Do Chemformers Dream of Organic Matter? Evaluating a Transformer Model for Multistep Retrosynthesis.

Synthesis planning of new pharmaceutical compounds is a well-known bottleneck in modern drug design. Template-free methods, such as transformers, have recently been proposed as an alternative to template-based methods for single-step retrosynthetic predictions. Here, we trained and evaluated a transformer model, called the Chemformer, for retrosynthesis predictions within drug discovery. The proprietary data set used for training comprised ∼18 M reactions from literature, patents, and electronic lab notebooks. Chemformer was evaluated for the purpose of both single-step and multistep retrosynthesis. We found that the single-step performance of Chemformer was especially good on reaction classes common in drug discovery, with most reaction classes showing a top-10 round-trip accuracy above 0.97. Moreover, Chemformer reached a higher round-trip accuracy compared to that of a template-based model. By analyzing multistep retrosynthesis experiments, we observed that Chemformer found synthetic routes, leading to commercial starting materials for 95% of the target compounds, an increase of more than 20% compared to the template-based model on a proprietary compound data set. In addition to this, we discovered that Chemformer suggested novel disconnections corresponding to reaction templates, which are not included in the template-based model. These findings were further supported by a publicly available ChEMBL compound data set. The conclusions drawn from this work allow for the design of a synthesis planning tool where template-based and template-free models work in harmony to optimize retrosynthetic recommendations.

Data-driven approaches for identifying hyperparameters in multi-step retrosynthesis.

The multi-step retrosynthesis problem can be solved by a search algorithm, such as Monte Carlo tree search (MCTS). The performance of multistep retrosynthesis, as measured by a trade-off in search time and route solvability, therefore depends on the hyperparameters of the search algorithm. In this paper, we demonstrated the effect of three MCTS hyperparameters (number of iterations, tree depth, and tree width) on metrics such as Linear integrated speed-accuracy score (LISAS) and Inverse efficiency score which consider both route solvability and search time. This exploration was conducted by employing three data-driven approaches, namely a systematic grid search, Bayesian optimization over an ensemble of molecules to obtain static MCTS hyperparameters, and a machine learning approach to dynamically predict optimal MCTS hyperparameters given an input target molecule. With the obtained results on the internal dataset, we demonstrated that it is possible to identify a hyperparameter set which outperforms the current AiZynthFinder default setting. It appeared optimal across a variety of target input molecules, both on proprietary and public datasets. The settings identified with the in-house dataset reached a solvability of 93 % and median search time of 151 s for the in-house dataset, and a 74 % solvability and 114 s for the ChEMBL dataset. These numbers can be compared to the current default settings which solved 85 % and 73 % during a median time of 110s and 84 s, for in-house and ChEMBL, respectively.

Markov state modelling reveals heterogeneous drug-inhibition mechanism of Calmodulin.

Calmodulin (CaM) is a calcium sensor which binds and regulates a wide range of target-proteins. This implicitly enables the concentration of calcium to influence many downstream physiological responses, including muscle contraction, learning and depression. The antipsychotic drug trifluoperazine (TFP) is a known CaM inhibitor. By binding to various sites, TFP prevents CaM from associating to target-proteins. However, the molecular and state-dependent mechanisms behind CaM inhibition by drugs such as TFP are largely unknown. Here, we build a Markov state model (MSM) from adaptively sampled molecular dynamics simulations and reveal the structural and dynamical features behind the inhibitory mechanism of TFP-binding to the C-terminal domain of CaM. We specifically identify three major TFP binding-modes from the MSM macrostates, and distinguish their effect on CaM conformation by using a systematic analysis protocol based on biophysical descriptors and tools from machine learning. The results show that depending on the binding orientation, TFP effectively stabilizes features of the calcium-unbound CaM, either affecting the CaM hydrophobic binding pocket, the calcium binding sites or the secondary structure content in the bound domain. The conclusions drawn from this work may in the future serve to formulate a complete model of pharmacological modulation of CaM, which furthers our understanding of how these drugs affect signaling pathways as well as associated diseases.

Vascular Tissue Specific miRNA Profiles Reveal Novel Correlations with Risk Factors in Coronary Artery Disease.

Cardiovascular disease (CVD) is the leading cause of morbidity and mortality worldwide. Non-coding RNAs have already been linked to CVD development and progression. While microRNAs (miRs) have been well studied in blood samples, there is little data on tissue-specific miRs in cardiovascular relevant tissues and their relation to cardiovascular risk factors. Tissue-specific miRs derived from Arteria mammaria interna (IMA) from 192 coronary artery disease (CAD) patients undergoing coronary artery bypass grafting (CABG) were analyzed. The aims of the study were 1) to establish a reference atlas which can be utilized for identification of novel diagnostic biomarkers and potential therapeutic targets, and 2) to relate these miRs to cardiovascular risk factors. Overall, 393 individual miRs showed sufficient expression levels and passed quality control for further analysis. We identified 17 miRs-miR-10b-3p, miR-10-5p, miR-17-3p, miR-21-5p, miR-151a-5p, miR-181a-5p, miR-185-5p, miR-194-5p, miR-199a-3p, miR-199b-3p, miR-212-3p, miR-363-3p, miR-548d-5p, miR-744-5p, miR-3117-3p, miR-5683 and miR-5701-significantly correlated with cardiovascular risk factors (correlation coefficient >0.2 in both directions, p-value (p < 0.006, false discovery rate (FDR) <0.05). Of particular interest, miR-5701 was positively correlated with hypertension, hypercholesterolemia, and diabetes. In addition, we found that miR-629-5p and miR-98-5p were significantly correlated with acute myocardial infarction. We provide a first atlas of miR profiles in IMA samples from CAD patients. In perspective, these miRs might play an important role in improved risk assessment, mechanistic disease understanding and local therapy of CAD.

Risk Prediction of Cardiovascular Events by Exploration of Molecular Data with Explainable Artificial Intelligence.

Cardiovascular diseases (CVD) annually take almost 18 million lives worldwide. Most lethal events occur months or years after the initial presentation. Indeed, many patients experience repeated complications or require multiple interventions (recurrent events). Apart from affecting the individual, this leads to high medical costs for society. Personalized treatment strategies aiming at prediction and prevention of recurrent events rely on early diagnosis and precise prognosis. Complementing the traditional environmental and clinical risk factors, multi-omics data provide a holistic view of the patient and disease progression, enabling studies to probe novel angles in risk stratification. Specifically, predictive molecular markers allow insights into regulatory networks, pathways, and mechanisms underlying disease. Moreover, artificial intelligence (AI) represents a powerful, yet adaptive, framework able to recognize complex patterns in large-scale clinical and molecular data with the potential to improve risk prediction. Here, we review the most recent advances in risk prediction of recurrent cardiovascular events, and discuss the value of molecular data and biomarkers for understanding patient risk in a systems biology context. Finally, we introduce explainable AI which may improve clinical decision systems by making predictions transparent to the medical practitioner.

Calmodulin acts as a state-dependent switch to control a cardiac potassium channel opening.

Calmodulin (CaM) and phosphatidylinositol 4,5-bisphosphate (PIP2) are potent regulators of the voltage-gated potassium channel KCNQ1 (KV7.1), which conducts the cardiac I Ks current. Although cryo-electron microscopy structures revealed intricate interactions between the KCNQ1 voltage-sensing domain (VSD), CaM, and PIP2, the functional consequences of these interactions remain unknown. Here, we show that CaM-VSD interactions act as a state-dependent switch to control KCNQ1 pore opening. Combined electrophysiology and molecular dynamics network analysis suggest that VSD transition into the fully activated state allows PIP2 to compete with CaM for binding to VSD. This leads to conformational changes that alter VSD-pore coupling to stabilize open states. We identify a motif in the KCNQ1 cytosolic domain, which works downstream of CaM-VSD interactions to facilitate the conformational change. Our findings suggest a gating mechanism that integrates PIP2 and CaM in KCNQ1 voltage-dependent activation, yielding insights into how KCNQ1 gains the phenotypes critical for its physiological function.

Network analysis reveals how lipids and other cofactors influence membrane protein allostery.

Many membrane proteins are modulated by external stimuli, such as small molecule binding or change in pH, transmembrane voltage, or temperature. This modulation typically occurs at sites that are structurally distant from the functional site. Revealing the communication, known as allostery, between these two sites is key to understanding the mechanistic details of these proteins. Residue interaction networks of isolated proteins are commonly used to this end. Membrane proteins, however, are embedded in a lipid bilayer, which may contribute to allosteric communication. The fast diffusion of lipids hinders direct use of standard residue interaction networks. Here, we present an extension that includes cofactors such as lipids and small molecules in the network. The novel framework is applied to three membrane proteins: a voltage-gated ion channel (KCNQ1), a G-protein coupled receptor (GPCR-ß2 adrenergic receptor), and a pH-gated ion channel (KcsA). Through systematic analysis of the obtained networks and their components, we demonstrate the importance of lipids for membrane protein allostery. Finally, we reveal how small molecules may stabilize different protein states by allosterically coupling and decoupling the protein from the membrane.

Tracking the motion of the KV1.2 voltage sensor reveals the molecular perturbations caused by a de novo mutation in a case of epilepsy.

KEY POINTS: KV1.2 channels, encoded by the KCNA2 gene, regulate neuronal excitability by conducting K+ upon depolarization. A new KCNA2 missense variant was discovered in a patient with epilepsy, causing amino acid substitution F302L at helix S4, in the KV1.2 voltage-sensing domain. Immunocytochemistry and flow cytometry showed that F302L does not impair KCNA2 subunit surface trafficking. Molecular dynamics simulations indicated that F302L alters the exposure of S4 residues to membrane lipids. Voltage clamp fluorometry revealed that the voltage-sensing domain of KV1.2-F302L channels is more sensitive to depolarization. Accordingly, KV1.2-F302L channels opened faster and at more negative potentials; however, they also exhibited enhanced inactivation: that is, F302L causes both gain- and loss-of-function effects. Coexpression of KCNA2-WT and -F302L did not fully rescue these effects. The proband's symptoms are more characteristic of patients with loss of KCNA2 function. Enhanced KV1.2 inactivation could lead to increased synaptic release in excitatory neurons, steering neuronal circuits towards epilepsy. ABSTRACT: An exome-based diagnostic panel in an infant with epilepsy revealed a previously unreported de novo missense variant in KCNA2, which encodes voltage-gated K+ channel KV1.2. This variant causes substitution F302L, in helix S4 of the KV1.2 voltage-sensing domain (VSD). F302L does not affect KCNA2 subunit membrane trafficking. However, it does alter channel functional properties, accelerating channel opening at more hyperpolarized membrane potentials, indicating gain of function. F302L also caused loss of KV1.2 function via accelerated inactivation onset, decelerated recovery and shifted inactivation voltage dependence to more negative potentials. These effects, which are not fully rescued by coexpression of wild-type and mutant KCNA2 subunits, probably result from the enhancement of VSD function, as demonstrated by optically tracking VSD depolarization-evoked conformational rearrangements. In turn, molecular dynamics simulations suggest altered VSD exposure to membrane lipids. Compared to other encephalopathy patients with KCNA2 mutations, the proband exhibits mild neurological impairment, more characteristic of patients with KCNA2 loss of function. Based on this information, we propose a mechanism of epileptogenesis based on enhanced KV1.2 inactivation leading to increased synaptic release preferentially in excitatory neurons, and hence the perturbation of the excitatory/inhibitory balance of neuronal circuits.

Molecular Insights from Conformational Ensembles via Machine Learning.

Biomolecular simulations are intrinsically high dimensional and generate noisy data sets of ever-increasing size. Extracting important features from the data is crucial for understanding the biophysical properties of molecular processes, but remains a big challenge. Machine learning (ML) provides powerful dimensionality reduction tools. However, such methods are often criticized as resembling black boxes with limited human-interpretable insight. We use methods from supervised and unsupervised ML to efficiently create interpretable maps of important features from molecular simulations. We benchmark the performance of several methods, including neural networks, random forests, and principal component analysis, using a toy model with properties reminiscent of macromolecular behavior. We then analyze three diverse biological processes: conformational changes within the soluble protein calmodulin, ligand binding to a G protein-coupled receptor, and activation of an ion channel voltage-sensor domain, unraveling features critical for signal transduction, ligand binding, and voltage sensing. This work demonstrates the usefulness of ML in understanding biomolecular states and demystifying complex simulations.

InfleCS: Clustering Free Energy Landscapes with Gaussian Mixtures.

Free energy landscapes provide insights into conformational ensembles of biomolecules. In order to analyze these landscapes and elucidate mechanisms underlying conformational changes, there is a need to extract metastable states with limited noise. This has remained a formidable task, despite a plethora of existing clustering methods. We present InfleCS, a novel method for extracting well-defined core states from free energy landscapes. The method is based on a Gaussian mixture free energy estimator and exploits the shape of the estimated density landscape. The core states that naturally arise from the clustering allow for detailed characterization of the conformational ensemble. The clustering quality is evaluated on three toy models with different properties, where the method is shown to consistently outperform other conventional and state-of-the-art clustering methods. Finally, the method is applied to a temperature enhanced molecular dynamics simulation of Ca2+-bound Calmodulin. Through the free energy landscape, we discover a pathway between a canonical and a compact state, revealing conformational changes driven by electrostatic interactions.

Effect of Ca2+ on the promiscuous target-protein binding of calmodulin.

Calmodulin (CaM) is a calcium sensing protein that regulates the function of a large number of proteins, thus playing a crucial part in many cell signaling pathways. CaM has the ability to bind more than 300 different target peptides in a Ca2+-dependent manner, mainly through the exposure of hydrophobic residues. How CaM can bind a large number of targets while retaining some selectivity is a fascinating open question. Here, we explore the mechanism of CaM selective promiscuity for selected target proteins. Analyzing enhanced sampling molecular dynamics simulations of Ca2+-bound and Ca2+-free CaM via spectral clustering has allowed us to identify distinct conformational states, characterized by interhelical angles, secondary structure determinants and the solvent exposure of specific residues. We searched for indicators of conformational selection by mapping solvent exposure of residues in these conformational states to contacts in structures of CaM/target peptide complexes. We thereby identified CaM states involved in various binding classes arranged along a depth binding gradient. Binding Ca2+ modifies the accessible hydrophobic surface of the two lobes and allows for deeper binding. Apo CaM indeed shows shallow binding involving predominantly polar and charged residues. Furthermore, binding to the C-terminal lobe of CaM appears selective and involves specific conformational states that can facilitate deep binding to target proteins, while binding to the N-terminal lobe appears to happen through a more flexible mechanism. Thus the long-ranged electrostatic interactions of the charged residues of the N-terminal lobe of CaM may initiate binding, while the short-ranged interactions of hydrophobic residues in the C-terminal lobe of CaM may account for selectivity. This work furthers our understanding of the mechanism of CaM binding and selectivity to different target proteins and paves the way towards a comprehensive model of CaM selectivity.

Inference of Calmodulin's Ca2+-Dependent Free Energy Landscapes via Gaussian Mixture Model Validation.

A free energy landscape estimation method based on the well-known Gaussian mixture model (GMM) is used to compare the efficiencies of thermally enhanced sampling methods with respect to regular molecular dynamics. The simulations are carried out on two binding states of calmodulin, and the free energy estimation method is compared with other estimators using a toy model. We show that GMM with cross-validation provides a robust estimate that is not subject to overfitting. The continuous nature of Gaussians provides better estimates on sparse data than canonical histogramming. We find that diffusion properties determine the sampling method effectiveness, such that diffusion-dominated apo calmodulin is most efficiently sampled by regular molecular dynamics, while holo calmodulin, with its rugged free energy landscape, is better sampled by enhanced sampling methods.