Sampling useful three-dimensional molecular structures along with their most favorable conformations is a key challenge in drug discovery. Current state-of-the-art 3D de-novo design flow matching or diffusion-based models are limited to generating a single conformation. However, the conformational landscape of a molecule determines its observable properties and how tightly it is able to bind to a given protein target. By generating a representative set of low-energy conformers, we can more directly assess these properties and potentially improve the ability to generate molecules with desired thermodynamic observables. Towards this aim, we propose \textit{FlexiFlow}, a novel architecture that extends flow-matching models, allowing for the joint sampling of molecules along with multiple conformations while preserving both equivariance and permutation invariance. We demonstrate the effectiveness of our approach on the QM9 and GEOM Drugs datasets, achieving state-of-the-art results in molecular generation tasks. Our results show that FlexiFlow can generate valid, unstrained, unique, and novel molecules with high fidelity to the training data distribution, while also capturing the conformational diversity of molecules. Moreover, we show that our model can generate conformational ensembles that provide similar coverage to state-of-the-art physics-based methods at a fraction of the inference time. Finally, FlexiFlow can be successfully transferred to the protein-conditioned ligand generation task, even when the dataset contains only static pockets without accompanying conformations.

Predicting which products can be made from a set of building blocks underpins library design, analogue generation, and synthesis-constrained generative modelling. Reaction templates encoded as SMARTS patterns are the standard tool for this forward synthesis enumeration task. However, existing approaches force a trade-off between chemical coverage, specificity, and the computational cost of applying large template sets at scale. We present SmartReact, an open-source Python library that resolves this trade-off through a template set built on SMARTS-RX, a hierarchical functional group ontology. The same ontology that gives each template a precise chemical environment also drives exact pre-filtering of incompatible reactant pairs, keeping enumeration fast as the template library grows. Every prediction is traceable to a named reaction template and an explicit set of functional group requirements, making results directly interpretable by chemists and straightforward to validate or audit. New reaction types can be incorporated by specifying a SMARTS pattern and its functional group requirements, making the library easy to extend. A case study enumerating 524,911 unique products in under a minute from 1,000 building blocks demonstrates real-world utility at scale. Scientific Contribution: SmartReact contributes both a large, expert-refined SMARTS reaction template library and a scalable open-source Python package designed to use it efficiently. The library’s 9,594 templates spanning 36 medicinal chemistry reaction types are grounded in the SMARTS-RX functional group ontology, giving them high chemical specificity and non-redundancy that auto-extracted libraries lack. The same ontology drives exact pre-filtering of incompatible reactant-template combinations, yielding a 43-fold algorithmic speedup that combines with near-linear parallel scaling and a classify-once API to reach over 11,000-fold throughput relative to single-core brute-force matching on a 32-core server.

Nickel/photoredox catalysis in cross-coupling reactions offers mild operating conditions for efficient C-C bond formation, expanding synthetic access to pharmaceutically relevant molecules. However, routine implementation of such reactions remains constrained by the intricate reaction mechanism and limited availability of experimental data, which complicate optimization tasks and the development of predictive models. The integration of quantum-mechanical (QM) calculations with machine learning (ML) has proven to be effective for developing predictive models of complex reactions with sparse experimental data. Here, we present a combined approach that integrates automated density functional theory calculations, ML, and parallel synthesis to develop quantum mechanics-machine learning (QM-ML) models for the nickel metallophotoredox cross-coupling reaction feasibility prediction. Random-Forest classification models are trained to predict the outcome of a given reaction using DFT-computed descriptors from automatically generated 3D structures of catalytic cycle intermediates. We demonstrate the broad applicability of this approach, applying it to a diverse data set encompassing four reaction subtypes, namely, bromide cross-electrophile couplings, chloride cross-electrophile couplings, deoxygenative couplings, and amino radical transfer (ART) couplings, augmented with additional experiments curated by a systematic cheminformatics method to broaden the alkyl halides scope. We show on a blind literature data set that such a QM-ML approach can successfully predict the feasibility of complex reactions from heterogeneous data sets with minimal data requirements and can generalize it to unseen reaction subtypes with a few-shot learning approach, affording a computational model for ART coupling. Together, these capabilities provide a data-efficient solution for rapidly predicting the outcome of cross-coupling reactions and facilitate the adoption of nickel photocatalysis in the MAKE stage of the DMTA cycle.

Abstract Machine learning (ML) models for molecular property prediction are increasingly deployed in drug discovery, yet their adoption in real-world scenarios requires an understanding of the conditions in which a model succeeds or fails. While standardized benchmarks are powerful instruments to measure and unlock progress in ML research, they should not be blindly treated as the end goal. Especially static and retrospective benchmarks, in which no true unknown test set is employed, limit our ability to robustly validate a model’s performance. Building on the collective expertise of a cross-industry consortium, we present a model validation framework consisting of five recommendations that would enable the community to move beyond aggregate metrics toward understanding where and why molecular property prediction models fail. We connect evaluation choices to real-world applications and case studies encountered in pharmaceutical research. The framework proposes splitting strategies that mimic realistic distribution shifts and expose common failure modes. We apply the recommended framework on a recently released dataset of absorption, distribution, metabolism, and excretion (ADME) properties. Across two complementary model algorithms, our case studies reveal four distinct failure modes (extrapolation, interpolation, representation, and evaluation) showing that model errors arise not only from distribution shift but also from limitations in molecular representations. Our results show that commonly used evaluation protocols can significantly overestimate performance and may not detect important model failure modes. All software and data are released via https://github.com/srijitseal/polaris .

Abstract Generative models based on diffusion and flow matching have recently been applied to structure-based drug design, but their outputs often include unrealistic protein–ligand interactions that do not obey the laws of physics. We present an energy guidance framework that incorporates a molecular mechanics force field (MMFF94) directly into the sampling process. The method steers molecular generation toward more physically plausible and energetically stable conformations without retraining the underlying model. We evaluate this approach using two state-of-the-art architectures, SemlaFlow, a flow matching model and EDM, a diffusion model, on the PDBBind dataset. Across both models, energy guidance improves enthalpic interaction energy, improves strain energy by up to 75 $$\%$$ % , and generates over 1000 ligands with better docking scores than native ligands. These results demonstrate that lightweight, physics-based guidance can significantly enhance generative drug design while preserving chemical validity and diversity. Scientific contribution We introduce a novel, training-free force field guidance framework that steers ligand generation using empirical molecular mechanics (e.g., MMFF94) during diffusion or flow-based sampling–without modifying or retraining the base generative model (e.g., EDM or Semflaflow by [24]). Our method operates as a plug-in during inference time, leveraging energy feedback to generate poses with lower strain and having better predicted interactions with the protein structure. Our main contributions are as follows: Energy-based guidance without retraining: Unlike methods that require gradients from neural affinity predictors (e.g., BADGER [26]), our approach injects classical force field feedback (MMFF94) directly during the posterior sampling step. Improved docking and strain metrics: In benchmarks against unconditional EDM and Semflaflow, our guided inference yields consistently better AutoDock Vina scores and lower ligand strain energy, even after optimizing the final structures using the same force field. Compatibility and flexibility: Because the guidance module is external, it can be applied broadly to multiple generative backbones–without retraining or architecture modifications, and can be applied to arbitrary differentiable potential energy functions. Theoretical guarantee of stability. We demonstrate in Appendix B that the gradient correction step corresponds to a descent step on the energy under standard smoothness assumptions. While the full sampling update also includes model-driven (and, in the diffusion case, stochastic) components, this result formalizes how the guidance term locally biases the trajectory toward lower-energy regions and provides a principled justification for its stabilizing effect.

A higher-level retrosynthesis strategy helps computers search synthesis space more like chemists do, improving route discovery for challenging molecular targets.

Agentic systems that are based on large language models (LLMs) have emerged as promising tools in the chemistry domain over the past few years. Early examples included work on CoScientist, Chemcrow, and LLM-RDF, which showcased the potential of agentic systems to assist in chemical research, in the orchestration of cheminformatics tools, and in synthetic reaction development. Despite this, the current literature lacks examples of the real-world adoption of such systems in drug discovery. We present such an example by describing our work on an agentic system called ChatInvent, which has been integrated into the discovery pipeline at AstraZeneca to aid in molecular design and synthesis planning. We discuss how the system evolved from a proof-of-concept single agent into an extensible, robust, and scalable multi-agent architecture with a graphical user interface. We emphasize the lessons learnt and the challenges that persist as we continue to work on this project, and share our perspectives on the future of agentic systems in our domain.

Generative models coupled with reinforcement learning (RL), such as REINVENT and PepINVENT, have emerged as a powerful framework for de novo molecular design. During the ideation process these generative frameworks utilize various predictive models as part of the optimization objectives. However, the utility of the predictive models can be limited by their domain of applicability. When RL is used to explore the chemical space with predictive models, it can suggest molecules that lie outside the predictor's domain of applicability. As a result, the predictions may become less reliable, potentially steering designs into high reward but also high uncertainty chemical spaces. This is particularly pronounced for cyclic peptides which show therapeutic promise due to their modifiability and large interaction surfaces but are understudied compared to small molecules. While passive membrane permeation in cyclic peptides has attracted interest, identifying optimal permeable designs remains challenging yet crucial for targeting intracellular sites. We present an RL-guided generative framework that designs permeable cyclic peptides using an uncertainty-aware permeability predictor as the scoring component. To address predictive uncertainty, especially impacted by novel chemistry, we integrate conformal prediction (CP) as our uncertainty quantification method. CP assesses designs based on the calibrated model under a user-defined confidence level. We demonstrate that rewarding generated peptides with CP-informed predictions improves both reliability and efficiency of peptide optimization process. This also discourages exploration outside the predictor's applicability domain. This approach bridges the gap between predictive uncertainty and RL-guided exploration, showing how generative modelling and conformal prediction can be combined for the first time.

Large-language models (LLMs) incorporated into agentic systems present exciting opportunities to accelerate drug discovery. In this study, we examine the modularity of LLM-based agentic systems for drug discovery, i.e., whether parts of the system are interchangeable, a topic that has received limited attention in drug discovery. We compare the performance of different LLMs and the effectiveness of tool-calling agents versus code-generating agents. Our case study, comparing performance in orchestrating tools for chemistry and drug discovery using an LLM-as-a-judge score, shows that Claude-3.5-Sonnet, Claude-3.7-Sonnet and GPT-4o outperform alternative LLMs such as Llama-3.1-8B, Llama-3.1-70B, GPT-3.5-Turbo, and Nova-Micro. Although we confirm that code-generating agents outperform the tool-calling ones on average, we show that this is highly question- and model-dependent. Furthermore, the impact of replacing the system promptl is dependent on the question and model, underscoring that even in this particular domain one cannot replace components of the system without re-engineering. Our study highlights the necessity of further research into the modularity of LLM-based agentic systems to enable the development of reliable and modular solutions for real-world problems.

Selecting efficient multi-step synthetic routes is a central challenge in organic synthesis, particularly in medicinal and process chemistry, where route choice directly impacts feasibility, cost, and development efficiency. Data-driven assessment systems often oversimplify the multi-objective nature of synthesis design and rely on proxy datasets, such as patent routes, rather than universally grounded criteria. To address this, we introduce an expert-augmented, data-driven scoring framework that integrates machine learning with chemists' domain knowledge for both numerical and explainable route assessment. A DeepSets-based model is trained using tree edit distance between reference and machine-generated routes, and then fine-tuned with expert evaluations to produce both quantitative scores and interpretable qualitative categories: Good, Plausible, and Bad. The resulting system achieves a Spearman correlation coefficient of 0.78 and a Pearson correlation of 0.77 for category assessment prediction, and 60.2

Selecting efficient multi-step synthetic routes is a central challenge in organic synthesis, particularly in medicinal and process chemistry, where route choice directly impacts feasibility, cost, and development efficiency. While computer-aided synthesis planning (CASP) tools can generate diverse pathways, the evaluation of these generated routes remains challenging. Existing heuristic or data-driven scoring systems often oversimplify the multi-objective nature of synthesis design and rely on proxy datasets (e.g., patent routes) rather than universal grounded criteria. To address this, we introduce an expert-augmented, data-driven scoring framework that integrates machine learning with chemists’ domain knowledge for both numerical and explainable route assessment. A DeepSets-based model is trained using tree edit distance between reference and machine-generated routes, then fine-tuned with expert evaluations to produce both quantitative scores and interpretable qualitative categories (Good, Plausible, Bad). Validated against expert assessments, the resulting system achieves a Spearman correlation coefficient of 0.78 ± 0.05 and a Pearson correlation of 0.77 ± 0.06 regarding its category assessment prediction. When ranking routes by predicted score, the model achieves an overall 60.2% top-1 accuracy (vs. a 17.5% baseline) relative to reference routes, while expert validation confirms that 75% of sampled top-1 selections are feasible, demonstrating the model correctly identifies diverse, expert-validated alternatives beyond the reference.

Free energy calculations are widely used tools in computational chemistry, but their dependence on the assignment of partial charges during force field parametrization reduces their accuracy and reproducibility. In this work, we highlight the direct connection between the low accuracy of AM1-BCC charges on polar species and the poor accuracy of corresponding hydration free energy calculations. We then propose an XGBoost regressor trained on atomic descriptors to rapidly predict charges obtained with high-fidelity density functional theory calculations at PBE0-D3(BJ)/def2-TZVP level. The more accurate electrostatic description results in more reliable free energy calculations than those obtained with semiempirical AM1-BCC charges. Finally, we leverage this predictive model in combination with a 1 ns gas-phase molecular dynamics simulation to propose the Boltzmann Percentile method for assigning charges representative of the conformational ensemble of a molecule. Charges obtained with this method are robust to different input conformations, and the resulting free energies, calculated on a subset of the FreeSolv data set, show a root mean squared error of 1.69 kcal/mol against the 3.05 kcal/mol obtained with semiempirical charges as well as a significantly better ranking. We provide a trained model, easily integrable into traditional workflows, enabling free energy calculations at the same computational cost as AM1-BCC charges. These two aspects make it a realistic tool for enhancing already expensive free energy calculations, and more in general, molecular dynamics simulations in condensed phase.

The estimation of uncertainties associated with predictions from quantitative structure–activity relationship (QSAR) models can accelerate the drug discovery process by identifying promising experiments and allowing an efficient allocation of resources. Several computational tools exist that estimate the predictive uncertainty in machine learning models. However, deviations from the i.i.d. setting have been shown to impair the performance of these uncertainty quantification methods. We use a real-world pharmaceutical dataset to address the pressing need for a comprehensive, large-scale evaluation of uncertainty quantification approaches in the context of realistic distribution shifts over time. We investigate the performance of several popular uncertainty estimation methods for classification models, including ensemble-based and Bayesian approaches. Furthermore, we use this real-world setting to systematically assess the distribution shifts in label and descriptor space and their impact on the capability of the uncertainty quantification methods. Our study reveals significant shifts over time in both label and descriptor space and a clear connection between the magnitude of the shift and the nature of the assay. Moreover, we show that pronounced distribution shifts impair the performance of popular uncertainty quantification methods used in QSAR models. This work highlights the challenges of identifying uncertainty quantification techniques that remain reliable under distribution shifts introduced by real-world data.

In the drug discovery process, where experiments can be costly and time-consuming, computational models that predict drug-target interactions are valuable tools to accelerate the development of new therapeutic agents. Estimating the uncertainty inherent in these neural network predictions provides valuable information that facilitates optimal decision-making when risk assessment is crucial. However, such models can be poorly calibrated, which results in unreliable uncertainty estimates that do not reflect the true predictive uncertainty. In this study, we compare different metrics, including accuracy and calibration scores, used for model hyperparameter tuning to investigate which model selection strategy achieves well-calibrated models. Furthermore, we propose to use a computationally efficient Bayesian uncertainty estimation method named HMC Bayesian Last Layer (HBLL), which generates Hamiltonian Monte Carlo (HMC) trajectories to obtain samples for the parameters of a Bayesian logistic regression fitted to the hidden layer of the baseline neural network. We report that this approach improves model calibration and achieves the performance of common uncertainty quantification methods by combining the benefits of uncertainty estimation and probability calibration methods. Finally, we show that combining post hoc calibration method with well-performing uncertainty quantification approaches can boost model accuracy and calibration.

Assay interference caused by small organic compounds continues to pose formidable challenges to early drug discovery. Various computational methods have been developed to identify compounds likely to cause assay interference. However, due to the scarcity of data available for model development, the predictive accuracy and applicability of these approaches are limited. In this work, we present E-GuARD (Expert-Guided Augmentation for Robust Interference Compound Detection), a novel framework seeking to address data scarcity and imbalance by integrating self-distillation, active learning, and expert-guided molecular generation. E-GuARD iteratively enriches the training data with interference-relevant molecules, resulting in quantitative structure-interference relationship (QSIR) models with superior performance. We demonstrate the utility of E-GuARD with the examples of four high-quality data sets on thiol reactivity, redox reactivity, nanoluciferase inhibition, and firefly luciferase inhibition. Our models reached MCC values of up to 0.47 for these data sets, with two-fold or higher improvements in enrichment factors compared to models trained without E-GuARD data augmentation. These results highlight the potential of E-GuARD as a scalable solution to mitigating assay interference in early drug discovery.

In the early stages of drug discovery, decisions regarding which experiments to pursue can be influenced by computational models for quantitative structure–activity relationships (QSAR). These decisions are critical due to the time-consuming and expensive nature of the experiments. Therefore, it is becoming essential to accurately quantify the uncertainty in machine learning predictions, such that resources can be used optimally and trust in the models improves. While computational methods for QSAR modeling often suffer from limited data and sparse experimental observations, additional information can exist in the form of censored labels that provide thresholds rather than precise values of observations. However, the standard approaches that quantify uncertainty in machine learning cannot fully utilize censored labels. In this work, we adapt ensemble-based, Bayesian, and Gaussian models with tools to learn from censored labels by using the Tobit model from survival analysis. Our results demonstrate that despite the partial information available in censored labels, they are essential to reliably estimate uncertainties in real pharmaceutical settings where approximately one-third or more of experimental labels are censored.

Accurate prediction of thermodynamic properties is essential in drug discovery and materials science. Molecular dynamics (MD) simulations provide a principled approach to this task, yet they typically rely on prohibitively long sequential simulations. Implicit Transfer Operator (ITO) Learning offers a promising approach to address this limitation by enabling stable simulation with time steps orders of magnitude larger than MD. However, to train ITOs, we need extensive, unbiased MD data, limiting the scope of this framework. Here, we introduce Boltzmann Priors for ITO (BoPITO) to enhance ITO learning in two ways. First, BoPITO enables more efficient data generation, and second, it embeds inductive biases for long-term dynamical behavior, simultaneously improving sample efficiency by one order of magnitude and guaranteeing asymptotically unbiased equilibrium statistics. Furthermore, we showcase the use of BoPITO in a new tunable sampling protocol interpolating between ITOs trained on off-equilibrium simulations and an equilibrium model by incorporating unbiased correlation functions. Code is available at https://github.com/olsson-group/bopito.

Fine-tuning a pre-trained generative model has demonstrated good performance in generating promising drug molecules. The fine-tuning task is often formulated as a reinforcement learning problem, where previous methods efficiently learn to optimize a reward function to generate potential drug molecules. Nevertheless, in the absence of an adaptive update mechanism for the reward function, the optimization process can become stuck in local optima. The efficacy of the optimal molecule in a local optimization may not translate to usefulness in the subsequent drug optimization process or as a potential standalone clinical candidate. Therefore, it is important to generate a diverse set of promising molecules. Prior work has modified the reward function by penalizing structurally similar molecules, primarily focusing on finding molecules with higher rewards. To date, no study has comprehensively examined how different adaptive update mechanisms for the reward function influence the diversity of generated molecules. In this work, we investigate a wide range of intrinsic motivation methods and strategies to penalize the extrinsic reward, and how they affect the diversity of the set of generated molecules. Our experiments reveal that combining structure- and prediction-based methods generally yields better results in terms of diversity.

Incorporating non-natural amino acids (NNAAs) into peptides enhances therapeutic properties, including binding affinity, metabolic stability, and in vivo half-life time. The pursuit of novel NNAAs for improved peptide designs faces the challenge of effective synthesis of these building blocks as well as the entire peptide itself. Solid-Phase Peptide Synthesis (SPPS) is an essential technology for the automated assembly of peptides with NNAAs, necessitating careful protection for effective coupling of amino acids in the peptide chain. This process requires orthogonal protection of the reactive groups in individual amino acids after synthesizing them, presenting a challenge in bridging in silico peptide design with chemical synthesis. To address this, we have developed a first-of-its-kind synthesis assistance tool, NNAA-Synth, that plans and evaluates the synthesis of individual SPPS-compatible NNAAs. Our tool unifies (i) introducing orthogonal protecting groups to NNAAs, (ii) retrosynthetic prediction to propose synthesis routes, and (iii) scoring the synthetic feasibility of these routes. We demonstrate how the tool facilitates optimal protection strategy selection for individual NNAAs. Additionally, it enables synthesizability-aware NNAA ranking and prioritization during computational screening, enhancing the quality of the in silico design by assessing the accessibility of individual building blocks.