Large language models (LLMs) are increasingly used for molecular tasks, but it remains unclear which molecular representation to use. We present a systematic benchmark evaluating LLM molecular competence across nine representations and eight chemical tasks. We benchmark 16 LLMs across five model families, including reasoning and non-reasoning variants, chemistry-specialized LLMs, and closed frontier models. Performance is strongly representation-dependent and no single representation wins across tasks, though CML is the best, followed by MolJSON, InChI, and then canonical SMILES. Explicit structured text representations (CML and MolJSON) dominate structural tasks; IUPAC dominates semantic tasks, winning molecule retrieval for all 16 LLMs; and SMILES variants are rarely optimal despite their prevalence in pretraining. Chemistry-specialized models perform well with SMILES at the cost of large degradations with structured text representations, suggesting SMILES-only evaluation rewards specialization that does not generalize. Using LLM-as-a-judge, we find that IUPAC produces the highest fraction of correct molecule generations. A mechanistic study via tokenization audits, linear probes and attention shows that representations are encoded differently inside the model; for example, structured representations require higher attention across the molecular span. Our results argue against representation-invariant evaluation and motivate task-aware representation routing for LLM-based chemistry.

Molecular Graph Neural Networks (GNNs) are increasingly common in drug discovery, particularly for Quantitative Structure-Activity Relationship (QSAR) studies; yet, their superiority compared to classical molecular featurisation approaches is disputed. We report a general strategy for improving GNNs for QSAR by pre-training to predict Extended-Connectivity Fingerprints (ECFP). We validate our approach with statistical tests and challenging out-of-distribution (OOD) splits. Across five out of six Biogen benchmarks, we observed a statistically significant improvement in standard performance metrics over all evaluated baselines when using ECFP pre-trained GNNs. However, for more heterogeneous datasets and more complex endpoints, such as binding affinity prediction, pre-trained GNNs underperformed in OOD settings. Importantly, we investigated the impact of substructure-level data leakage during pre-training on downstream performance. While we identified scenarios where pre-training on ECFPs was less effective, our findings show that ECFP-based pre-training can enhance downstream OOD performance on a diverse set of practically relevant QSAR tasks.

Graph neural networks (GNNs) are a natural choice to represent chemical data, due to their inherent ability to handle arbitrary input topologies. They avoid the need to convert molecules into molecular fingerprints with a fixed vector length. However, like most deep learning models, GNNs are not interpretable and common explainability methods fail because of the variable input size. We introduce a novel method to interpret the predictions of GNNs based on Myerson values from cooperative game theory. Myerson values are closely related to Shapley values , which have been adapted to explain a wide variety of machine learning model predictions. Applying these approaches to GNNs have, however, proven to be challenging because of their varying graph size. Our approach treats a GNN as a coalition game and the nodes of an input layer graph as players. The Myerson value of a node then determines the contribution to the prediction of the model, with only connected nodes contributing to coalitions. All Myerson values add up to the predicted value of the model allowing for a simple and intuitive interpretation of the prediction. Because calculating Myerson values becomes computationally infeasible for large graphs, we have also implemented a scalable approximation technique using Monte Carlo sampling. We developed the technique for applications in cheminformatics and drug discovery, but it can also be used in any application that uses GNNs. The effectiveness of our approach is validated through successful applications to two proof-of-concept datasets (logP and molecular weight) as well as a real-world dataset featuring kinase inhibitors, highlighting its broad applicability and promise in explaining graph-based cheminformatic models.

Determining the binding pose of a ligand to a protein, known as molecular docking, is a fundamental task in drug discovery. Generative approaches promise faster, improved, and more diverse pose sampling than physics-based methods, but are often hindered by chemically implausible outputs, poor generalisability, and high computational cost. To address these challenges, we introduce a novel fragmentation scheme, leveraging inductive biases from structural chemistry, to decompose ligands into rigid-body fragments. Building on this decomposition, we present SigmaDock, an SE(3) Riemannian diffusion model that generates poses by learning to reassemble these rigid bodies within the binding pocket. By operating at the level of fragments in SE(3), SigmaDock exploits well-established geometric priors while avoiding overly complex diffusion processes and unstable training dynamics. Experimentally, we show SigmaDock achieves state-of-the-art performance, reaching Top-1 success rates (RMSD<2 & PB-valid) above 79.9% on the PoseBusters set, compared to 12.7-30.8% reported by recent deep learning approaches, whilst demonstrating consistent generalisation to unseen proteins. SigmaDock is the first deep learning approach to surpass classical physics-based docking under the PB train-test split, marking a significant leap forward in the reliability and feasibility of deep learning for molecular modelling.

Machine learning offers great promise for fast and accurate binding affinity predictions. However, current models lack robust evaluation and fail on tasks encountered in (hit-to-) lead optimisation, such as ranking the binding affinity of a congeneric series of ligands, thereby limiting their application in drug discovery. Here, we address these issues by first introducing a novel attention-based graph neural network model called AEV-PLIG (atomic environment vector-protein ligand interaction graph). Second, we introduce a new and more realistic out-of-distribution test set called the OOD Test. We benchmark our model on this set, CASF-2016, and a test set used for free energy perturbation (FEP) calculations, that not only highlights the competitive performance of AEV-PLIG, but provides a realistic assessment of machine learning models with rigorous physics-based approaches. Moreover, we demonstrate how leveraging augmented data (generated using template-based modelling or molecular docking) can significantly improve binding affinity prediction correlation and ranking on the FEP benchmark (weighted mean PCC and Kendall's tau increases from 0.41 and 0.26 to 0.59 and 0.42). These strategies together are closing the performance gap with FEP calculations (FEP+ achieves weighted mean PCC and Kendall's tau of 0.68 and 0.49 on the FEP benchmark) while being similar to 400,000 times faster.

Unconditional molecular generation is a stepping stone for conditional molecular generation, which is important in de novo drug design. Recent unconditional 3D molecular generation methods report saturated benchmarks, suggesting it is time to re-evaluate our benchmarks and compare the latest models. We assess five recent high-performing 3D molecular generation methods (EQGAT-diff, FlowMol, GCDM, GeoLDM, and SemlaFlow), in terms of both standard benchmarks and chemical and physical validity. Overall, the best method, SemlaFlow, has a success rate of 87 molecules without post-processing and 92.4

Machine learning offers a promising approach for fast and accurate binding affin- ity predictions. However, current models often fail to generalise beyond their training data and are not robustly evaluated on a diverse range of benchmarks, limiting their application in drug discovery projects. In this work, we address these issues by intro- ducing a novel graph neural network model called AEV-PLIG (Atomic Environment Vector - Protein Ligand Interaction Graph), which encodes protein-ligand interactions via atomic environment vectors to improve generalisation. We evaluate our model on improved benchmarks, including our new out-of-distribution test set we call OOD Test, and two alternative benchmark systems used for free energy perturbation (FEP) calculations, and highlight competitive performance of AEV-PLIG across the board. Moreover, we demonstrate how augmented data can be leveraged to enhance predic- tion accuracy, and how enriching the training data with three complexes from a con- generic series of ligands binding to a target of interest improves performance further. Altogether, we show that these strategies improve the applicability of machine learn- ing scoring functions and enable state-of-the-art performance nearing the accuracy of physics-based simulation methods—but at a fraction of their computational cost. This practical approach extends the predictive capabilities of machine learning for molecular discovery, paving the way for its broader use in computer-aided drug design.

A novel class of protein misfolding characterized by either the formation of non-native noncovalent lasso entanglements in the misfolded structure or loss of native entanglements has been predicted to exist and found circumstantial support through biochemical assays and limited-proteolysis mass spectrometry data. Here, we examine whether it is possible to design small molecule compounds that can bind to specific folding intermediates and thereby avoid these misfolded states in computer simulations under idealized conditions (perfect drug-binding specificity, zero promiscuity, and a smooth energy landscape). Studying two proteins, type III chloramphenicol acetyltransferase (CAT-III) and D-alanyl-D-alanine ligase B (DDLB), that were previously suggested to form soluble misfolded states through a mechanism involving a failure-to-form of native entanglements, we explore two different drug design strategies using coarse-grained structure-based models. The first strategy, in which the native entanglement is stabilized by drug binding, failed to decrease misfolding because it formed an alternative entanglement at a nearby region. The second strategy, in which a small molecule was designed to bind to a non-native tertiary structure and thereby destabilize the native entanglement, succeeded in decreasing misfolding and increasing the native state population. This strategy worked because destabilizing the entanglement loop provided more time for the threading segment to position itself correctly to be wrapped by the loop to form the native entanglement. Further, we computationally identified several FDA-approved drugs with the potential to bind these intermediate states and rescue misfolding in these proteins. This study suggests it is possible for small molecule drugs to prevent protein misfolding of this type.

Generative models have the potential to accelerate key steps in the discovery of novel molecular therapeutics and materials. Diffusion models have recently emerged as a powerful approach, excelling at unconditional sample generation and, with data-driven guidance, conditional generation within their training domain. Reliably sampling from high-value regions beyond the training data, however, remains an open challenge---with current methods predominantly focusing on modifying the diffusion process itself. In this paper, we develop context-guided diffusion (CGD), a simple plug-and-play method that leverages unlabeled data and smoothness constraints to improve the out-of-distribution generalization of guided diffusion models. We demonstrate that this approach leads to substantial performance gains across various settings, including continuous, discrete, and graph-structured diffusion processes with applications across drug discovery, materials science, and protein design.

Extended-connectivity fingerprints (ECFPs) are a ubiquitous tool in current cheminformatics and molecular machine learning, and one of the most prevalent molecular feature extraction techniques used for chemical prediction. Atom features learned by graph neural networks can be aggregated to compound-level representations using a large spectrum of graph pooling methods. In contrast, sets of detected ECFP substructures are by default transformed into bit vectors using only a simple hash-based folding procedure. We introduce a general mathematical framework for the vectorisation of structural fingerprints via a formal operation called substructure pooling that encompasses hash-based folding, algorithmic substructure selection, and a wide variety of other potential techniques. We go on to describe Sort Slice, an easy-to-implement and bit-collision-free alternative to hash-based folding for the pooling of ECFP substructures. Sort Slice first sorts ECFP substructures according to their relative prevalence in a given set of training compounds and then slices away all but the L most frequent substructures which are subsequently used to generate a binary fingerprint of desired length, L. We computationally compare the performance of hash-based folding, Sort Slice, and two advanced supervised substructure-selection schemes (filtering and mutual-information maximisation) for ECFP-based molecular property prediction. Our results indicate that, despite its technical simplicity, Sort Slice robustly (and at times substantially) outperforms traditional hash-based folding as well as the other investigated substructure-pooling methods across distinct prediction tasks, data splitting techniques, machine-learning models and ECFP hyperparameters. We thus recommend that Sort Slice canonically replace hash-based folding as the default substructure-pooling technique to vectorise ECFPs for supervised molecular machine learning. Scientific contribution A general mathematical framework for the vectorisation of structural fingerprints called substructure pooling; and the technical description and computational evaluation of Sort Slice, a conceptually simple and bit-collision-free method for the pooling of ECFP substructures that robustly and markedly outperforms classical hash-based folding at molecular property prediction.

The last few years have seen the development of numerous deep learning-based protein-ligand docking methods. They offer huge promise in terms of speed and accuracy. However, despite claims of state-of-the-art performance in terms of crystallographic root-mean-square deviation (RMSD), upon closer inspection, it has become apparent that they often produce physically implausible molecular structures. It is therefore not sufficient to evaluate these methods solely by RMSD to a native binding mode. It is vital, particularly for deep learning-based methods, that they are also evaluated on steric and energetic criteria. We present PoseBusters, a Python package that performs a series of standard quality checks using the well-established cheminformatics toolkit RDKit. Only methods that both pass these checks and predict native-like binding modes should be classed as having "state-of-the-art" performance. We use PoseBusters to compare five deep learning-based docking methods (DeepDock, DiffDock, EquiBind, TankBind, and Uni-Mol) and two well-established standard docking methods (AutoDock Vina and CCDC Gold) with and without an additional post-prediction energy minimisation step using a molecular mechanics force field. We show that both in terms of physical plausibility and the ability to generalise to examples that are distinct from the training data, no deep learning-based method yet outperforms classical docking tools. In addition, we find that molecular mechanics force fields contain docking-relevant physics missing from deep-learning methods. PoseBusters allows practitioners to assess docking and molecular generation methods and may inspire new inductive biases still required to improve deep learning-based methods, which will help drive the development of more accurate and more realistic predictions.

CC and CXC-chemokines are the primary drivers of chemotaxis in inflammation, but chemokine network redundancy thwarts pharmacological intervention. Tick evasins promiscuously bind CC and CXC-chemokines, overcoming redundancy. Here we show that short peptides that promiscuously bind both chemokine classes can be identified from evasins by phage-display screening performed with multiple chemokines in parallel. We identify two conserved motifs within these peptides and show using saturation-mutagenesis phage-display and chemotaxis studies of an exemplar peptide that an anionic patch in the first motif and hydrophobic, aromatic and cysteine residues in the second are functionally necessary. AlphaFold2-Multimer modelling suggests that the peptide occludes distinct receptor-binding regions in CC and in CXC-chemokines, with the first and second motifs contributing ionic and hydrophobic interactions respectively. Our results indicate that peptides with broad-spectrum anti-chemokine activity and therapeutic potential may be identified from evasins, and the pharmacophore characterised by phage display, saturation mutagenesis and computational modelling.

Accelerating the discovery of novel and more effective therapeutics is an important pharmaceutical problem in which deep learning is playing an increasingly significant role. However, real-world drug discovery tasks are often characterized by a scarcity of labeled data and significant covariate shift$\unicode{x2013}\unicode{x2013}$a setting that poses a challenge to standard deep learning methods. In this paper, we present Q-SAVI, a probabilistic model able to address these challenges by encoding explicit prior knowledge of the data-generating process into a prior distribution over functions, presenting researchers with a transparent and probabilistically principled way to encode data-driven modeling preferences. Building on a novel, gold-standard bioactivity dataset that facilitates a meaningful comparison of models in an extrapolative regime, we explore different approaches to induce data shift and construct a challenging evaluation setup. We then demonstrate that using Q-SAVI to integrate contextualized prior knowledge of drug-like chemical space into the modeling process affords substantial gains in predictive accuracy and calibration, outperforming a broad range of state-of-the-art self-supervised pre-training and domain adaptation techniques.

Pairs of similar compounds that only differ by a small structural modification but exhibit a large difference in their binding affinity for a given target are known as activity cliffs (ACs). It has been hypothesised that quantitative structure-activity relationship (QSAR) models struggle to predict ACs and that ACs thus form a major source of prediction error. However, a study to explore the AC-prediction power of modern QSAR methods and its relationship to general QSAR-prediction performance is lacking. We systematically construct nine distinct QSAR models by combining three molecular representation methods (extended-connectivity fingerprints, physicochemical-descriptor vectors and graph isomorphism networks) with three regression techniques (random forests, k-nearest neighbours and multilayer perceptrons); we then use each resulting model to classify pairs of similar compounds as ACs or non-ACs and to predict the activities of individual molecules in three case studies: dopamine receptor D2, factor Xa, and SARS-CoV-2 main protease. We observe low AC-sensitivity amongst the tested models when the activities of both compounds are unknown, but a substantial increase in AC-sensitivity when the actual activity of one of the compounds is given. Graph isomorphism features are found to be competitive with or superior to classical molecular representations for AC-classification and can thus be employed as baseline AC-prediction models or simple compound-optimisation tools. For general QSAR-prediction, however, extended-connectivity fingerprints still consistently deliver the best performance. Our results provide strong support for the hypothesis that indeed QSAR methods frequently fail to predict ACs. We propose twin-network training for deep learning models as a potential future pathway to increase AC-sensitivity and thus overall QSAR performance.

The rapid and accurate in silico prediction of protein-ligand binding free energies or binding affinities has the potential to transform drug discovery. In recent years, there has been a rapid growth of interest in deep learning methods for the prediction of protein-ligand binding affinities based on the structural information of protein-ligand complexes. These structure-based scoring functions often obtain better results than classical scoring functions when applied within their applicability domain. Here we review structure-based scoring functions for binding affinity prediction based on deep learning, focussing on different types of architectures, featurization strategies, data sets, methods for training and evaluation, and the role of explainable artificial intelligence in building useful models for real drug-discovery applications.

Graph Neural Networks (GNNs) have recently gained in popularity, challenging molecular fingerprints or SMILES-based representations as the predominant way to represent molecules for binding affinity prediction. Although simple ligand-based graphs alone are already useful for affinity prediction, better performance on multi-target datasets has been achieved with models that incorporate 3D structural information. Most recent advances utilize complex GNN architectures to capture 3D protein-ligand information by incorporating ligand-interacting protein atoms as additional nodes in the graphs; or by building a second protein-based graph in parallel. This expands the graph considerably while obfuscating the shape of the underlying ligand, diminishing the advantage that GNNs have when encoding molecular structures. There is therefore a need for a simple and elegant molecular graph representation that retains the topology of the ligand while simultaneously encoding 3D protein-ligand interactions. We present Protein-Ligand Interaction Graphs (PLIGs): a simple way of representing atom-atom contacts of 3D protein-ligand complexes as node features for GNNs. PLIGs featurize an atom node in the molecular graph by describing each atom’s properties as well as all atom-atom contacts made with protein atoms within a distance threshold. The edges of the graph are therefore identical to ligand-based graphs, but the nodes encode the 3D protein-ligand contacts. Since PLIGs are applicable to any GNN architecture, we have benchmarked their performance with six different GNN architectures, and compared them to conventional ligand-based graphs and fingerprint-based multi-layer perceptron (MLP) models using the CASF-2016 benchmark set where we found PLIG-based Graph Attention Networks (GATNet) to be the best performing model ( ρ =0.84, RMSE=1.22 pK). In summary, we created a novel graph-based representation that incorporates 3D structural information into the node features of ligand-shaped molecular graphs. The PLIG representation is simple, elegant, flexible and easily customizable, opening up many possibilities of incorporating other 2D and 3D properties into the graph. Access The code and implementation for PLIGs and all models can be found at github.com/MarcMoesser/Protein-Ligand-Interaction-Graphs .

The SARS-CoV-2 coronavirus (CoV) causes COVID-19, a current global pandemic. SARS-CoV-2 belongs to an order of Nidovirales with very large RNA genomes. It is proposed that the fidelity of CoV genome replication is aided by an RNA nuclease complex, formed of non-structural proteins 14 and 10 (nsp14-nsp10), an attractive target for antiviral inhibition. Here, we confirm that the SARS-CoV-2 nsp14-nsp10 complex is an RNase. Detailed functional characterisation reveals nsp14-nsp10 is a highly versatile nuclease capable of digesting a wide variety of RNA structures, including those with a blocked 3’-terminus. We propose that the role of nsp14-nsp10 in maintaining replication fidelity goes beyond classical proofreading and purges the nascent replicating RNA strand of a range of potentially replication terminating aberrations. Using the developed assays, we identify a series of drug and drug-like molecules that potently inhibit nsp14-nsp10, including the known Sars-Cov-2 major protease (M pro ) inhibitor ebselen and the HIV integrase inhibitor raltegravir, revealing the potential for bifunctional inhibitors in the treatment of COVID-19.

TRIM33 is a member of the tripartite motif (TRIM) family of proteins, some of which possess E3 ligase activity and are involved in the ubiquitin-dependent degradation of proteins. Four of the TRIM family proteins, TRIM24 (TIF1α), TRIM28 (TIF1β), TRIM33 (TIF1γ) and TRIM66, contain C-terminal plant homeodomain (PHD) and bromodomain (BRD) modules, which bind to methylated lysine (KMen) and acetylated lysine (KAc), respectively. Here we investigate the differences between the two isoforms of TRIM33, TRIM33α and TRIM33β, using structural and biophysical approaches. We show that the N1039 residue, which is equivalent to N140 in BRD4(1) and which is conserved in most BRDs, has a different orientation in each isoform. In TRIM33β, this residue coordinates KAc, but this is not the case in TRIM33α. Despite these differences, both isoforms show similar affinities for H31-27K18Ac, and bind preferentially to H31-27K9Me3K18Ac. We used this information to develop an AlphaScreen assay, with which we have identified four new ligands for the TRIM33 PHD-BRD cassette. These findings provide fundamental new information regarding which histone marks are recognized by both isoforms of TRIM33 and suggest starting points for the development of chemical probes to investigate the cellular function of TRIM33.

Drug resistance caused by mutations is a public health threat for existing and emerging viral diseases. A wealth of evidence about these mutations and their clinically associated phenotypes is scattered across the literature, but a comprehensive perspective is usually lacking. This work aimed to produce a clinically relevant view for the case of Hepatitis B virus (HBV) mutations by combining a chronic HBV clinical study with a compendium of genetic mutations systematically gathered from the scientific literature. We enriched clinical mutation data by systematically mining 2,472,725 scientific articles from PubMed Central in order to gather information about the HBV mutational landscape. By performing this analysis, we were able to identify mutational hotspots for each HBV genotype (A-E) and gene (C, X, P, S), as well as the location of disulfide bonds associated with these mutations. Through a modelling study, we also identified a mutation position common in both the clinical data and the literature that is located at the binding pocket for a known anti-HBV drug, namely entecavir. The results of this novel approach show the potential of integrated analyses to assist in the development of new drugs for viral diseases that are more robust to resistance. Such analyses should be of particular interest due to the increasing importance of viral resistance in established and emerging viruses, such as for newly developed drugs against SARS-2.

Machine learning scoring functions for protein-ligand binding affinity have been found to consistently outperform classical scoring functions when trained and tested on crystal structures of bound protein-ligand complexes. However, it is less clear how these methods perform when applied to docked poses of complexes.We explore how the use of docked, rather than crystallographic, poses for both training and testing affects the performance of machine learning scoring functions. Using the PDBbind Core Sets as benchmarks, we show that the performance of a structure-based machine learning scoring function trained and tested on docked poses is lower than that of the same scoring function trained and tested on crystallographic poses. We construct a hybrid scoring function by combining both structure-based and ligand-based features, and show that its ability to predict binding affinity using docked poses is comparable to that of purely structure-based scoring functions trained and tested on crystal poses. Despite strong performance on docked poses of the PDBbind Core Sets, we find that our hybrid scoring function fails to generalise to anew data set, demonstrating the need for improved scoring functions and additional validation benchmarks. Code and data to reproduce our results are available from https://github.com/oxpig/learning-from-docked-poses.