Portal de Periódicos da CAPES

Real-World Applications and Experiences of AI/ML Deployment for Drug Discovery

2025; American Chemical Society; Linguagem: Inglês

10.1021/acs.jmedchem.4c03044

1520-4804

William R. Pitt, Jonathan Bentley, Christophe Boldron, Lionel Colliandre, Carmen Esposito, Elizabeth H. Frush, J. Kopec, Stéphanie Labouille, Jérôme Meneyrol, D. Pardoe, Ferruccio Palazzesi, Alfonso Pozzan, Jacob M. Remington, René Rex, Michelle Southey, Sachin Vishwakarma, Paul Walker,

Innovative Microfluidic and Catalytic Techniques Innovation

InfoMetricsFiguresRef. Journal of Medicinal ChemistryASAPArticle This publication is free to access through this site. Learn More CiteCitationCitation and abstractCitation and referencesMore citation options ShareShare onFacebookX (Twitter)WeChatLinkedInRedditEmailJump toExpandCollapse EditorialJanuary 8, 2025Real-World Applications and Experiences of AI/ML Deployment for Drug DiscoveryClick to copy article linkArticle link copied!Will R. Pitt*Will R. PittMolecular Architects, Evotec Ltd, Dorothy Crowfoot Hodgkin Campus, 114 Innovation Drive, Milton Park, Abingdon, Oxfordshire OX14 4RZ, U.K.*Email: [email protected]More by Will R. Pitthttps://orcid.org/0000-0001-8164-4550Jonathan BentleyJonathan BentleyDiscovery Chemistry, Evotec Ltd, Dorothy Crowfoot Hodgkin Campus, 114 Innovation Drive, Milton Park, Abingdon, Oxfordshire OX14 4RZ, U.K.More by Jonathan BentleyChristophe BoldronChristophe BoldronMolecular Architects, Evotec SAS, Campus Curie, 195, route d'Espagne, Toulouse 31095, FranceMore by Christophe BoldronLionel ColliandreLionel Colliandrein silico R&D, Evotec SAS, Campus Curie, 195 route d'Espagne, 31100 Toulouse, FranceMore by Lionel ColliandreCarmen EspositoCarmen Espositoin silico R&D, Aptuit Srl, Via Alessandro Fleming, 4, 37135 Verona, ItalyMore by Carmen EspositoElizabeth H. FrushElizabeth H. FrushMolecular Architects, Evotec Inc., 303B College Road East, Princeton, New Jersey 08540, United StatesMore by Elizabeth H. Frushhttps://orcid.org/0000-0003-3611-132XJola KopecJola Kopecin silico R&D, Evotec Ltd, Dorothy Crowfoot Hodgkin Campus, 114 Innovation Drive, Milton Park, Abingdon, Oxfordshire OX14 4RZ, U.K.More by Jola KopecStéphanie LabouilleStéphanie LabouilleMolecular Architects, Evotec Ltd, Dorothy Crowfoot Hodgkin Campus, 114 Innovation Drive, Milton Park, Abingdon, Oxfordshire OX14 4RZ, U.K.More by Stéphanie LabouilleJerome MeneyrolJerome MeneyrolMolecular Architects, Evotec SAS, Campus Curie, 195, route d'Espagne, Toulouse 31095, FranceMore by Jerome MeneyrolDavid A. PardoeDavid A. PardoeMolecular Architects, Evotec Ltd, Dorothy Crowfoot Hodgkin Campus, 114 Innovation Drive, Milton Park, Abingdon, Oxfordshire OX14 4RZ, U.K.More by David A. Pardoehttps://orcid.org/0009-0005-0807-2994Ferruccio PalazzesiFerruccio Palazzesiin silico R&D, Aptuit Srl, Via Alessandro Fleming, 4, 37135 Verona, ItalyMore by Ferruccio PalazzesiAlfonso PozzanAlfonso PozzanMolecular Architects, Aptuit Srl, Via Alessandro Fleming, 4, 37135 Verona, ItalyMore by Alfonso PozzanJacob M. RemingtonJacob M. RemingtonMolecular Architects, Evotec Inc., 303B College Road East, Princeton, New Jersey 08540, United StatesMore by Jacob M. RemingtonRené RexRené RexEvotec International GmbH, Marie-Curie-Str. 7, Göttingen D-37079, GermanyMore by René RexMichelle SoutheyMichelle SoutheyMolecular Architects, Evotec Ltd, Dorothy Crowfoot Hodgkin Campus, 114 Innovation Drive, Milton Park, Abingdon, Oxfordshire OX14 4RZ, U.K.More by Michelle SoutheySachin VishwakarmaSachin Vishwakarmain silico R&D, Evotec SAS, Campus Curie, 195 route d'Espagne, 31100 Toulouse, FranceMore by Sachin VishwakarmaPaul WalkerPaul WalkerCyprotex Discovery Ltd, No. 24 Mereside, Alderley Park, Macclesfield, Cheshire SK10 4TG, U.K.More by Paul WalkerOpen PDFJournal of Medicinal ChemistryCite this: J. Med. Chem. 2025, XXXX, XXX, XXX-XXXClick to copy citationCitation copied!https://pubs.acs.org/doi/10.1021/acs.jmedchem.4c03044https://doi.org/10.1021/acs.jmedchem.4c03044Published January 8, 2025 Publication History Received 11 December 2024Published online 8 January 2025editorialPublished 2025 by American Chemical Society. This publication is available under these Terms of Use. Request reuse permissionsThis publication is licensed for personal use by The American Chemical Society. ACS PublicationsPublished 2025 by American Chemical SocietySubjectswhat are subjectsArticle subjects are automatically applied from the ACS Subject Taxonomy and describe the scientific concepts and themes of the article.Bioinformatics and computational biologyDrug discoveryMedicinal chemistryOptimizationStructure activity relationshipThe emergence of artificial intelligence (AI) and machine learning (ML) in the field of drug discovery has been propelled by significant advances in computer science, infrastructure, and the surge of "big data". There is also an expectation that AI-related progress in other fields, such as virtual assistants, image generation, autonomous vehicles, and protein structure prediction, can be replicated elsewhere. The continuous desire to bring novel treatments to market has driven drug discovery companies, including large pharmaceutical firms, biotechs, and contract research organizations (CROs), to deploy AI/ML technology to both strengthen and accelerate drug pipelines. These companies face the decision of whether to build or to buy, either to invest in internal staff and infrastructure and establish in-house capabilities or to collaborate with AI-enabled companies. (1) It is noteworthy that the use of ML in medicinal chemistry began more than 40 years ago. (2) However, with recent advances in the field, particularly the rise of deep learning, these methods are now impacting every stage of the drug discovery process, from early target identification, to hit finding and lead optimization. Examples include virtual screening (VS) of ultralarge chemical databases for hit identification, ML models that predict potency and other relevant end points, as well as generative design algorithms that build molecular structures from scratch. In this paper we will present our perspective as a CRO involved in drug discovery (and development) partnerships. Given the competitive landscape, organizations such as ours need to stay abreast of AI/ML technological advancements because potential partners seek the advantage of integrating these tools into their discovery projects to guide the generation and exploitation of high-quality experimental data. For us, the commitment to do this is crucial to ensure a comprehensive and robust drug discovery process.However, the accurate prediction of experimental data remains challenging due to the intrinsic complexity of biological systems, the availability of quality training data, and the limited ability of chemical descriptors to fully capture the nature of chemical interactions. There are also cultural challenges in the adoption of AI. (3) The inherent biases in decision-making within drug discovery are well documented. (4,5) Such biases can hinder progress and prevent the integration of AI/ML technologies because they implicitly challenge well-established working practices. The situation is further complicated by the often-exaggerated claims regarding the effectiveness of AI/ML technologies and their impact in accelerating the drug discovery process. It remains premature to draw definitive conclusions as the market has not yet witnessed the introduction of a treatment developed solely through AI/ML methods. (6)In our experience, blending well-established in silico approaches, AI/ML technologies, and human experience produces the best outcomes. The decision to enhance our in-house capabilities was consistent with our company's ethos of innovation. Building our own capabilities provides a cost-efficient opportunity to evaluate and/or develop the most appropriate AI/ML technologies and foster internal talent development. Our organization covers the whole process from early target discovery to clinical trial support, including a broad range of therapeutic modalities. AI/ML is impacting our work in many ways, including aiding therapeutic antibody (7) and small molecule targeted degrader design. However, in this paper we focus on small molecule medicinal chemistry from hit identification to late lead optimization.AI/ML Methodologies and ApplicationsClick to copy section linkSection link copied!Briefly summarized below are our and others' experiences with the AI/ML applications that currently have the greatest impact on our work. Machine Representations of Chemical SpaceUsing deep learning to represent chemical space is a major recent development in chemical informatics. Compounds can now be represented by vectors, generated through training deep neural networks on large compound databases. Such representations are termed latent space because they are derived mathematically from the data set and encapsulate its essential features. A given vector (position in this latent space) can be decoded into a chemical structure, which is a great benefit over older representations like molecular fingerprints. It enables the rapid identification of compounds of interest in new regions. For instance, the interpolation between vectors allows the exploration of intermediate chemical structures, which can be a way to move into patentable chemical space.One of the pioneering examples is Continuous and Data-Driven Descriptors (CDDD), (8) which we have used extensively for generating compound designs (see other ways in the section on Generative Design below). CDDD, which is an autoencoder (AE), is simultaneously trained on SMILES (9) and constrained by chemical properties (e.g., polar surface area and lipophilicity) that push chemically and physically similar molecules into similar latent subspaces. This way of training predisposes the representations for transfer learning (TL), i.e. changing the task of a pretrained model by adding new, project-specific training data and thereby focusing on project-specific objectives and chemical properties. (10,11) The linkage of molecular similarity and calculated properties provided by this AE architecture is another advantage over fingerprints.We have developed our own in-house AE-based Seq2Seq (12) models, utilizing both recurrent neural networks (RNNs) (13) and transformer architectures. (14−17) By training these models on data sets curated in-house, we have achieved improved performance and flexibility for downstream tasks. The improvements include coverage of compounds with molecular weight greater than 600 Da, which is necessary for some projects. They also include the extraction of the latent features of molecules for quantitative structure–activity relationship (QSAR) model building. Combining QSAR and deep generative chemistry (DGC) in the same latent space, we employ optimization algorithms such as Bayesian optimization (BO) (18,19) and particle swarm optimization (PSO) (20) to perform inverse QSAR (21)/inverse design. (2) This means we can generate compound designs which are optimized against QSAR model predictions.The quality of these representations is critical, as it directly impacts the reliability and accuracy of subsequent applications. We validate our representation models based on their DGC SMILES validity, novelty, and drug-likeness, along with metrics quantifying QSAR performance and the smoothness of latent space objective functions. (22) Together, these validations allow our scientists to make informed decisions and build ML models with confidence. Machine Learning (ML)In this section, we briefly describe how we use ML to predict activity and absorption, distribution, metabolism, excretion, and toxicity (ADMET) end points (23) and physicochemical properties of compounds directly from the molecular structure ─ approaches commonly referred to as QSAR and quantitative structure–property relationship (QSPR) modeling, respectively.The quality of predictive models depends on the quality of the training data. Our experimental data, generated through standardized assays, are carefully curated to remove unreliable or inconsistent measurements. These assays include logD, aqueous solubility, Caco2 permeability, microsomal clearance, and hERG channel inhibition. Specific curation processes are implemented for both regression (continuous predictions) and classification tasks (discrete predictions), ensuring that only high-quality data are used. To streamline ML activities and facilitate the regular training and updating of models, we have implemented an automated ML workflow that encompasses chemical structure preparation, descriptors calculation, model selection, hyperparameter optimization, and model delivery. ML-generated predictions are finally interpreted using explainability techniques, which estimate the contribution of the input features to the model decision. (24)In recent years, the application of deep learning techniques to QSAR/QSPR modeling has shown great promise. Graph Neural Networks (GNN) in particular, have been shown to outperform traditional ML algorithms like Random Forest (RF) for certain end points. (25,26) However, in our experience with data sets typically spanning from a few hundred to over ten thousand data points, traditional ML algorithms usually outperform deep learning models. Nonetheless, GNNs have proven useful to enhance the performance and robustness of models when applied to larger data sets.Predictive QSAR and QSPR models play a pivotal role in discovery projects, aiding compound idea selection and prioritization. One application in this context is scoring functions for our generative tools. Generative DesignThe use of DGC to design compounds with targeted properties has recently emerged as a powerful approach in medicinal chemistry. Our previous review (27) identified over 100 deep learning de novo design methods published between 2017 and 2020. Since then, the explosion of interest in this topic has made it hard to keep track of all the new articles. We find that these papers often lack a real-world application perspective, since many researchers are not fortunate enough to be able to synthesize and test their designs. We make use of our opportunity and routinely and successfully employ state-of-the-art 2D and 3D DGC tools to design compounds that are then made and tested. Due to time constraints on the vetting of new tools, we focus on methods from reputable sources.One tool we have adopted and modified based upon internal feedback is REINVENT. (28,29) This is a reinforcement learning method which generates compound designs with improved scores using a positive feedback loop. Our findings suggest that its ability to generate relevant molecules can be highly connected to the scoring components used to drive toward project specific goals. In particular, 3D components like pharmacophore-based matches or docking scores, produce designs in the desired chemical space more rapidly than using 2D scores alone. (30) In subsequent iterations, advanced QSAR models of physicochemical properties and ADMET end points and more standard computational chemistry tools can be used to improve the properties of the generated compounds. In agreement with other authors, (31) we find that the use of these generative tools cannot be simplified to a simple button-clicking exercise.Postprocessing of the results obtained by any generative tools is crucial, for three main reasons. First, some scoring components can only be used a posteriori, due to their intrinsic computational cost. Examples of these scoring methods include relative binding free energies (RBFE) (32) and fragment molecular orbital (FMO) interaction energies. (33) Second, deep generative tools cannot always optimize multiple components simultaneously, and for this reason, some of them must be applied sequentially during the postprocessing stage. For instance, methods which grow ligands within a pocket usually focus on enthalpic contributions to potency, such as protein–ligand interactions. Finally, medicinal chemistry projects evolve over time, and so do the targeted compound properties. Given the importance of this postprocessing step, we are developing automated pipelines that integrate conventional computational chemistry, AI/ML, and physics-based calculations to speed up this process (see Computational Pipeline below). Protein ModelingAn accurate protein model can be incredibly useful for a drug discovery project. Usually, such models are obtained using experimental methods such as X-ray crystallography or cryogenic electron microscopy (cryo-EM). Until very recently, only non-AI methods were used to build homology models of proteins for which experimental models were not available. However, recently AlphaFold 2 (AF2), a member of a family of methods for predicting protein structure utilizing AI, demonstrated remarkable accuracy in its predictions. (34) Our local installation is a great resource for generating models for iterative protein construct design and preparation of models to be fitted into experimentally obtained density. We combined AF2 and ProteinMPNN (35) to increase protein stability and production yield. This approach can transform projects where it is only possible to isolate miniscule amounts of protein. The ability of AF Multimer (36) to predict the 3D structures of protein–protein complexes helps structural biologists to obtain initial models of the targets. Such models can be fitted into the experimental density and further refined. Novel complexes can be modeled using FoldDock, (37) which optimizes multiple sequence alignments for AlphaFold multimer run, producing better predictions based on a score separating acceptable and incorrect models.The AlphaFoldDB (34,38) database of AF2 models, made available by DeepMind and hosted by the EBI, combined with our installation of AF Multimer, are tremendous resources for many aspects of drug design from target ligandability estimation to VS and docking. However, we aim to produce our own experimental structures for our drug targets in complex with ligands of interest. When this is not possible, we often build homology models using classical methods in the presence of a known ligand so that side chains in the binding site are in a suitable conformation for docking.Recent advances in deep learning have also enabled prediction methods for ligand-protein complexes. Methods such as RoseTTAFold-AllAtom, (39) Umol, (40) and AF3 (41) claim to predict structural details of target proteins' interactions with small molecule ligands, metal ions, nucleic acids, and covalent binders with a precision surpassing established docking methods. We watch for developments in this area with great interest. Active LearningMedicinal chemistry often operates with limited experimental data. This is especially true for the hit-to-lead phase of projects working on novel targets. Where data are thin on the ground and expensive to generate, active learning (AL) can be very useful because its purpose is to generate sufficient data in the most efficient manner. To be precise, AL is a ML-based strategy that aims to maximize learning performance with respect to a specific task (objective function) with minimal data. The algorithm iteratively selects from a predefined pool of unlabeled items (in this case compound ideas) according to a so-called acquisition function, that balances exploitation (selection of the most promising, based on current knowledge) and exploration (selection from less known or unknown regions of chemical space to enhance the model's overall knowledge). (42) Analogously, BO seeks to identify the next compounds to test within a fully defined parameter space to find the optimum of the objective, (18,19) which in this context could be a multiparameter optimization (MPO) score. These MPO scores can contain primary assay components with more data points such as potency, lipophilicity, metabolic stability, and permeability measurements or follow on assays with fewer data points such as off-target activity against enzymes, receptors, and transporters, depending on the project requirements. In medicinal chemistry, AL is used to guide the selection of informative compounds from the vast chemical space. (43,44) We use AL both to enable the VS of ultralarge make-on-demand compound libraries such as Enamine REAL (45) and to reduce the number of compounds needed to reach project goals.Traditional structure-based and ligand-based methods are too computationally expensive and time-consuming for the brute-force screening of billions of compounds. (46) Additionally, VS costs increase with the complexity of the scoring function. Our solution, built on the open source MolPal, (46) combines BO with VS tools and advanced molecular dynamics (MD)-based scoring functions to focus exploration on the highest performing compounds.The Design-Make-Test-Analyze (DMTA) cycle can be configured as an AL process that explores chemical space. (47) We employ BO in this way, assisting in the decision-making process by selecting compounds to make and validate experimentally. The selection of informative compounds like this should ultimately lead to a reduction in the number of cycles. In its AL form, BO ranks a predefined list of compounds coming from other tools or from medicinal chemists' ideas. While this approach limits the exploratory capacity, it can increase the acceptance of the proposed solutions by drug designers and reduce the search space to a more manageable size. In its generative form, BO proposes new points to test in a machine-based chemical space representation (see section above). The proposed points must be decoded to chemical structures. These designs can challenge the mindset of the team and avoid unwanted human bias. However, they are not always very easy to synthesize. (8,48,49) Feedback from the medicinal chemistry team can highlight synergistic opportunities for improvement, e.g., flagging multiple designs derived from a single, outlier result and new opportunities arising from synthetic improvements. Synthetic Tractability and Retrosynthesis PredictionThe synthesis of compounds, or the "Make" phase, is often the rate-limiting step in the DMTA cycle. (50) Therefore, synthetic tractability is a key aspect of the "Design" phase. This applies to human and AI-generated designs alike. Currently most generative design tools do not explicitly encode this criterion in their algorithms for growing or scoring compounds. However, one of the most exciting developments in this domain has been the invention of AI computer-aided synthesis planning (CASP) tools. (51) This has enabled the scoring or filtering by synthetic tractability using full blown retrosynthesis analysis or faster ML models. (52) Medicinal chemists usually design compounds with a synthetic route in mind or at least with a mental estimate of the difficulty involved. The most advanced AI tools have not yet reached the sophistication and efficiency of a team of medicinal chemists sharing expertise and knowledge daily, for example on the availability and reactivity of building blocks and intermediates. However, the addition of in-house data, such as from electronic laboratory notebooks (ELNs) and building block inventories, does increase the effectiveness of the tools. (53,54) AI retrosynthesis is increasingly being used by medicinal and computational chemists, for example for scaffold-hopping, inspiration, and easier planning of simple routes. As with AI in other areas, retrosynthesis output can make a disappointing first impression, if parity with users' own expertise and specific experience is expected. (50,54) A commercial AI CASP tool via a web interface is used by our chemists for inspiration or to cross-check their route planning; they find its quick and easy links to background literature very useful. Evaluation of tools, some of which are very expensive, has proved difficult for us, perhaps because we had unrealistic expectations of performance. For generative design workflows, ML synthetic complexity scores have some utility, but we always apply a manual assessment of synthetic tractability as one of the last steps. Safety AssessmentIn addition to synthetic tractability, the safety risks of a given compound design must also be considered. Safety remains a major concern for drug development programs. Often, safety risks become apparent only late in drug development after significant resources have been deployed. Hence, increasingly AI/ML approaches, which can flag safety risks earlier and more cheaply, are receiving considerable attention. (55) Pure in silico models have been developed for example to reduce the probability of Drug-Induced Liver Injury (DILI), based on compound descriptors like the number of carbon atoms in sp3 hybridization. (56) In silico models are desirable as they can aid design prior to synthesizing compounds, potentially reducing the costs associated with exploratory safety assays. These models tend to be either rule-based or employ traditional supervised ML algorithms. (56,57) However, to increase predictive performance, it is beneficial to incorporate in vitro data (e.g., bile salt export pump (BSEP) transporter inhibition and cellular cytotoxicity data) to build more sophisticated systems, such as Bayesian models. (58)In contrast to individual in vitro assays, which cover only limited aspects of toxicity, omics technologies provide a more comprehensive snapshot of the cellular state in response to drug exposure. Fortunately, new high-throughput omics technologies allow the creation of data sets that are of sufficient size to train AI models. (59) These models can identify complex patterns in the omics profiles which are associated with adverse outcomes resulting in organ toxicity. Once trained, they can predict the toxicity risk of new compounds with high accuracy, outperforming existing in vitro methods. (60) Moreover, this approach is not limited to small molecules but works equally well for other modalities including biologics. To create a training data set for our AI model, we utilized our high-throughput transcriptomics platform (ScreenSeq) to produce a database of transcriptomics profiles obtained from cell models. The profiles generated by hundreds of well-characterized compounds of different types serve as useful reference points. Computational PipelinesThe advent of de novo design methods, particularly deep generative AI methods, has increased the need to evaluate and prioritize large numbers of virtual compounds. This is often done by applying predictive models alongside simpler calculated property and/or more sophisticated physics-based scores. Each virtual molecule is scored according to multiple criteria (drug-likeness, predicted activity and ADMET attributes, novelty, physicochemical properties, synthetic tractability, etc.), then the different scores are aggregated using an ad-hoc, project specific MPO function. When correctly parametrized, this MPO score can be used to rank virtual molecules and prioritize the most promising compounds for the next round of synthesis. One technical challenge when deploying such a pipeline is the orchestration between the different tasks, given the number and diversity of tools that are typically involved. A good orchestrator needs to be able to interface between different file formats, handle multiple environments, manage resources efficiently, scale-up jobs where needed and be robust. Because the AI/ML field is evolving rapidly, the orchestrator also needs to be designed in a way that makes it easy to add new components or change the infrastructure it is deployed on.Automation of the DMTA cycle can save time and resources, while encoding best practices and improving reproducibility, which facilitates objectivity in the selection of designs for synthesis. BRADSHAW (61) was designed so that human and machine-generated ideas are processed in the same way, thus trying to avoid some of the unwanted biases that can be involved in selection. There are several commercial (62−64) and open source (65−67) platforms that were designed with automating drug design in mind. We are very much influenced by the work of Green et al. (61) and Besnard et al. (47) and seek to automate our workflows wherever possible, using Knime or our in-house high-performance computing (HPC) pipelining solution. We face the same challenges cited by the authors of BRADSHAW (61) of integration, robustness, simplicity, and flexibility. Each pipeline needs to be adapted to the ever-changing needs of a project while at the same time be reusable, at least in parts, by other projects. AI in the Context of Medicinal Chemistry ProjectsThe emergence of AI design tools, combined with the increasing impact of physics-based methods and lower HPC costs, has led some pharma companies to explore different ways of working. (61,68) At Evotec, we have an AI/ML research and development group (in silico R&D or isRD) responsible for adapting and integrating cutting-edge techniques into our technology stack, and an operational group (Molecular Architects or MAs), who apply these techniques to discovery projects in collaboration with the chemistry team and our partners. The concept of MAs (illustrated in Figure 1) is to fuse medicinal and computational chemistry experience and expertise, working on a foundation of data science and in silico tools. We consider it a powerful facilitator for establishing trust, attaining ambitious goals, and expediting the discovery of potential drug candidates. MAs ensure that (i) the right tools and methods are used, irrespective of their origin and whether they use AI/ML or not, (ii) that the data are clean and well understood, (iii) that project objectives are clear and met, and (iv) bespoke computational pipelines combined with efficient operational DMTA workflows are created to test the given design hypothesis with the minimum number of compounds.Figure 1Figure 1. Secret sauce of excellence in molecular design at Evotec.High Resolution ImageDownload MS PowerPoint SlideThe concept of D2MTL (Design-Decide-Make-Test-Learn) was introduced by MAs as an evolution of the well-establish