [58]; enhanced sampling by means of collective variables [59]; Boltzmann generators trained to identify transition states [60]; protein-ligand interaction fingerprints [61] such as SPLIF [62] or ECIF [63], and scoring functions like GNINA [64]. Recently, the geometric DL approach was used to learn distance distributions and ligand-target interactions and to predict the binding conformation of bioactive compounds. This potential performs as well as or better than well-established scoring functions [27]. Geometry DL uses a mesh on the protein surface [65] as a molecular representation.

New approaches to CADD based on AI methodologies

Chemoinformatics helps transform data into information and subsequently into knowledge in support of decision making. New techniques and methodologies have contributed significantly to encoding and analyzing chemical, biological, and clinical data patterns. For example, different types of neural networks (e.g., neural, deep neural, Kohonen-Self Organizing Maps (SOM), and graph-based) [7] support multitask learning, which facilitates the exploration and exploitation of synergies between prediction tasks in complex systems. This potentially alleviates the need for system reduction or approximation, an attractive approach for holistic drug discovery and design. Furthermore, it is possible to use these new techniques and methodologies for improving graph-based pharmacophoric representations, fragment-based drug design, de novo drug design, binding energy predictions, and consensus classification models [18]. However, there are a number of caveats associated with these approaches that must be addressed in order for them to be fully mature.

De novo drug design and generative models

De novo drug design is one of the areas benefiting from DL. For example, DeepSARM is a deep learning extension of SARM for generative fragment-based analog design. DeepSARM [26] introduces chemical novelty into the design process based on recent developments in generative modeling adaptation and the further development of chemical language models. Iterative DeepSARM (iDeepSARM) [25] can rationally modify and extend sequence-to-sequence models and add iterative compound optimization and core-structure modifications.

Deep Graph Learning (DGL) which is based on ANNs, is capable of learning from graph-structured data [66]. It is included as part of the ProSurfScan platform developed by Chemotargets. This platform has been successfully applied to the identification of novel compounds for different targets. It yielded the first AI-designed drug for Huntington’s disease, which is currently in clinical trials [67]. ProSurfScan allows estimation of the compatibility and binding mode of fragments on different regions of a protein surface. Therefore, the protein surface is represented as a complete graph consisting of nodes with pharmacophoric features derived from the analysis of a triangulated mesh representation of the protein surface [68, 69]. Two complementary methods are employed to carry out the predictions. A clique detection algorithm is used to compare the protein surface with known surfaces associated with fragments from ligands present in structures from the Protein Data Bank (PDB) (aka fragment environments). This allows placement of the fragment based on the largest subgraph found between the fragment-environment and the protein surface. In addition, a series of DGL models is built using Graph Convolutional Neural Networks (GCNN) that estimate the compatibility of the fragments with respect to distinct regions of the protein surface.

Fernandez-de Gortari discussed the use of generators [16, 18] based on Variational Autoencoders (VAE), a deep neural network architecture. He discussed their advantage for constructing molecules with multi-target profiles and properties of pharmaceutical interest from lead molecule seeds. The methodology is based on using generators obtained from reasonable mutations of fragments [17], obtained by exchanging structurally similar fragments on the lead molecule seed based on a hypothetical continuous SAR for the development of a ML-based virtual screening classifier of Sarco(endo)plasmic reticulum Ca²⁺-ATPase (SERCA) inhibitors.

Machine learning for the prediction of ADME-Tox properties

Low efficacy associated with bioavailability problems and adverse drug effects have been recognized as one of the main causes of attrition during clinical trials [70]. Thus, the number of possible causes for a compound to fail or to have barely tolerable adverse effects is quite large. Moreover, in vitro and in vivo characterization of a compound’s properties can become very costly and time-consuming. For all of these reasons, considerable effort has been made to develop computational models for predicting ADME-Tox properties [70]. AI models have leveraged the information available in heterogeneous ADME-Tox data sets and helped to improve the accuracy of early drug efficacy and safety predictions. There is an increasing number of public and private sector initiatives aimed at the generation and evaluation of prospective models to assist decision-making processes and to generate future innovations for predicting ADME-Tox properties. Initiatives are also underway to permit public use and comparison of ML/DL models to increase confidence in and acceptance of these predictions. For example, Therapeutics Data commons (TDC) was introduced as a platform to systematically access and evaluate ML models across the entire range of therapeutics, accessible via an open python library [71, 72]. TDC encompasses AI-ready datasets and learning tasks for therapeutics; sets of tools to support data processing, model development, validation, and evaluation; and a collection of ‘leaderboards’ to support model comparison and benchmarking.

Other ML models derive hypothetical properties such as brain penetration (Kp) from limited experimental data or characterize in vivo properties from in vitro assay data. In a study conducted by Rodríguez-Pérez’s group, multitask learning based on Graph Neural Networks (MT-GNN) showed superior performance to other ML approaches based solely on in vitro brain penetration data [20]. These promising models have considerable potential for practical applications in other property prediction tasks.

To provide a partial solution to the data issues and improve early drug safety assessment, an effort has been made to integrate preclinical and post-marketing drug safety data with other commonly used sources of information, such as chemical structure data and preclinical assays. Current trends focus on developing novel systems approaches to drug safety that offer a more mechanistic view of predictive safety based on similarity to drug classes, interaction with secondary targets, and interference with biological pathways beyond the traditional identification of chemical fragments associated with selected toxicity criteria [53]. An example of the integration of this information is CLARITY_PV [73], a web platform for translational safety and pharmacovigilance studies that track side effects throughout all phases of the drug discovery and development process.

Importance of natural products in drug discovery

Natural products have historically contributed to drug discovery as a source of diverse, structurally complex bioactive molecules that have evolved to fulfill specific biological functions. However, drug development from NPs is more complex, costly, and inefficient than drug development from small molecules [74]. Similarly, the small amount of bioactivity data associated with NPs has limited potential applications of ML and DL in the study of naturally occurring compounds. Initiatives such as the NuBBE_DB, a virtual database of NPs and their derivatives from the Brazilian biodiversity [75, 76], have paved the way for developing new NP databases and projects like LOTUS [77] for NP storage, search, and analysis. A number of different chemoinformatics [78] and AI [32] applications have been proposed for analyzing the data collected to date. The main applications have focused on understanding the biological activity of NPs, carrying out the systematic search for bioactive NPs with respect to a molecular target of interest, and guiding the chemical synthesis of NP analogs with simplified structures and improved activity. The NuBBE_DB database has been expanded in collaboration with CAS (Chemical Abstracts Service). Currently, more than 54,000 substances are described with information on chemical, biological, and pharmacology data that can be explored in order to analyze their medicinal chemistry potential. Recent work on target predictions for compounds in the NuBBE_DB led to the identification of chalcones with potential application for the treatment of Chagas disease [79].

General opportunities

Access to AI technology and international networking can also accelerate the development of drugs for neglected diseases, Alzheimer’s disease, and antibiotic resistance. The research group of Oprea developed ML models to identify a potential gene relevant to susceptibility to Alzheimer’s disease [29]. This analysis also identified potential risk genes including FRRS1, CTRAM, SCGB3A1, FAM92B/CIBAR2, and TMEFF2.

Other chemoinformatics, ML, and DL models were proposed as a means of identifying compounds to combat antibiotic resistance, which is found in all parts of the world [80]. Peptides have been proposed as suitable alternatives since they display biological activity against bacteria, viruses, fungi, and parasites [81, 82]. Antimicrobial peptides (AMP) have a low propensity for bacteria resistance [83, 84]. The research group of Rondón-Villarreal [12] developed an AMP library using the CAMP_R3 [85] database, and genetic algorithms. The peptide library was designed with specific physicochemical properties (charge, hydrophobicity, isoelectric point, and stability index) and tested against Escherichia coli, Pseudomonas aeruginosa and methicillin-resistant Staphylococcus aureus. This library could potentially lead to the discovery of potent antimicrobial peptides.

However, the challenges of peptide design might require addressing multiple parameters such as high toxicity, poor oral bioavailability, thermal and pH stability, and functional promiscuity in concert. In addition, costs associated with experimental time, human resources, and equipment involved [13], must also be accounted for. Chemoinformatics, ML, and DL approaches should provide a means for developing safe AMPs with reduced toxicity, predict their antibacterial activity and drug-likeness profile, and accelerate antibiotic discovery [86, 87]. Plisson et al. [13] proposed an ML-guided discovery and design project related to non-hemolytic peptides. The workflow is composed of collecting compounds for an AMP database, computing 56 physicochemical descriptors; developing binary-classifier models to predict hemolytic nature and activity; estimating the domain of applicability, and applying optimized models to the discovery of non-hemolytic AMPs from a known database (e.g., APD3) or design novel sequences. The models used in this study include support vector machines, decision trees, random forest, gradient boosting, and k-nearest neighbors. This research is part of a growing series of predictive and generative ML models applied to support the discovery and design of bioactive peptides, including antimicrobial peptides [56, 63]. The authors applied multivariate outlier detection to delineate the boundaries of their predictive models (i.e., applicability domain) leading to the identification of outlying sequences [9]. To date, little work is being carried out on estimating the domain(s) of applicability of peptide modeling, although it is necessary for the parallel application of multiple predictors on a given sequence space.

Recommendations for new generations of scientists

Some speakers shared their experiences as scientists. This section summarizes some general recommendations for future scientists. The early-career scientist should choose topics that open new possibilities and should not adhere to a single approach or technology. “If you have your data, run your own benchmarks tests, build your own models, and try to interpret them in context. Metrics are irrelevant. The only proof is unbiased predictivity”.

One should always review the original publications to ensure integrity of information sources and avoid dilution or subjective bias. “Verify what you see, doubt what you find, and always obtain independent confirmation of your observations to validate your work”.

Do not be afraid to say, “I do not know.” Omniscient human beings are rare. Be ready to learn continuously. Focus on problem-solving skills; they are more important than static learning and memorization of facts. Always prize creativity and out-of-the-box thinking. As you progress in your career, you will learn that people are the most important asset. If someone “steals” your ideas, which does happen, remember that this is a form of flattery. It is not sufficient to only generate one great idea in your scientific life (the, indeed, it should be taken away …). Rather, one needs to generate new ideas continuously to cultivate individual creativity.