Convolutional Neural Networks Enable Highly Accurate and Automated Subvisible Particulate Classification of Biopharmaceuticals.

Quantification of subvisible particles, which are generally defined as those ranging in size from 2 to 100 µm, is important as critical characteristics for biopharmaceutical formulation development. Micro Flow Imaging (MFI) provides quantifiable morphological parameters to study both the size and type of subvisible particles, including proteinaceous particles as well as non-proteinaceous features incl. silicone oil droplets, air bubble droplets, etc., thus enabling quantitative and categorical particle attribute reporting for quality control. However, limitations in routine MFI image analysis can hinder accurate subvisible particle classification. In this work, we custom-built a subvisible particle-aware Convolutional Neural Network, SVNet, which has a very small computational footprint, and achieves comparable performance to prior state-of-art image classification models. SVNet significantly improves upon current standard operating procedures for subvisible particulate assessments as confirmed by thorough real-world validation studies.

Discovery of non-boronic acid Arginase 1 inhibitors through virtual screening and biophysical methods.

Inhibiting Arginase 1 (ARG1), a metalloenzyme that hydrolyzes l-arginine in the urea cycle, has been demonstrated as a promising therapeutic avenue in immuno-oncology through the restoration of suppressed immune response in several types of cancers. Most of the currently reported small molecule inhibitors are boronic acid based. Herein, we report the discovery of non-boronic acid ARG1 inhibitors through virtual screening. Biophysical and biochemical methods were used to experimentally profile the hits while X-ray crystallography confirmed a class of trisubstituted pyrrolidine derivatives as optimizable alternatives for the development of novel classes of immuno-oncology agents targeting this enzyme.

Selection of an Optimal In Vitro Model to Assess P-gp Inhibition: Comparison of Vesicular and Bidirectional Transcellular Transport Inhibition Assays.

The multidrug resistance protein 1 (MDR1) P-glycoprotein (P-gp) is a clinically important transporter. In vitro P-gp inhibition assays have been routinely conducted to predict the potential for clinical drug-drug interactions (DDIs) mediated by P-gp. However, high interlaboratory and intersystem variability of P-gp IC50 data limits accurate prediction of DDIs using static models and decision criteria recommended by regulatory agencies. In this study, we calibrated two in vitro P-gp inhibition models: vesicular uptake of N-methyl-quinidine (NMQ) in MDR1 vesicles and bidirectional transport (BDT) of digoxin in Lilly Laboratories Cell Porcine Kidney 1 cells overexpressing MDR1 (LLC-MDR1) using a total of 48 P-gp inhibitor and noninhibitor drugs and digoxin DDI data from 70 clinical studies. Refined thresholds were derived using receiver operating characteristic analysis, and their predictive performance was compared with the decision frameworks proposed by regulatory agencies and selected reference. Furthermore, the impact of various IC50 calculation methods and nonspecific binding of drugs on DDI prediction was evaluated. Our studies suggest that the concentration of inhibitor based on highest approved dose dissolved in 250 ml divided by IC50(I2/IC50) is sufficient to predict P-gp related intestinal DDIs. IC50 obtained from vesicular inhibition assay with a refined threshold of I2/IC50 ≥ 25.9 provides comparable predictive power over those measured by net secretory flux and efflux ratio in LLC-MDR1 cells. We therefore recommend vesicular P-gp inhibition as our preferred method given its simplicity, lower variability, higher assay throughput, and more direct estimation of in vitro kinetic parameters, rather than BDT assay. SIGNIFICANCE STATEMENT: This study has conducted comprehensive calibration of two in vitro P-gp inhibition models: uptake in MDR1 vesicles and bidirectional transport in LLC-MDR1 cell monolayers to predict DDIs. This study suggests that IC50s obtained from vesicular inhibition with a refined threshold of I2/IC50 ≥ 25.9 provide comparable predictive power over those in LLC-MDR1 cells. Therefore, vesicular P-gp inhibition is recommended as the preferred method given its simplicity, lower variability, higher assay throughput, and more direct estimation of in vitro kinetic parameters.

Publisher Correction: A mass spectrometry-based proteome map of drug action in lung cancer cell lines.

An amendment to this paper has been published and can be accessed via a link at the top of the paper.

A mass spectrometry-based proteome map of drug action in lung cancer cell lines.

Mass spectrometry-based discovery proteomics is an essential tool for the proximal readout of cellular drug action. Here, we apply a robust proteomic workflow to rapidly profile the proteomes of five lung cancer cell lines in response to more than 50 drugs. Integration of millions of quantitative protein-drug associations substantially improved the mechanism of action (MoA) deconvolution of single compounds. For example, MoA specificity increased after removal of proteins that frequently responded to drugs and the aggregation of proteome changes across cell lines resolved compound effects on proteostasis. We leveraged these findings to demonstrate efficient target identification of chemical protein degraders. Aggregating drug response across cell lines also revealed that one-quarter of compounds modulated the abundance of one of their known protein targets. Finally, the proteomic data led us to discover that inhibition of mitochondrial function is an off-target mechanism of the MAP2K1/2 inhibitor PD184352 and that the ALK inhibitor ceritinib modulates autophagy.

Nearest Neighbor Gaussian Process for Quantitative Structure-Activity Relationships.

While Gaussian process models are typically restricted to smaller data sets, we propose a variation which extends its applicability to the larger data sets common in the industrial drug discovery space, making it relatively novel in the quantitative structure-activity relationship (QSAR) field. By incorporating locality-sensitive hashing for fast nearest neighbor searches, the nearest neighbor Gaussian process model makes predictions with time complexity that is sub-linear with the sample size. The model can be efficiently built, permitting rapid updates to prevent degradation as new data is collected. Given its small number of hyperparameters, it is robust against overfitting and generalizes about as well as other common QSAR models. Like the usual Gaussian process model, it natively produces principled and well-calibrated uncertainty estimates on its predictions. We compare this new model with implementations of random forest, light gradient boosting, and k-nearest neighbors to highlight these promising advantages. The code for the nearest neighbor Gaussian process is available at https://github.com/Merck/nngp.

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.

Deep Dive into Machine Learning Models for Protein Engineering.

Protein redesign and engineering has become an important task in pharmaceutical research and development. Recent advances in technology have enabled efficient protein redesign by mimicking natural evolutionary mutation, selection, and amplification steps in the laboratory environment. For any given protein, the number of possible mutations is astronomical. It is impractical to synthesize all sequences or even to investigate all functionally interesting variants. Recently, there has been an increased interest in using machine learning to assist protein redesign, since prediction models can be used to virtually screen a large number of novel sequences. However, many state-of-the-art machine learning models, especially deep learning models, have not been extensively explored. Moreover, only a small selection of protein sequence descriptors has been considered. In this work, the performance of prediction models built using an array of machine learning methods and protein descriptor types, including two novel, single amino acid descriptors and one structure-based three-dimensional descriptor, is benchmarked. The predictions were evaluated on a diverse collection of public and proprietary data sets, using a variety of evaluation metrics. The results of this comparison suggest that Convolution Neural Network models built with amino acid property descriptors are the most widely applicable to the types of protein redesign problems faced in the pharmaceutical industry.

Can Galactose Be Converted to Glucose in HepG2 Cells? Improving the in Vitro Mitochondrial Toxicity Assay for the Assessment of Drug Induced Liver Injury.

Human hepatocellular carcinoma cells, HepG2, are often used for drug mediated mitochondrial toxicity assessments. Glucose in HepG2 culture media is replaced by galactose to reveal drug-induced mitochondrial toxicity as a marked shift of drug IC50 values for the reduction of cellular ATP. It has been postulated that galactose sensitizes HepG2 mitochondria by the additional ATP consumption demand in the Leloir pathway. However, our NMR metabolomics analysis of HepG2 cells and culture media showed very limited galactose metabolism. To clarify the role of galactose in HepG2 cellular metabolism, U-13C6-galactose or U-13C6-glucose was added to HepG2 culture media to help specifically track the metabolism of those two sugars. Conversion to U-13C3-lactate was hardly detected when HepG2 cells were incubated with U-13C6-galactose, while an abundance of U-13C3-lactate was produced when HepG2 cells were incubated with U-13C6-glucose. In the absence of glucose, HepG2 cells increased glutamine consumption as a bioenergetics source. The requirement of additional glutamine almost matched the amount of glucose needed to maintain a similar level of cellular ATP in HepG2 cells. This improved understanding of galactose and glutamine metabolism in HepG2 cells helped optimize the ATP-based mitochondrial toxicity assay. The modified assay showed 96% sensitivity and 97% specificity in correctly discriminating compounds known to cause mitochondrial toxicity from those with prior evidence of not being mitochondrial toxicants. The greatest significance of the modified assay was its improved sensitivity in detecting the inhibition of mitochondrial fatty acid ß-oxidation (FAO) when glutamine was withheld. Use of this improved assay for an empirical prediction of the likely contribution of mitochondrial toxicity to human DILI (drug induced liver injury) was attempted. According to testing of 65 DILI positive compounds representing numerous mechanisms of DILI together with 55 DILI negative compounds, the overall prediction of mitochondrial mechanism-related DILI showed 25% sensitivity and 95% specificity.

Building Quantitative Structure-Activity Relationship Models Using Bayesian Additive Regression Trees.

Quantitative structure-activity relationship (QSAR) is a very commonly used technique for predicting the biological activity of a molecule using information contained in the molecular descriptors. The large number of compounds and descriptors and the sparseness of descriptors pose important challenges to traditional statistical methods and machine learning (ML) algorithms (such as random forest (RF)) used in this field. Recently, Bayesian Additive Regression Trees (BART), a flexible Bayesian nonparametric regression approach, has been demonstrated to be competitive with widely used ML approaches. Instead of only focusing on accurate point estimation, BART is formulated entirely in a hierarchical Bayesian modeling framework, allowing one to also quantify uncertainties and hence to provide both point and interval estimation for a variety of quantities of interest. We studied BART as a model builder for QSAR and demonstrated that the approach tends to have predictive performance comparable to RF. More importantly, we investigated BART's natural capability to analyze truncated (or qualified) data, generate interval estimates for molecular activities as well as descriptor importance, and conduct model diagnosis, which could not be easily handled through other approaches.

Demystifying Multitask Deep Neural Networks for Quantitative Structure-Activity Relationships.

Deep neural networks (DNNs) are complex computational models that have found great success in many artificial intelligence applications, such as computer vision1,2 and natural language processing.3,4 In the past four years, DNNs have also generated promising results for quantitative structure-activity relationship (QSAR) tasks.5,6 Previous work showed that DNNs can routinely make better predictions than traditional methods, such as random forests, on a diverse collection of QSAR data sets. It was also found that multitask DNN models-those trained on and predicting multiple QSAR properties simultaneously-outperform DNNs trained separately on the individual data sets in many, but not all, tasks. To date there has been no satisfactory explanation of why the QSAR of one task embedded in a multitask DNN can borrow information from other unrelated QSAR tasks. Thus, using multitask DNNs in a way that consistently provides a predictive advantage becomes a challenge. In this work, we explored why multitask DNNs make a difference in predictive performance. Our results show that during prediction a multitask DNN does borrow "signal" from molecules with similar structures in the training sets of the other tasks. However, whether this borrowing leads to better or worse predictive performance depends on whether the activities are correlated. On the basis of this, we have developed a strategy to use multitask DNNs that incorporate prior domain knowledge to select training sets with correlated activities, and we demonstrate its effectiveness on several examples.

Dose-dependent effects of siRNA-mediated inhibition of SCAP on PCSK9, LDLR, and plasma lipids in mouse and rhesus monkey.

SREBP cleavage-activating protein (SCAP) is a key protein in the regulation of lipid metabolism and a potential target for treatment of dyslipidemia. SCAP is required for activation of the transcription factors SREBP-1 and -2. SREBPs regulate the expression of genes involved in fatty acid and cholesterol biosynthesis, and LDL-C clearance through the regulation of LDL receptor (LDLR) and PCSK9 expression. To further test the potential of SCAP as a novel target for treatment of dyslipidemia, we used siRNAs to inhibit hepatic SCAP expression and assess the effect on PCSK9, LDLR, and lipids in mice and rhesus monkeys. In mice, robust liver Scap mRNA knockdown (KD) was achieved, accompanied by dose-dependent reduction in SREBP-regulated gene expression, de novo lipogenesis, and plasma PCSK9 and lipids. In rhesus monkeys, over 90% SCAP mRNA KD was achieved resulting in approximately 75, 50, and 50% reduction of plasma PCSK9, TG, and LDL-C, respectively. Inhibition of SCAP function was demonstrated by reduced expression of SREBP-regulated genes and de novo lipogenesis. In conclusion, siRNA-mediated inhibition of SCAP resulted in a significant reduction in circulating PCSK9 and LDL-C in rodent and primate models supporting SCAP as a novel target for the treatment of dyslipidemia.

Extreme Gradient Boosting as a Method for Quantitative Structure-Activity Relationships.

In the pharmaceutical industry it is common to generate many QSAR models from training sets containing a large number of molecules and a large number of descriptors. The best QSAR methods are those that can generate the most accurate predictions but that are not overly expensive computationally. In this paper we compare eXtreme Gradient Boosting (XGBoost) to random forest and single-task deep neural nets on 30 in-house data sets. While XGBoost has many adjustable parameters, we can define a set of standard parameters at which XGBoost makes predictions, on the average, better than those of random forest and almost as good as those of deep neural nets. The biggest strength of XGBoost is its speed. Whereas efficient use of random forest requires generating each tree in parallel on a cluster, and deep neural nets are usually run on GPUs, XGBoost can be run on a single CPU in less than a third of the wall-clock time of either of the other methods.

Evaluation of cynomolgus monkeys for the identification of endogenous biomarkers for hepatic transporter inhibition and as a translatable model to predict pharmacokinetic interactions with statins in humans.

Inhibition of hepatic transporters such as organic anion transporting polypeptides (OATPs) 1B can cause drug-drug interactions (DDIs). Determining the impact of perpetrator drugs on the plasma exposure of endogenous substrates for OATP1B could be valuable to assess the risk for DDIs early in drug development. As OATP1B orthologs are well conserved between human and monkey, we assessed in cynomolgus monkeys the endogenous OATP1B substrates that are potentially suitable to assess DDI risk in humans. The effect of rifampin (RIF), a potent inhibitor for OATP1B, on plasma exposure of endogenous substrates of hepatic transporters was measured. From the 18 biomarkers tested, RIF (18 mg/kg, oral) caused significant elevation of plasma unconjugated and conjugated bilirubin, which may be attributed to inhibition of cOATP1B1 and cOATP1B3 based on in vitro to in vivo extrapolation analysis. To further evaluate whether cynomolgus monkeys are a suitable translational model to study OATP1B-mediated DDIs, we determined the inhibitory effect of RIF on in vitro transport and pharmacokinetics of rosuvastatin (RSV) and atorvastatin (ATV). RIF strongly inhibited the uptake of RSV and ATV by cOATP1B1 and cOATP1B3 in vitro. In agreement with clinical observations, RIF (18 mg/kg, oral) significantly decreased plasma clearance and increased the area under the plasma concentration curve (AUC) of intravenously administered RSV by 2.8- and 2.7-fold, and increased the AUC and maximum plasma concentration of orally administered RSV by 6- and 10.3-fold, respectively. In contrast to clinical findings, RIF did not significantly increase plasma exposure of either intravenous or orally administered ATV, indicating species differences in the rate-limiting elimination pathways.

Deep neural nets as a method for quantitative structure-activity relationships.

Neural networks were widely used for quantitative structure-activity relationships (QSAR) in the 1990s. Because of various practical issues (e.g., slow on large problems, difficult to train, prone to overfitting, etc.), they were superseded by more robust methods like support vector machine (SVM) and random forest (RF), which arose in the early 2000s. The last 10 years has witnessed a revival of neural networks in the machine learning community thanks to new methods for preventing overfitting, more efficient training algorithms, and advancements in computer hardware. In particular, deep neural nets (DNNs), i.e. neural nets with more than one hidden layer, have found great successes in many applications, such as computer vision and natural language processing. Here we show that DNNs can routinely make better prospective predictions than RF on a set of large diverse QSAR data sets that are taken from Merck's drug discovery effort. The number of adjustable parameters needed for DNNs is fairly large, but our results show that it is not necessary to optimize them for individual data sets, and a single set of recommended parameters can achieve better performance than RF for most of the data sets we studied. The usefulness of the parameters is demonstrated on additional data sets not used in the calibration. Although training DNNs is still computationally intensive, using graphical processing units (GPUs) can make this issue manageable.

Development and Evaluation of Conformal Prediction Methods for Quantitative Structure-Activity Relationship.

The quantitative structure-activity relationship (QSAR) regression model is a commonly used technique for predicting the biological activities of compounds using their molecular descriptors. Besides accurate activity estimation, obtaining a prediction uncertainty metric like a prediction interval is highly desirable. Quantifying prediction uncertainty is an active research area in statistical and machine learning (ML), but the implementation for QSAR remains challenging. However, most ML algorithms with high predictive performance require add-on companions for estimating the uncertainty of their prediction. Conformal prediction (CP) is a promising approach as its main components are agnostic to the prediction modes, and it produces valid prediction intervals under weak assumptions on the data distribution. We proposed computationally efficient CP algorithms tailored to the most widely used ML models, including random forests, deep neural networks, and gradient boosting. The algorithms use a novel approach to the derivation of nonconformity scores from the estimates of prediction uncertainty generated by the ensembles of point predictions. The validity and efficiency of proposed algorithms are demonstrated on a diverse collection of QSAR data sets as well as simulation studies. The provided software implementing our algorithms can be used as stand-alone or easily incorporated into other ML software packages for QSAR modeling.

Dynamic sampling in autonomous process optimization.

Autonomous process optimization (APO) is a technology that has recently found utility in a multitude of process optimization challenges. In contrast to most APO examples in microflow reactor systems, we recently presented a system capable of optimization in high-throughput batch reactor systems. The drawback of APO in a high-throughput batch reactor system is the reliance on reaction sampling at a predetermined static timepoint rather than a dynamic endpoint. Static timepoint sampling can lead to the inconsistent capture of the process performance under each process parameter permutation. This is important because critical process behaviors such as rate acceleration accompanied by decomposition could be missed entirely. To address this drawback, we implemented a dynamic reaction endpoint determination strategy to capture the product purity once the process stream stabilized. We accomplished this through the incorporation of a real-time plateau detection algorithm into the APO workflow to measure and report the product purity at the dynamically determined reaction endpoint. We then applied this strategy to the autonomous optimization of a photobromination reaction towards the synthesis of a pharmaceutically relevant intermediate. In doing so, we not only uncovered process conditions to access the desired monohalogenation product in 85 UPLC area % purity with minimal decomposition risk, but also measured the effect of each parameter on the process performance. Our results highlight the advantage of incorporating dynamic sampling in APO workflows to drive optimization toward a stable and high-performing process.

Quantitative analysis of intact apolipoproteins in human HDL by top-down differential mass spectrometry.

Top-down mass spectrometry holds tremendous potential for the characterization and quantification of intact proteins, including individual protein isoforms and specific posttranslationally modified forms. This technique does not require antibody reagents and thus offers a rapid path for assay development with increased specificity based on the amino acid sequence. Top-down MS is efficient whereby intact protein mass measurement, purification by mass separation, dissociation, and measurement of product ions with ppm mass accuracy occurs on the seconds to minutes time scale. Moreover, as the analysis is based on the accurate measurement of an intact protein, top-down mass spectrometry opens a research paradigm to perform quantitative analysis of "unknown" proteins that differ in accurate mass. As a proof of concept, we have applied differential mass spectrometry (dMS) to the top-down analysis of apolipoproteins isolated from human HDL(3). The protein species at 9415.45 Da demonstrates an average fold change of 4.7 (p-value 0.017) and was identified as an O-glycosylated form of apolipoprotein C-III [NANA-(2 --> 3)-Gal-beta(1 --> 3)-GalNAc, +656.2037 Da], a protein associated with coronary artery disease. This work demonstrates the utility of top-down dMS for quantitative analysis of intact protein mixtures and holds potential for facilitating a better understanding of HDL biology and complex biological systems at the protein level.

High-Throughput Screening to Identify Small Molecules That Selectively Inhibit APOL1 Protein Level in Podocytes.

High-throughput phenotypic screening is a key driver for the identification of novel chemical matter in drug discovery for challenging targets, especially for those with an unclear mechanism of pathology. For toxic or gain-of-function proteins, small-molecule suppressors are a targeting/therapeutic strategy that has been successfully applied. As with other high-throughput screens, the screening strategy and proper assays are critical for successfully identifying selective suppressors of the target of interest. We executed a small-molecule suppressor screen to identify compounds that specifically reduce apolipoprotein L1 (APOL1) protein levels, a genetically validated target associated with increased risk of chronic kidney disease. To enable this study, we developed homogeneous time-resolved fluorescence (HTRF) assays to measure intracellular APOL1 and apolipoprotein L2 (APOL2) protein levels and miniaturized them to 1536-well format. The APOL1 HTRF assay served as the primary assay, and the APOL2 and a commercially available p53 HTRF assay were applied as counterscreens. Cell viability was also measured with CellTiter-Glo to assess the cytotoxicity of compounds. From a 310,000-compound screening library, we identified 1490 confirmed primary hits with 12 different profiles. One hundred fifty-three hits selectively reduced APOL1 in 786-O, a renal cell adenocarcinoma cell line. Thirty-one of these selective suppressors also reduced APOL1 levels in conditionally immortalized human podocytes. The activity and specificity of seven resynthesized compounds were validated in both 786-O and podocytes.