Designing Drugs with AI: A New Recipe for Discovery

Author: Denis Avetisyan

Researchers have developed a novel AI framework that leverages chemical building blocks and reinforcement learning to generate promising new drug candidates.

The ReACT-Drug pipeline demonstrates an illustrative example of reasoning and acting for drug discovery.

ReACT-Drug utilizes protein and molecular embeddings with reaction templates to achieve target-agnostic de novo drug design with competitive binding affinities and drug-like properties.

Despite advances in computational methods, discovering novel drug candidates within vast chemical spaces remains a significant challenge. This is addressed in ‘ReACT-Drug: Reaction-Template Guided Reinforcement Learning for de novo Drug Design’, which introduces a target-agnostic framework leveraging reinforcement learning to generate synthetically accessible molecules. By integrating protein embeddings, fragment-based search spaces, and reaction-template guided molecular transformations, ReACT-Drug efficiently produces de novo drug candidates with competitive binding affinities and drug-like properties. Could this approach unlock a new era of automated, rational drug design, accelerating the discovery of life-saving therapeutics?

The Challenge of Discovery: Beyond Chance and Screening

The historical process of discovering new pharmaceuticals has long been characterized by substantial costs and protracted timelines, frequently depending as much on chance observations as on rational design. Early successes often arose from accidental findings – serendipity – where a compound’s unexpected biological effect sparked further investigation. More recently, high-throughput screening has become a dominant approach, involving the automated testing of vast libraries of compounds against biological targets. While significantly accelerating the pace of investigation, this method remains incredibly resource-intensive, requiring substantial investment in robotics, reagents, and data analysis. The sheer volume of compounds tested often yields many false positives, necessitating extensive validation – and even with these advanced techniques, the average time from target identification to approved drug can still exceed a decade, with billions of dollars spent per successful medication.

The sheer scale of potential drug candidates presents a formidable obstacle to modern pharmaceutical research. Chemical space, encompassing all theoretically possible small molecules, is estimated to contain upwards of 10⁶⁰ compounds – a number dwarfing any conceivable capacity for traditional screening. Consequently, a brute-force approach to identifying promising drug leads is not only financially prohibitive but computationally impossible. This necessitates the development of intelligent design strategies, leveraging computational methods like machine learning and artificial intelligence to navigate this vast landscape. These techniques prioritize the exploration of chemical space based on predicted properties and biological activity, effectively focusing resources on the most promising candidates and accelerating the discovery of novel therapeutics. Rather than randomly sampling possibilities, researchers now aim to proactively design molecules with a high probability of success.

Contemporary de novo drug design faces a significant trilemma: simultaneously optimizing for a molecule’s ‘drug-likeness’ – its absorption, distribution, metabolism, and excretion properties – while ensuring it can be realistically synthesized in a laboratory, and crucially, that it effectively interacts with the intended biological target. Many computational approaches excel at predicting activity or optimizing for physicochemical properties, but often falter when considering the practicalities of chemical synthesis; a theoretically potent molecule remains useless if it cannot be created. Conversely, methods prioritizing synthetic accessibility may generate compounds with suboptimal binding affinity or poor pharmacokinetic profiles. This inherent tension necessitates increasingly sophisticated algorithms and scoring functions capable of navigating this complex multi-objective optimization landscape, representing a major hurdle in the rational design of novel therapeutics.

This de novo molecular drug discovery pipeline leverages reinforcement learning with a PPO agent to iteratively optimize molecules-guided by rewards for binding affinity (<span class="katex-eq" data-katex-display="false">\Delta G</span>), drug-likeness (QED), synthetic accessibility, and novelty-starting from a target protein and utilizing embeddings from similar proteins identified via ESM-2. — This de novo molecular drug discovery pipeline leverages reinforcement learning with a PPO agent to iteratively optimize molecules-guided by rewards for binding affinity ( $\Delta G$ ), drug-likeness (QED), synthetic accessibility, and novelty-starting from a target protein and utilizing embeddings from similar proteins identified via ESM-2.

Generative Models: Sculpting Molecules from Data

Generative models represent a significant advancement in de novo drug design by shifting from traditional screening methods to the algorithmic creation of novel molecules. These models operate by analyzing large datasets of known chemical structures – typically sourced from databases like ChEMBL and PDBbind – and identifying the statistical relationships between molecular features, such as bond connectivity, atomic composition, and three-dimensional conformation. The learned patterns are then used to generate new molecular structures that statistically resemble the training data, but are not exact copies, thereby expanding the chemical space for potential drug candidates. This approach allows for the design of molecules with desired properties, optimized for specific targets, and potentially bypassing the limitations of existing compound libraries.

Multiple generative model architectures are employed for de novo molecular design, each utilizing distinct approaches to generate novel compounds. AutoEncoders learn compressed representations of molecules, reconstructing them from this reduced dimensionality; Variational AutoEncoders introduce probabilistic encoding, allowing for the generation of molecules by sampling from the latent space. Generative Adversarial Networks (GANs) pit a generator network against a discriminator, iteratively refining the generator’s ability to produce realistic molecular structures. Diffusion Models, conversely, operate by progressively adding noise to a molecule and then learning to reverse this process, enabling the creation of new molecules from random noise; these models have recently demonstrated state-of-the-art performance in molecular generation tasks.

The Transformer architecture, while demonstrating significant capability in generative modeling tasks including molecular design, presents substantial computational demands. These requirements stem from the self-attention mechanism, which scales quadratically with the input sequence length – in this case, the number of atoms in a molecule. This quadratic scaling impacts both training and inference times, necessitating the use of specialized hardware like GPUs or TPUs for practical application. Consequently, research efforts are actively focused on developing alternative strategies, such as simplified attention mechanisms, graph neural networks, or recurrent neural networks, to achieve comparable generative performance with reduced computational cost and improved scalability for larger and more complex molecular structures.

Effective training of generative molecular models is contingent upon access to large, high-quality datasets characterizing known chemical compounds and their properties. Databases such as ChEMBL and PDBbind serve as critical resources, providing curated collections of molecular structures, associated bioactivity data, and physicochemical properties. ChEMBL, a manually curated database, contains over 2.2 million bioactive molecules with associated potency data against over 1,800 targets. PDBbind, conversely, focuses on experimentally determined protein-ligand binding affinities, containing over 19,000 complexes. The size and quality of these datasets directly influence the model’s ability to learn valid chemical patterns and generate novel molecules with desired characteristics; insufficient or biased data can lead to the generation of chemically invalid or unrealistic structures.

ReACT-Drug successfully generated novel, chemically valid, and synthetically accessible high-affinity molecules for six diverse protein targets, exceeding the scope of the MOSES benchmark.

Reinforcement Learning: Navigating Chemical Space with Purpose

Reinforcement Learning (RL) offers a computational framework for molecular optimization by treating the process of molecule generation as a sequential decision-making problem. An “agent” explores the vast chemical space – the set of all possible molecules – by iteratively proposing molecular structures. Each proposed molecule is evaluated based on its properties, and a “reward” signal is provided to the agent, indicating the desirability of the molecule based on predefined criteria, such as binding affinity or solubility. The agent then learns a “policy” – a strategy for selecting actions (molecular modifications) – that maximizes cumulative reward, effectively guiding the search towards molecules exhibiting optimal properties. This approach contrasts with traditional methods by actively learning to navigate chemical space, rather than relying on pre-defined search algorithms or extensive pre-screening of compounds.

ReACT-Drug employs a reinforcement learning (RL) methodology where an agent learns to directly generate molecular structures. This is achieved by defining a policy that maps states – representing partially constructed molecules – to actions – corresponding to the addition of specific chemical building blocks. The agent’s actions are evaluated based on a reward function that quantifies the desirability of the resulting molecule, incorporating characteristics such as predicted binding affinity, drug-likeness, and synthetic accessibility. Through iterative training, the agent optimizes its policy to maximize cumulative rewards, effectively learning to generate molecules with improved properties as defined by the reward function.

ReACT-Drug demonstrated the ability to generate molecules with competitive binding affinities, achieving mean values ranging from -9.13 to -10.4 kcal/mol against six distinct protein targets. Critically, this performance was obtained without any target-specific training data; the model generalized across diverse protein structures based solely on the reward function. This target-agnostic capability suggests ReACT-Drug can facilitate de novo drug design for novel targets where limited prior knowledge exists, reducing the need for extensive target-specific datasets and accelerating the drug discovery process.

MOLRL (Molecular Reinforcement Learning) advances drug design by employing reinforcement learning within a continuous latent space representation of molecules. This approach utilizes an autoencoder to map molecules to a lower-dimensional latent vector, allowing the RL agent to explore and optimize molecular structures more efficiently than direct manipulation of SMILES strings or other discrete representations. The agent learns to navigate this latent space, generating latent vectors that, when decoded, produce molecules predicted to possess improved properties. This method enables the discovery of novel compounds with desired characteristics by circumventing the limitations of traditional discrete search spaces and leveraging the continuous nature of the latent representation for smoother and more effective optimization.

Proximal Policy Optimization (PPO) is a reinforcement learning algorithm designed to improve training stability and sample efficiency. It achieves this by limiting the policy update at each step; specifically, PPO constrains the ratio between the new and old policies to remain within a specified range, typically using a clipped surrogate objective function. This constraint prevents excessively large policy updates that can lead to instability and performance degradation. By ensuring more conservative updates, PPO allows for more reliable learning and reduces the variance in the training process, leading to faster convergence and improved overall performance in complex chemical space exploration tasks like molecular optimization.

Molecular Representation and Synthesis: Building with Defined Rules

Accurate molecular representation is foundational for both generative modeling and reinforcement learning approaches in chemical space. The performance of these algorithms is directly correlated with the quality of the molecular encoding. ChemBERTa, a BERT-based language model trained on a large corpus of SMILES strings, provides a powerful and context-aware embedding of molecular structures. This embedding captures complex chemical information, enabling more effective prediction of molecular properties and facilitating the generation of valid and diverse chemical compounds. Unlike simpler representations, ChemBERTa’s SMILES-based embedding accounts for the sequential nature of SMILES notation, providing a richer and more nuanced understanding of molecular structure and relationships.

Reaction template libraries are constructed from databases such as ChEMBL, providing a curated collection of known and experimentally verified chemical transformations. These libraries define allowable reactions by mapping reactant substructures to product substructures, effectively encoding synthetic feasibility rules. Each template specifies the chemical changes that are considered valid, including atom mapping, bond changes, and any required reagents or conditions. Utilizing these pre-defined templates during de novo molecular generation constrains the search space to synthetically accessible compounds, increasing the probability of producing molecules that can be realistically synthesized in a laboratory setting. This approach differs from unconstrained generation methods which may propose chemically impossible structures.

Fragmenting molecules into meaningful building blocks is essential for retrosynthetic analysis and the subsequent creation of reaction templates. Algorithms such as BRICS (Building blocks for Reaction Information Classification System) and RECAP (Recursive Enumeration of Chemically Accessible Pathways) facilitate this process by systematically breaking down a molecule into smaller, commercially available fragments. BRICS operates by identifying potential disconnection points based on bond types and atom environments, while RECAP employs a recursive approach to enumerate all possible fragmentations. The resulting fragments, paired with the identified disconnection reactions, form the basis of reaction templates used in generative models to ensure the proposed synthetic steps are chemically plausible and align with established transformations.

Generated molecules, as evaluated by established metrics, demonstrate an average Quantitative Estimate of Drug-likeness (QED) score of 0.307. This score assesses properties associated with potential pharmaceutical candidates, with higher values indicating increased drug-likeness. Simultaneously, the molecules achieve an average Synthetic Accessibility (SA) score of 3.15, as determined by an eight-rule-based assessment. This SA score indicates the relative ease of synthesizing the generated compounds, with lower numbers representing higher synthetic accessibility; a score of 3.15 suggests moderate synthetic feasibility. These metrics provide quantitative benchmarks for evaluating the quality and practicality of the generated molecular structures.

Policy-guided Unbiased REpresentations utilize reinforcement learning to generate novel molecules with specific structural properties. This approach trains an agent to iteratively construct molecular graphs, guided by a policy network that assigns probabilities to valid bond formations. The reward function incorporates both a novelty component, encouraging exploration of chemical space, and a constraint component, directing generation towards desired structural features. By learning to navigate the chemical space through iterative refinement, the model achieves structure-constrained molecular generation without being limited by predefined molecular scaffolds or templates, leading to the creation of compounds with tailored properties.

Generated molecules exhibit adaptable physicochemical properties, maintaining consistent molecular weight <span class="katex-eq" data-katex-display="false"> \approx 0.8 </span> normalized units while tailoring hydrogen bond acceptor/donor counts to meet the specific demands of diverse targets like 5-HT1B and DRD2. — Generated molecules exhibit adaptable physicochemical properties, maintaining consistent molecular weight $\approx 0.8$ normalized units while tailoring hydrogen bond acceptor/donor counts to meet the specific demands of diverse targets like 5-HT1B and DRD2.

The Future of AI-Driven Drug Discovery: A New Paradigm

The future of pharmaceutical innovation is increasingly shaped by a powerful synergy between artificial intelligence techniques. Generative models, capable of creating novel molecular structures, are now coupled with reinforcement learning algorithms that optimize these structures for desired properties, such as binding affinity and drug-likeness. Crucially, this process relies on advanced molecular representations – sophisticated ways of describing molecules that allow AI to ‘understand’ their characteristics and predict their behavior. This convergence dramatically accelerates drug discovery by intelligently navigating the vast chemical space, identifying promising candidates far more efficiently than traditional methods, and ultimately reducing both the time and expense associated with bringing new treatments to market.

Artificial intelligence is revolutionizing drug discovery by efficiently exploring the vast landscape of potential molecules – a concept known as chemical space. Traditionally, identifying promising drug candidates involved painstakingly synthesizing and testing countless compounds, a process that is both time-consuming and expensive. However, AI algorithms, particularly those employing generative models and reinforcement learning, can intelligently navigate this space, predicting molecular properties and prioritizing compounds likely to exhibit desired therapeutic effects. This computational approach not only accelerates the identification of novel candidates but also optimizes their characteristics – enhancing potency, selectivity, and bioavailability – while simultaneously reducing the overall costs associated with preclinical development. By focusing resources on the most promising molecules, AI promises to dramatically shorten the drug development pipeline and deliver innovative treatments to patients more quickly.

Recent studies showcase the tangible impact of artificial intelligence on pharmaceutical innovation, notably through the ReACT-Drug model. This system achieved a predicted binding affinity of -11.3 kcal/mol when targeting the kappa opioid receptor (KOR), a significant result indicating strong potential for drug interaction. Importantly, ReACT-Drug’s performance extended beyond this single target; it demonstrably outperformed established dopamine receptor D2 (DRD2) inhibitors, achieving a mean binding affinity of -10.7 kcal/mol compared to their average of -7.753 kcal/mol. These findings suggest that AI-driven approaches are not merely theoretical exercises but can generate novel molecular candidates with superior binding characteristics, offering a promising pathway toward the development of more effective therapeutic interventions.

The trajectory of artificial intelligence in drug discovery is inextricably linked to ongoing progress in several key areas. Improvements in algorithmic design, particularly within generative models and reinforcement learning, are enabling more efficient exploration of vast chemical spaces and more accurate prediction of molecular properties. Simultaneously, exponential growth in computational power-driven by advancements in hardware and cloud computing-allows for the training of increasingly complex models and the simulation of molecular interactions with unprecedented detail. Crucially, the availability of large, high-quality datasets-including genomic information, protein structures, and clinical trial results-provides the necessary fuel for these algorithms to learn and refine their predictive capabilities. As these three forces – smarter algorithms, greater computing resources, and more comprehensive data – continue to converge, the potential for AI to revolutionize drug discovery, leading to faster development of more effective treatments, will be increasingly realized.

The ultimate ambition driving artificial intelligence in drug discovery extends beyond incremental improvements to existing methods; it envisions a paradigm shift in how therapeutics are identified and brought to patients. By dramatically compressing the timeline and reducing the expense traditionally associated with bringing a novel drug to market, AI promises to address unmet medical needs with greater agility. This isn’t merely about faster screening of known compounds, but the de novo design of molecules tailored to specific biological targets, potentially circumventing limitations of current treatment options. The efficiency gains aren’t limited to the initial discovery phase, as AI can also optimize clinical trial design, predict patient responses, and personalize treatment regimens – ultimately fostering a healthcare landscape where effective therapies reach those who need them, with unprecedented speed and precision.

The pursuit of novel drug candidates, as detailed in this work concerning ReACT-Drug, often falls prey to unnecessary complexity. The framework’s reliance on reaction templates and protein embeddings, while sophisticated, ultimately serves to distill the generative process into manageable components. This echoes Bertrand Russell’s sentiment: “The point of education is not to increase the amount of information, but to create the capacity for critical thinking.” ReACT-Drug, similarly, isn’t merely about generating molecules; it’s about constructing a system capable of intelligently navigating chemical space, prioritizing clarity and efficiency in the search for viable therapeutic options. The target-agnostic approach further embodies this principle, fostering a generalized capacity for drug discovery rather than focusing on isolated instances.

What Remains?

The pursuit of de novo molecular generation, even when guided by reinforcement learning and tempered by reaction templates, inevitably reveals the limitations of current understanding. ReACT-Drug demonstrates a capacity for producing viable candidates, yet viability, as a metric, feels suspiciously close to luck. The framework’s target-agnostic approach, while elegant in its generality, skirts the deeper issue: true drug design isn’t about creating molecules that simply bind; it’s about predicting complex biological interactions with a precision currently beyond reach.

Future iterations must confront the inherent ambiguity of ‘drug-likeness’. Current proxies, however sophisticated, are merely statistical echoes of past successes, not guarantees of future efficacy. The true advancement lies not in generating more molecules, but in developing a more fundamental understanding of the protein-ligand landscape – a landscape currently mapped by approximations.

Simplicity, then, is the critical path. The field would benefit from abandoning the quest for ever-more-complex models and instead focusing on distilling existing knowledge into a minimal, predictive core. If a molecular property cannot be explained in one sentence, it is likely not understood, and any framework built upon it will remain, at best, a sophisticated exercise in pattern recognition.

Original article: https://arxiv.org/pdf/2512.20958.pdf

Contact the author: https://www.linkedin.com/in/avetisyan/