AbLEF: antibody language ensemble fusion for thermodynamically empowered property predictions.

MOTIVATION: Pre-trained protein language and/or structural models are often fine-tuned on drug development properties (i.e. developability properties) to accelerate drug discovery initiatives. However, these models generally rely on a single structural conformation and/or a single sequence as a molecular representation. We present a physics-based model, whereby 3D conformational ensemble representations are fused by a transformer-based architecture and concatenated to a language representation to predict antibody protein properties. Antibody language ensemble fusion enables the direct infusion of thermodynamic information into latent space and this enhances property prediction by explicitly infusing dynamic molecular behavior that occurs during experimental measurement. RESULTS: We showcase the antibody language ensemble fusion model on two developability properties: hydrophobic interaction chromatography retention time and temperature of aggregation (Tagg). We find that (i) 3D conformational ensembles that are generated from molecular simulation can further improve antibody property prediction for small datasets, (ii) the performance benefit from 3D conformational ensembles matches shallow machine learning methods in the small data regime, and (iii) fine-tuned large protein language models can match smaller antibody-specific language models at predicting antibody properties. AVAILABILITY AND IMPLEMENTATION: AbLEF codebase is available at https://github.com/merck/AbLEF.

AbMelt: Learning antibody thermostability from molecular dynamics.

Antibody thermostability is challenging to predict from sequence and/or structure. This difficulty is likely due to the absence of direct entropic information. Herein, we present AbMelt where we model the inherent flexibility of homologous antibody structures using molecular dynamics simulations at three temperatures and learn the relevant descriptors to predict the temperatures of aggregation (Tagg), melt onset (Tm,on), and melt (Tm). We observed that the radius of gyration deviation of the complementarity determining regions at 400 K is the highest Pearson correlated descriptor with aggregation temperature (rp = -0.68 ± 0.23) and the deviation of internal molecular contacts at 350 K is the highest correlated descriptor with both Tm,on (rp = -0.74 ± 0.04) as well as Tm (rp = -0.69 ± 0.03). Moreover, after descriptor selection and machine learning regression, we predict on a held-out test set containing both internal and public data and achieve robust performance for all endpoints compared with baseline models (Tagg R2 = 0.57 ± 0.11, Tm,on R2 = 0.56 ± 0.01, and Tm R2 = 0.60 ± 0.06). In addition, the robustness of the AbMelt molecular dynamics methodology is demonstrated by only training on <5% of the data and outperforming more traditional machine learning models trained on the entire data set of more than 500 internal antibodies. Users can predict thermostability measurements for antibody variable fragments by collecting descriptors and using AbMelt, which has been made available.

Exploring Deep Learning of Quantum Chemical Properties for Absorption, Distribution, Metabolism, and Excretion Predictions.

Quantum mechanical (QM) descriptors of small molecules have wide applicability in understanding organic reactivity and molecular properties, but the substantial compute cost required for ab initio QM calculations limits their broad usage. Here, we investigate the use of deep learning for predicting QM descriptors, with the goal of enabling usage of near-QM accuracy electronic properties on large molecular data sets such as those seen in drug discovery. Several deep learning approaches have previously been benchmarked on a published data set called QM9, where 12 ground-state properties have been calculated for molecules with up to nine heavy atoms, limited to C, H, N, O, and F elements. To advance the work beyond the QM9 chemical space and enable application to molecules encountered in drug discovery, we extend the QM9 data set by creating a QM9-extended data set covering an additional ∼20,000 molecules containing S and Cl atoms. Using this extended set, we generate new deep learning models as well as leverage ANI-2x models to provide predictions on larger, more diverse molecules common in drug discovery, and we find the models estimate 11 of 12 ground-state properties reasonably. We use the predicted QM descriptors to augment graph convolutional neural network (GCNN) models for selected ADME end points (rat microsomal clearance, hepatic clearance, total clearance, and P-glycoprotein efflux) and found varying degrees of performance improvement compared to nonaugmented GCNN models, including pronounced improvement in P-glycoprotein efflux prediction.

Experimental Error, Kurtosis, Activity Cliffs, and Methodology: What Limits the Predictivity of Quantitative Structure-Activity Relationship Models?

Given a particular descriptor/method combination, some quantitative structure-activity relationship (QSAR) datasets are very predictive by random-split cross-validation while others are not. Recent literature in modelability suggests that the limiting issue for predictivity is in the data, not the QSAR methodology, and the limits are due to activity cliffs. Here, we investigate, on in-house data, the relative usefulness of experimental error, distribution of the activities, and activity cliff metrics in determining how predictive a dataset is likely to be. We include unmodified in-house datasets, datasets that should be perfectly predictive based only on the chemical structure, datasets where the distribution of activities is manipulated, and datasets that include a known amount of added noise. We find that activity cliff metrics determine predictivity better than the other metrics we investigated, whatever the type of dataset, consistent with the modelability literature. However, such metrics cannot distinguish real activity cliffs due to large uncertainties in the activities. We also show that a number of modern QSAR methods, and some alternative descriptors, are equally bad at predicting the activities of compounds on activity cliffs, consistent with the assumptions behind "modelability." Finally, we relate time-split predictivity with random-split predictivity and show that different coverages of chemical space are at least as important as uncertainty in activity and/or activity cliffs in limiting predictivity.

Collaborating to improve the use of free-energy and other quantitative methods in drug discovery.

In May and August, 2016, several pharmaceutical companies convened to discuss and compare experiences with Free Energy Perturbation (FEP). This unusual synchronization of interest was prompted by Schrödinger's FEP+ implementation and offered the opportunity to share fresh studies with FEP and enable broader discussions on the topic. This article summarizes key conclusions of the meetings, including a path forward of actions for this group to aid the accelerated evaluation, application and development of free energy and related quantitative, structure-based design methods.

Structure-based druggability assessment of the mammalian structural proteome with inclusion of light protein flexibility.

Advances reported over the last few years and the increasing availability of protein crystal structure data have greatly improved structure-based druggability approaches. However, in practice, nearly all druggability estimation methods are applied to protein crystal structures as rigid proteins, with protein flexibility often not directly addressed. The inclusion of protein flexibility is important in correctly identifying the druggability of pockets that would be missed by methods based solely on the rigid crystal structure. These include cryptic pockets and flexible pockets often found at protein-protein interaction interfaces. Here, we apply an approach that uses protein modeling in concert with druggability estimation to account for light protein backbone movement and protein side-chain flexibility in protein binding sites. We assess the advantages and limitations of this approach on widely-used protein druggability sets. Applying the approach to all mammalian protein crystal structures in the PDB results in identification of 69 proteins with potential druggable cryptic pockets.

Development of 2-aminooxazoline 3-azaxanthenes as orally efficacious ß-secretase inhibitors for the potential treatment of Alzheimer's disease.

The ß-site amyloid precursor protein (APP) cleaving enzyme 1 (BACE1) is one of the most hotly pursued targets for the treatment of Alzheimer's disease. We used a structure- and property-based drug design approach to identify 2-aminooxazoline 3-azaxanthenes as potent BACE1 inhibitors which significantly reduced CSF and brain Aß levels in a rat pharmacodynamic model. Compared to the initial lead 2, compound 28 exhibited reduced potential for QTc prolongation in a non-human primate cardiovascular safety model.

MolPROP: Molecular Property prediction with multimodal language and graph fusion.

Pretrained deep learning models self-supervised on large datasets of language, image, and graph representations are often fine-tuned on downstream tasks and have demonstrated remarkable adaptability in a variety of applications including chatbots, autonomous driving, and protein folding. Additional research aims to improve performance on downstream tasks by fusing high dimensional data representations across multiple modalities. In this work, we explore a novel fusion of a pretrained language model, ChemBERTa-2, with graph neural networks for the task of molecular property prediction. We benchmark the MolPROP suite of models on seven scaffold split MoleculeNet datasets and compare with state-of-the-art architectures. We find that (1) multimodal property prediction for small molecules can match or significantly outperform modern architectures on hydration free energy (FreeSolv), experimental water solubility (ESOL), lipophilicity (Lipo), and clinical toxicity tasks (ClinTox), (2) the MolPROP multimodal fusion is predominantly beneficial on regression tasks, (3) the ChemBERTa-2 masked language model pretraining task (MLM) outperformed multitask regression pretraining task (MTR) when fused with graph neural networks for multimodal property prediction, and (4) despite improvements from multimodal fusion on regression tasks MolPROP significantly underperforms on some classification tasks. MolPROP has been made available at https://github.com/merck/MolPROP . SCIENTIFIC CONTRIBUTION: This work explores a novel multimodal fusion of learned language and graph representations of small molecules for the supervised task of molecular property prediction. The MolPROP suite of models demonstrates that language and graph fusion can significantly outperform modern architectures on several regression prediction tasks and also provides the opportunity to explore alternative fusion strategies on classification tasks for multimodal molecular property prediction.

Structure guided design of a series of sphingosine kinase (SphK) inhibitors.

Sphingosine-1-phosphate (S1P) signaling plays a vital role in mitogenesis, cell migration and angiogenesis. Sphingosine kinases (SphKs) catalyze a key step in sphingomyelin metabolism that leads to the production of S1P. There are two isoforms of SphK and observations made with SphK deficient mice show the two isoforms can compensate for each other's loss. Thus, inhibition of both isoforms is likely required to block SphK dependent angiogenesis. A structure based approach was used to design and synthesize a series of SphK inhibitors resulting in the identification of the first potent inhibitors of both isoforms of human SphK. Additionally, to our knowledge, this series of inhibitors contains the only sufficiently potent inhibitors of murine SphK1 with suitable physico-chemical properties to pharmacologically interrogate the role of SphK1 in rodent models and to reproduce the phenotype of SphK1 (-/-) mice.

A machine learning strategy for the identification of key in silico descriptors and prediction models for IgG monoclonal antibody developability properties.

Identification of favorable biophysical properties for protein therapeutics as part of developability assessment is a crucial part of the preclinical development process. Successful prediction of such properties and bioassay results from calculated in silico features has potential to reduce the time and cost of delivering clinical-grade material to patients, but nevertheless has remained an ongoing challenge to the field. Here, we demonstrate an automated and flexible machine learning workflow designed to compare and identify the most powerful features from computationally derived physiochemical feature sets, generated from popular commercial software packages. We implement this workflow with medium-sized datasets of human and humanized IgG molecules to generate predictive regression models for two key developability endpoints, hydrophobicity and poly-specificity. The most important features discovered through the automated workflow corroborate several previous literature reports, and newly discovered features suggest directions for further research and potential model improvement.

Structural insights into selective small molecule activation of PKG1α.

cGMP-dependent protein kinase I-α (PKG1α) is a target for pulmonary arterial hypertension due to its role in the regulation of smooth muscle function. While most work has focused on regulation of cGMP turnover, we recently described several small molecule tool compounds which were capable of activating PKG1α via a cGMP independent pathway. Selected molecules were crystallized in the presence of PKG1α and were found to bind to an allosteric site proximal to the low-affinity nucleotide binding domain. These molecules act to displace the switch helix and cause activation of PKG1α representing a new mechanism for the activation and control of this critical therapeutic path. The described structures are vital to understanding the function and control of this key regulatory pathway.

Discovery and multimerization of cross-reactive single-domain antibodies against SARS-like viruses to enhance potency and address emerging SARS-CoV-2 variants.

Coronaviruses have been the causative agent of three epidemics and pandemics in the past two decades, including the ongoing COVID-19 pandemic. A broadly-neutralizing coronavirus therapeutic is desirable not only to prevent and treat COVID-19, but also to provide protection for high-risk populations against future emergent coronaviruses. As all coronaviruses use spike proteins on the viral surface to enter the host cells, and these spike proteins share sequence and structural homology, we set out to discover cross-reactive biologic agents targeting the spike protein to block viral entry. Through llama immunization campaigns, we have identified single domain antibodies (VHHs) that are cross-reactive against multiple emergent coronaviruses (SARS-CoV, SARS-CoV-2, and MERS). Importantly, a number of these antibodies show sub-nanomolar potency towards all SARS-like viruses including emergent CoV-2 variants. We identified nine distinct epitopes on the spike protein targeted by these VHHs. Further, by engineering VHHs targeting distinct, conserved epitopes into multi-valent formats, we significantly enhanced their neutralization potencies compared to the corresponding VHH cocktails. We believe this approach is ideally suited to address both emerging SARS-CoV-2 variants during the current pandemic as well as potential future pandemics caused by SARS-like coronaviruses.

Discovery and optimization of potent and selective imidazopyridine and imidazopyridazine mTOR inhibitors.

mTOR is a critical regulator of cellular signaling downstream of multiple growth factors. The mTOR/PI3K/AKT pathway is frequently mutated in human cancers and is thus an important oncology target. Herein we report the evolution of our program to discover ATP-competitive mTOR inhibitors that demonstrate improved pharmacokinetic properties and selectivity compared to our previous leads. Through targeted SAR and structure-guided design, new imidazopyridine and imidazopyridazine scaffolds were identified that demonstrated superior inhibition of mTOR in cellular assays, selectivity over the closely related PIKK family and improved in vivo clearance over our previously reported benzimidazole series.

Hexamerization of Anti-SARS CoV IgG1 Antibodies Improves Neutralization Capacity.

The SARS-CoV-2 pandemic and particularly the emerging variants have deepened the need for widely available therapeutic options. We have demonstrated that hexamer-enhancing mutations in the Fc region of anti-SARS-CoV IgG antibodies lead to a noticeable improvement in IC50 in both pseudo and live virus neutralization assay compared to parental molecules. We also show that hexamer-enhancing mutants improve C1q binding to target surface. To our knowledge, this is the first time this format has been explored for application in viral neutralization and the studies provide proof-of-concept for the use of hexamer-enhanced IgG1 molecules as potential anti-viral therapeutics.

Optimization and Mechanistic Investigations of Novel Allosteric Activators of PKG1α.

Activation of PKG1α is a compelling strategy for the treatment of cardiovascular diseases. As the main effector of cyclic guanosine monophosphate (cGMP), activation of PKG1α induces smooth muscle relaxation in blood vessels, lowers pulmonary blood pressure, prevents platelet aggregation, and protects against cardiac stress. The development of activators has been mostly limited to cGMP mimetics and synthetic peptides. Described herein is the optimization of a piperidine series of small molecules to yield activators that demonstrate in vitro phosphorylation of vasodilator-stimulated phosphoprotein as well as antiproliferative effects in human pulmonary arterial smooth muscle cells. Hydrogen/deuterium exchange mass spectrometry experiments with the small molecule activators revealed a mechanism of action consistent with cGMP-induced activation, and an X-ray co-crystal structure with a construct encompassing the regulatory domains illustrated a binding mode in an allosteric pocket proximal to the low-affinity cyclic nucleotide-binding domain.

Discovery of triazine-benzimidazoles as selective inhibitors of mTOR.

mTOR is part of the PI3K/AKT pathway and is a central regulator of cell growth and survival. Since many cancers display mutations linked to the mTOR signaling pathway, mTOR has emerged as an important target for oncology therapy. Herein, we report the discovery of triazine benzimidazole inhibitors that inhibit mTOR kinase activity with up to 200-fold selectivity over the structurally homologous kinase PI3Kα. When tested in a panel of cancer cell lines displaying various mutations, a selective inhibitor from this series inhibited cellular proliferation with a mean IC(50) of 0.41 µM. Lead compound 42 demonstrated up to 83% inhibition of mTOR substrate phosphorylation in a murine pharmacodynamic model.

Discovery of a new class of integrin antibodies for fibrosis.

Lung fibrosis, or the scarring of the lung, is a devastating disease with huge unmet medical need. There are limited treatment options and its prognosis is worse than most types of cancer. We previously discovered that MK-0429 is an equipotent pan-inhibitor of αv integrins that reduces proteinuria and kidney fibrosis in a preclinical model. In the present study, we further demonstrated that MK-0429 significantly inhibits fibrosis progression in a bleomycin-induced lung injury model. In search of newer integrin inhibitors for fibrosis, we characterized monoclonal antibodies discovered using Adimab's yeast display platform. We identified several potent neutralizing integrin antibodies with unique human and mouse cross-reactivity. Among these, Ab-31 blocked the binding of multiple αv integrins to their ligands with IC50s comparable to those of MK-0429. Furthermore, both MK-0429 and Ab-31 suppressed integrin-mediated cell adhesion and latent TGFß activation. In IPF patient lung fibroblasts, TGFß treatment induced profound αSMA expression in phenotypic imaging assays and Ab-31 demonstrated potent in vitro activity at inhibiting αSMA expression, suggesting that the integrin antibody is able to modulate TGFß action though mechanisms beyond the inhibition of latent TGFß activation. Together, our results highlight the potential to develop newer integrin therapeutics for the treatment of fibrotic lung diseases.

Enhanced selectivity profile of pyrazole-urea based DFG-out p38alpha inhibitors.

By targeting an extended region of the conventional 'DFG-out' pocket of p38alpha, while minimizing interactions with the specificity pocket and eliminating interactions with the adenine binding site, we are able to design and synthesize a number of pyrazole-urea based DFG-out p38alpha inhibitors with good potencies, and excellent selectivity.

Structure-based maximal affinity model predicts small-molecule druggability.

Lead generation is a major hurdle in small-molecule drug discovery, with an estimated 60% of projects failing from lack of lead matter or difficulty in optimizing leads for drug-like properties. It would be valuable to identify these less-druggable targets before incurring substantial expenditure and effort. Here we show that a model-based approach using basic biophysical principles yields good prediction of druggability based solely on the crystal structure of the target binding site. We quantitatively estimate the maximal affinity achievable by a drug-like molecule, and we show that these calculated values correlate with drug discovery outcomes. We experimentally test two predictions using high-throughput screening of a diverse compound collection. The collective results highlight the utility of our approach as well as strategies for tackling difficult targets.

Improvement in ADMET Prediction with Multitask Deep Featurization.

The absorption, distribution, metabolism, elimination, and toxicity (ADMET) properties of drug candidates are important for their efficacy and safety as therapeutics. Predicting ADMET properties has therefore been of great interest to the computational chemistry and medicinal chemistry communities in recent decades. Traditional cheminformatics approaches, using learners such as random forests and deep neural networks, leverage fingerprint feature representations of molecules. Here, we learn the features most relevant to each chemical task at hand by representing each molecule explicitly as a graph. By applying graph convolutions to this explicit molecular representation, we achieve, to our knowledge, unprecedented accuracy in prediction of ADMET properties. By challenging our methodology with rigorous cross-validation procedures and prognostic analyses, we show that deep featurization better enables molecular predictors to not only interpolate but also extrapolate to new regions of chemical space.