The automated design of three-dimensional (3D) molecular structures is a rapidly advancing field with major applications in drug discovery, catalysis, and materials science. Despite this progress, there remains a lack of standardized benchmarks for objectively evaluating and comparing generative methods for 3D molecular design beyond the domain of organic drug-like compounds. Here, we present 3DOpt, the first benchmark designed to assess the ability of generative methods to identify optimal 3D structures across the full chemical spectrum, including organometallic and transition-metal complexes. Each benchmark task is defined by a target 3D structure, a rigorously curated starting population of pre-evaluated candidate molecules, and a scoring function that quantifies similarity to the target in terms of both geometry and composition. We demonstrate the utility of 3DOpt by applying baseline generative methods and provide reference performance metrics for widely used molecular design strategies. Overall, 3DOpt establishes a general-purpose framework for the systematic evaluation of 3D molecular generative design methods across diverse chemical spaces.

A key challenge in drug discovery is the efficient search of chemical space to identify small molecules with desirable biological activity and developability. Computational approaches address this challenge by applying in silico evaluators to prioritise compounds for experimental validation. Traditionally, this has been achieved through brute-force virtual screening of large, predefined compound libraries. More recently, advances in generative molecular design (GMD) have enabled models that directly propose molecules optimised for specific objectives, offering favourable computational scaling but often at the expense of progressability, evaluability, and synthetic feasibility. In this review, we analyse molecular search strategies across a spectrum ranging from brute-force search to de novo generative design, focusing on how chemical space is specified and explored. We emphasise three practical axes-computational scaling, progressability, and evaluability-and discuss how different paradigms trade off these properties. In particular, we highlight Generative Virtual Screening (GVS) as a hybrid approach in which generative optimisation is constrained to a fixed, explicit compound library. By combining the computational efficiency of generative methods with the progressability and retrospective evaluability of library-based search, GVS enables scalable exploration of chemical space while providing ground-truth reference points for algorithmic assessment. We argue that GVS offers both immediate practical benefits for hit discovery and a controlled framework for developing and benchmarking generative search algorithms, facilitating their principled extension to broader and less constrained chemical spaces.

The angiotensin-converting enzyme 2 (ACE2) is pivotal as the cellular receptor for SARS-CoV-2 (severe acute respiratory syndrome coronavirus 2), the virus responsible for COVID-19. This study presents a novel synthetic route for four analogues of MLN-4760, a known inhibitor of ACE2, guided by in silico docking predictions. These synthetic advances enabled in vitro pIC50 assays confirming the inhibitory potency of the synthesized analogues. Lastly, this route was applied to the synthesis of novel 18F-labeled ACE2 inhibitors for PET imaging applications.

Chemical Language Models (CLMs) leveraging reinforcement learning (RL) have shown promise in de novo molecular design, yet often suffer from mode collapse, limiting their exploration capabilities. Inspired by Test-Time Training (TTT) in large language models, we propose scaling TTT for CLMs to enhance chemical space exploration. We introduce MolExp, a novel benchmark emphasizing the discovery of structurally diverse molecules with similar bioactivity, simulating real-world drug design challenges. Our results demonstrate that scaling TTT by increasing the number of independent RL agents follows a log-linear scaling law, significantly improving exploration efficiency as measured by MolExp. In contrast, increasing TTT training time yields diminishing returns, even with exploration bonuses. We further evaluate cooperative RL strategies to enhance exploration efficiency. These findings provide a scalable framework for generative molecular design, offering insights into optimizing AI-driven drug discovery.

Generative chemical language models (CLMs) have demonstrated success in learning language-based molecular representations for de novo drug design. Here, we integrate structure-based drug design (SBDD) principles with CLMs to go from protein structure to novel small-molecule ligands, without a priori knowledge of ligand chemistry. Using Augmented Hill-Climb, we successfully optimise multiple objectives within a practical timeframe, including protein-ligand complementarity. Resulting de novo molecules contain known or promising adenosine A2A receptor ligand chemistry that is not available in commercial vendor libraries, accessing commercially novel areas of chemical space. Experimental validation demonstrates a binding hit rate of 88%, with 50% having confirmed functional activity, including three nanomolar ligands and two novel chemotypes. The two strongest binders are co-crystallised with the A2A receptor, revealing their binding mechanisms that can be used to inform future iterations of structure-based de novo design, closing the AI SBDD loop.

Chemical language models, combined with reinforcement learning (RL), have shown significant promise to efficiently traverse large chemical spaces for drug discovery. However, the performance of various RL algorithms and their best practices for practical drug discovery are still unclear. Here, starting from the principles of the REINFORCE algorithm, we investigate the effect of different components from RL theory including experience replay, hill-climbing, baselines to reduce variance, and alternative reward shaping. We propose a new regularization method more aligned to REINFORCE than current standard practices, and demonstrate how RL hyperparameters can be fine-tuned for effectiveness and efficiency. Lastly, we apply our learnings to practical drug discovery by demonstrating enhanced learning efficiency on frontier binding affinity models by using Boltz2 as a reward model. We share our RL models used in the ACEGEN repository, and hope the experiments here act as a guide to researchers applying RL to chemical language models for drug discovery.

In recent years, reinforcement learning (RL) has emerged as a valuable tool in drug design, offering the potential to propose and optimize molecules with desired properties. However, striking a balance between capabilities, flexibility, reliability, and efficiency remains challenging due to the complexity of advanced RL algorithms and the significant reliance on specialized code. In this work, we introduce ACEGEN, a comprehensive and streamlined toolkit tailored for generative drug design, built using TorchRL, a modern RL library that offers thoroughly tested reusable components. We validate ACEGEN by benchmarking against other published generative modeling algorithms and show comparable or improved performance. We also show examples of ACEGEN applied in multiple drug discovery case studies. ACEGEN is accessible at https://github.com/acellera/acegen-open and available for use under the MIT license.

Generative models are undergoing rapid research and application to de novo drug design. To facilitate their application and evaluation, we present MolScore. MolScore already contains many drug-design-relevant scoring functions commonly used in benchmarks such as, molecular similarity, molecular docking, predictive models, synthesizability, and more. In addition, providing performance metrics to evaluate generative model performance based on the chemistry generated. With this unification of functionality, MolScore re-implements commonly used benchmarks in the field (such as GuacaMol, MOSES, and MolOpt). Moreover, new benchmarks can be created trivially. We demonstrate this by testing a chemical language model with reinforcement learning on three new tasks of increasing complexity related to the design of 5-HT2a ligands that utilise either molecular descriptors, 266 pre-trained QSAR models, or dual molecular docking. Lastly, MolScore can be integrated into an existing Python script with just three lines of code. This framework is a step towards unifying generative model application and evaluation as applied to drug design for both practitioners and researchers. The framework can be found on GitHub and downloaded directly from the Python Package Index. Scientific Contribution MolScore is an open-source platform to facilitate generative molecular design and evaluation thereof for application in drug design. This platform takes important steps towards unifying existing benchmarks, providing a platform to share new benchmarks, and improves customisation, flexibility and usability for practitioners over existing solutions.

Generative chemical language models have demonstrated success in learning language-based molecular representations for de novo drug design. Here, we integrate structure-based drug design (SBDD) principles with chemical language models to present a modern hit-finding workflow to go from protein structure to novel small-molecule ligands, without a priori knowledge of ligand chemistry. Using Augmented Hill-Climb we successfully optimised multiple objectives within a practical timeframe, including protein-ligand complementarity. Generated de novo molecules contained both known and promising adenosine A2A receptor ligand chemistry that is not available in commercial vendor libraries, accessing commercially novel areas of chemical space. Experimental validation identified three nanomolar ligands with confirmed functional activity, two of which contain novel chemotypes. Overall, demonstrating a binding hit rate of 88% with 50% of the binders demonstrating confirmed functional activity emphasising the complex relationships in translating binding to downstream pharmacology. Lastly, the two strongest binders were co-crystallised with the A2A receptor revealing their binding mechanisms that can be used to inform future iterations of structure-guided de novo design, closing the AI SBDD loop.

SMILES-based generative models are amongst the most robust and successful recent methods used to augment drug design. They are typically used for complete de novo generation, however, scaffold decoration and fragment linking applications are sometimes desirable which requires a different grammar, architecture, training dataset and therefore, re-training of a new model. In this work, we describe a simple procedure to conduct constrained molecule generation with a SMILES-based generative model to extend applicability to scaffold decoration and fragment linking by providing SMILES prompts, without the need for re-training. In combination with reinforcement learning, we show that pre-trained, decoder-only models adapt to these applications quickly and can further optimize molecule generation towards a specified objective. We compare the performance of this approach to a variety of orthogonal approaches and show that performance is comparable or better. For convenience, we provide an easy-to-use python package to facilitate model sampling which can be found on GitHub and the Python Package Index. Scientific contribution This novel method extends an autoregressive chemical language model to scaffold decoration and fragment linking scenarios. This doesn’t require re-training, the use of a bespoke grammar, or curation of a custom dataset, as commonly required by other approaches.

Retrained PIDGIN (https://github.com/BenderGroup/PIDGINv4) models using PIDGINv5 written in Python3 and completely refactored.

Generative molecular design for drug discovery and development has seen a recent resurgence promising to improve the efficiency of the design-make-test-analyse cycle; by computationally exploring much larger chemical spaces than traditional virtual screening techniques. However, most generative models thus far have only utilized small-molecule information to train and condition de novo molecule generators. Here, we instead focus on recent approaches that incorporate protein structure into de novo molecule optimization in an attempt to maximize the predicted on-target binding affinity of generated molecules. We summarize these structure integration principles into either distribution learning or goal-directed optimization and for each case whether the approach is protein structure-explicit or implicit with respect to the generative model. We discuss recent approaches in the context of this categorization and provide our perspective on the future direction of the field.

While a multitude of deep generative models have recently emerged there exists no best practice for their practically relevant validation. On the one hand, novel de novo-generated molecules cannot be refuted by retrospective validation (so that this type of validation is biased); but on the other hand prospective validation is expensive and then often biased by the human selection process. In this case study, we frame retrospective validation as the ability to mimic human drug design, by answering the following question: Can a generative model trained on early-stage project compounds generate middle/late-stage compounds de novo? To this end, we used experimental data that contains the elapsed time of a synthetic expansion following hit identification from five public (where the time series was pre-processed to better reflect realistic synthetic expansions) and six in-house project datasets, and used REINVENT as a widely adopted RNN-based generative model. After splitting the dataset and training REINVENT on early-stage compounds, we found that rediscovery of middle/late-stage compounds was much higher in public projects (at 1.60%, 0.64%, and 0.21% of the top 100, 500, and 5000 scored generated compounds) than in in-house projects (where the values were 0.00%, 0.03%, and 0.04%, respectively). Similarly, average single nearest neighbour similarity between early- and middle/late-stage compounds in public projects was higher between active compounds than inactive compounds; however, for in-house projects the converse was true, which makes rediscovery (if so desired) more difficult. We hence show that the generative model recovers very few middle/late-stage compounds from real-world drug discovery projects, highlighting the fundamental difference between purely algorithmic design and drug discovery as a real-world process. Evaluating de novo compound design approaches appears, based on the current study, difficult or even impossible to do retrospectively.Scientific Contribution This contribution hence illustrates aspects of evaluating the performance of generative models in a real-world setting which have not been extensively described previously and which hopefully contribute to their further future development.

MolScore is an open-source Python framework for scoring and evaluating molecules in the context of goal-directed generative models as used in de novo drug design. MolScore includes many relevant scoring functions for de novo drug design such as molecular similarity, docking software, predictive models, and synthesizability, as well as commonly used performance metrics to evaluate generative model performance based on chemistry generated. Integration into an existing generative model framework is simple, requiring just three lines of code, and graphical user interfaces are provided for objective configuration and for monitoring de novo molecules generated. As a real-world demonstration of its use, we use it to design selective 5-HT2a ligands using 266 pre-trained off-target predictive models, as well as docking into two co-crystal structures. MolScore can also be used for generative model evaluation as we demonstrate by analysing and selecting fine-tuning epochs of an RNN-based generative model. Moreover, the use of configuration files allows the sharing of objectives within the community for the purposes of reproducibility, comparison, and benchmarking; making it easier to propose drug discovery relevant objective functions as benchmark tasks. The code is freely available and hosted on GitHub, https://github.com/MorganCThomas/MolScore.

De novo molecule generation can suffer from data inefficiency; requiring large amounts of training data or many sampled data points to conduct objective optimization. The latter is a particular disadvantage when combining deep generative models with computationally expensive molecule scoring functions (a.k.a. oracles) commonly used in computer-aided drug design. Recent works have therefore focused on methods to improve sample efficiency in the context of de novo molecule drug design, or to benchmark it. In this work, we discuss and adapt a recent sample efficiency benchmark to better reflect realistic goals also with respect to the quality of chemistry generated, which must always be considered in the context of small-molecule drug design; we then re-evaluate all benchmarked generative models. We find that accounting for molecular weight and LogP with respect to the training data, and the diversity of chemistry proposed, re-orders the ranking of generative models. In addition, we benchmark a recently proposed method to improve sample efficiency (Augmented Hill-Climb) and found it ranked top when considering both the sample efficiency and chemistry of molecules generated. Continual improvements in sample efficiency and chemical desirability enable more routine integration of computationally expensive scoring functions on a more realistic timescale.

The EAG (ether-à-go-go) family of voltage-gated K+ channels are important regulators of neuronal and cardiac action potential firing (excitability) and have major roles in human diseases such as epilepsy, schizophrenia, cancer, and sudden cardiac death. A defining feature of EAG (Kv10–12) channels is a highly conserved domain on the N terminus, known as the eag domain, consisting of a Per–ARNT–Sim (PAS) domain capped by a short sequence containing an amphipathic helix (Cap domain). The PAS and Cap domains are both vital for the normal function of EAG channels. Using heme-affinity pulldown assays and proteomics of lysates from primary cortical neurons, we identified that an EAG channel, hERG3 (Kv11.3), binds to heme. In whole-cell electrophysiology experiments, we identified that heme inhibits hERG3 channel activity. In addition, we expressed the Cap and PAS domain of hERG3 in Escherichia coli and, using spectroscopy and kinetics, identified the PAS domain as the location for heme binding. The results identify heme as a regulator of hERG3 channel activity. These observations are discussed in the context of the emerging role for heme as a regulator of ion channel activity in cells.

Artificial intelligence (AI) has undergone rapid development in recent years and has been successfully applied to real-world problems such as drug design. In this chapter, we review recent applications of AI to problems in drug design including virtual screening, computer-aided synthesis planning, and de novo molecule generation, with a focus on the limitations of the application of AI therein and opportunities for improvement. Furthermore, we discuss the broader challenges imposed by AI in translating theoretical practice to real-world drug design; including quantifying prediction uncertainty and explaining model behavior.

A plethora of AI-based techniques now exists to conduct de novo molecule generation that can devise molecules conditioned towards a particular endpoint in the context of drug design. One popular approach is using reinforcement learning to update a recurrent neural network or language-based de novo molecule generator. However, reinforcement learning can be inefficient, sometimes requiring up to 10^5 molecules to be sampled to optimize more complex objectives, which poses a limitation when using computationally expensive scoring functions like docking or computer-aided synthesis planning models. In this work, we propose a reinforcement learning strategy called Augmented Hill-Climb based on a simple, hypothesis-driven hybrid between REINVENT and Hill-Climb that improves sample-efficiency by addressing the limitations of both currently used strategies. We compare its ability to optimize several docking tasks with REINVENT and benchmark this strategy against other commonly used reinforcement learning strategies including REINFORCE, REINVENT (version 1 & 2), Hill-Climb and best agent reminder. We find that optimization ability is improved ~1.5-fold and sample-efficiency is improved ~45-fold compared to REINVENT while still delivering appealing chemistry as output. Diversity filters were used, and their parameters were tuned to overcome observed failure modes that take advantage of certain diversity filter configurations. Lastly, we find that Augmented Hill-Climb outperforms the other reinforcement learning strategies used on six tasks, especially in the early stages of training or for more difficult objectives. Overall, we hence show that AHC improves sample-efficiency for language-based de novo molecule generation conditioning via reinforcement learning, compared to the current state-of-the-art. This makes more computationally expensive scoring functions, such as docking, more accessible on a relevant timescale.

PARP inhibitors are synthetically lethal with BRCA1/2 mutations, and in this setting, accumulation of DNA damage leads to cell death. Because increased DNA damage and subsequent immune activation can prime an anti-tumor immune response, we studied the impact of olaparib +/- immune checkpoint blockade (ICB) on anti-tumor activity and the immune microenvironment. Concurrent combination of olaparib, at clinically relevant exposures, with ICB gave durable and deeper anti-tumor activity in the Brca1m BR5 model vs. monotherapies. Olaparib and combination treatment modulated the immune microenvironment, including increases in CD8+ T cells and NK cells, and upregulation of immune pathways, including type I IFN and STING signaling. Olaparib also induced a dose-dependent upregulation of immune pathways, including JAK/STAT, STING and type I IFN, in the tumor cell compartment of a BRCA1m (HBCx-10) but not a BRCA WT (HBCx-9) breast PDX model. In vitro, olaparib induced BRCAm tumor cell-specific dendritic cell transactivation. Relevance to human disease was assessed using patient samples from the MEDIOLA (NCT02734004) trial, which showed increased type I IFN, STING, and JAK/STAT pathway expression following olaparib treatment, in line with preclinical findings. These data together provide evidence for a mechanism and schedule underpinning potential benefit of ICB combination with olaparib.

An efficient macrocyclisation approach based on the double aromatic nucleophilic substitution (SNACK) was developed. This methodology allows a facile incorporation of heterocyclic motifs into macrocyclic rings and rapid synthesis of a significant number of structurally diverse macrocycles. SNACK macrocyclisation enables preparation of stable diastereoisomers of conformationally restricted macrocycles (atropisomers). Practical application of SNACK macrocyclisation in a drug discovery project was exemplified by the identification of high affinity macrocyclic binders of B-cell lymphoma 6 (BCL6).