Context Aware Data-Driven Retrosynthetic Analysis.

Modern drug discovery is an iterative process relying on hypothesis generation through exploitation of available data and hypothesis testing that produces informative results necessary for subsequent rounds of exploration. In this setting, hypothesis generation consists of designing chemical structures likely to meet the pharmaceutically relevant objectives of the discovery project pursued while hypothesis testing involves the compound synthesis and biological assays to query the hypothesis. While much attention has been placed on effective compound design, it is often the case that hypothesis generation efforts lead to novel chemical structure designs with no established chemical synthesis route. We introduce a chemical context aware data-driven method built upon millions of available reactions, with attractive run-time characteristics, to recommend synthetic routes matching a precedent-derived template. Coupled with modern automated synthesis platforms and available building block collections, the method enables drug discovery researchers to identify easy to interpret and implement routes for target compounds. Results of this in-house computer-aided synthesis platform termed ChemoPrint are presented here demonstrating how such tools can bridge chemical synthesis knowledge with synthetic resources and facilitate hypothesis testing, thereby reducing the time required to complete an idea-to-data drug discovery cycle.

High-throughput hydrogen bond strength calculation and its applications in optimizing drug ADME properties.

AIM: Modifying the molecule's intrinsic hydrogen bond strength (HBS) is a useful approach in optimizing its permeability and P-glycoprotein (P-gp) efflux. Quantum mechanics (QM) based computation has been utilized to estimate the molecular intrinsic HBS. Despite its usefulness, the computation is time consuming for a large set of molecules. METHODOLOGY/RESULTS: We introduced a fragment-based high-throughput HBS calculation method and validated it with internal and external datasets. Examples have been presented where the P-gp efflux and permeability can be optimized by modulating calculated HBS. CONCLUSION: The results will enable medicinal chemists to calculate HBS in a high-throughput manner while optimizing permeability and P-gp efflux. This will further improve the efficiency of balancing multiple properties during drug discovery process.

A retrosynthetic analysis algorithm implementation.

The need for synthetic route design arises frequently in discovery-oriented chemistry organizations. While traditionally finding solutions to this problem has been the domain of human experts, several computational approaches, aided by the algorithmic advances and the availability of large reaction collections, have recently been reported. Herein we present our own implementation of a retrosynthetic analysis method and demonstrate its capabilities in an attempt to identify synthetic routes for a collection of approved drugs. Our results indicate that the method, leveraging on reaction transformation rules learned from a large patent reaction dataset, can identify multiple theoretically feasible synthetic routes and, thus, support research chemist everyday efforts.

Investigating the Behavior of Published PAINS Alerts Using a Pharmaceutical Company Data Set.

Biochemical assay interference is becoming increasingly recognized as a significant waste of resource in drug discovery, both in industry and academia. A seminal publication from Baell and Holloway raised the awareness of this issue, and they published a set of alerts to identify what they described as PAINS (pan-assay interference compounds). These alerts have been taken up by drug discovery groups, even though the original paper had a somewhat limited data set. Here, we have taken Lilly's far larger internal data set to assess the PAINS alerts on four criteria: promiscuity (over six assay formats including AlphaScreen), compound stability, cytotoxicity, and presence of a high Hill slope as a surrogate for non-1:1 protein-ligand binding. It was found that only three of the alerts show pan-assay promiscuity, and the alerts appear to encode primarily AlphaScreen promiscuous molecules. Although not enriching for pan-assay promiscuity, many of the alerts do encode molecules that are unstable, show cytotoxicity, and increase the prevalence of high Hill slopes.

Drug wastage and costs to the healthcare system in the care of patients with non-small cell lung cancer in the United States.

BACKGROUND: Lung cancer is one of the most prevalent cancers in the US. This study was designed to evaluate the actual drug wastage and cost to the healthcare system using patient-level retrospective observational electronic medical record (EMR) data from a cohort of lung cancer patients in the US. METHODS: Data from the Flatiron Health advanced non-small cell lung cancer (NSCLC) cohort was used for this study. Drug administered amount (in mg) was used to determine an optimal set of available vial sizes to minimize waste. Drug wastage was defined as the difference between the drug amount in the optimal set of vials and the administered amount. Wholesale acquisition costs were used to value the cost of drugs, with and without vial sharing assumptions. The amount and cost of waste were quantified over the 2-year study period (January 2015-December 2016). RESULTS: There were 8,467 eligible patients included in this study, providing data from 103,826 unique drug administrations across multiple lines of therapy. Overall wastage was 4.37% of the total medication used to care for patients. While costs per administration were low, the total cost of wastage for the study population represented $16,630,112 across the 2-year study period. Assuming that vial sharing occurred at the site level slightly reduced waste to 3.74% (reducing costs to $15,953,212 over 2 years). CONCLUSIONS: Drug wastage is an important concern and has implications on healthcare costs in NSCLC. Evaluation of these real-world data suggest that pharmacists and physicians are able to reduce drug wastage by optimizing vial combinations and sharing vials among patients. Even small amounts of reduction in wastage could be useful in reducing healthcare costs in the US; however, caution is needed with drug rounding efforts to ensure patients do not receive a sub-optimal dose of medication.

Dopamine Transporter Neuroimaging as an Enrichment Biomarker in Early Parkinson's Disease Clinical Trials: A Disease Progression Modeling Analysis.

Given the recognition that disease-modifying therapies should focus on earlier Parkinson's disease stages, trial enrollment based purely on clinical criteria poses significant challenges. The goal herein was to determine the utility of dopamine transporter neuroimaging as an enrichment biomarker in early motor Parkinson's disease clinical trials. Patient-level longitudinal data of 672 subjects with early-stage Parkinson's disease in the Parkinson's Progression Markers Initiative (PPMI) observational study and the Parkinson Research Examination of CEP-1347 Trial (PRECEPT) clinical trial were utilized in a linear mixed-effects model analysis. The rate of worsening in the motor scores between subjects with or without a scan without evidence of dopamine transporter deficit was different both statistically and clinically. The average difference in the change from baseline of motor scores at 24 months between biomarker statuses was -3.16 (90% confidence interval [CI] = -0.96 to -5.42) points. Dopamine transporter imaging could identify subjects with a steeper worsening of the motor scores, allowing trial enrichment and 24% reduction of sample size.

Minimization of olaratumab drug waste using real-world data.

PURPOSE: Results of a study in which population-based body weight and body surface area (BSA) data were used for vial size optimization to reduce drug waste associated with administration of the i.v. anticancer agent olaratumab are reported. METHODS: A retrospective observational study was conducted to determine weight and BSA distributions in a large sample of U.S. oncology patients using data from a large electronic medical record database. Body weight and BSA values at the time of initial systemic anticancer therapy were used to compute olaratumab dose requirements in a cohort of patients with soft tissue sarcoma; those data were analyzed to derive estimates of drug waste likely to result from the use of various proposed olaratumab vial sizes in combination with an existing 500-mg size. Weight and BSA distributions were calculated for additional cohorts of patients with 7 other cancer types. RESULTS: Median weight values in men (n = 1,179) and women (n = 1,078) with soft tissue sarcoma were 82.55 kg (interquartile range [IQR], 72.58-95.53 kg) and 68.69 kg (IQR, 58.51-84.28 kg), respectively. Modeling of olaratumab dosing scenarios indicated that use of the 500-mg vial only would result in estimated average drug waste of 234 mg per patient per administration; analysis of various potential vial size combinations showed that waste could be reduced by 87.6% with the addition of a 190-mg vial size. CONCLUSION: Analysis of real-world patient weight and BSA data allowed olaratumab vial size optimization to enable maximal dosing flexibility with minimal drug waste.

QSAR Model of Unbound Brain-to-Plasma Partition Coefficient, Kp,uu,brain: Incorporating P-glycoprotein Efflux as a Variable.

We report development and prospective validation of a QSAR model of the unbound brain-to-plasma partition coefficient, Kp,uu,brain, based on the in-house data set of ∼1000 compounds. We discuss effects of experimental variability, explore the applicability of both regression and classification approaches, and evaluate a novel, model-within-a-model approach of including P-glycoprotein efflux prediction as an additional variable. When tested on an independent test set of 91 internal compounds, incorporation of P-glycoprotein efflux information significantly improves the model performance resulting in an R2 of 0.53, RMSE of 0.57, Spearman's Rho correlation coefficient of 0.73, and qualitative prediction accuracy of 0.8 (kappa = 0.6). In addition to improving the performance, one of the key advantages of this approach is the larger chemical space coverage provided indirectly through incorporation of the in vitro, higher throughput data set that is 4 times larger than the in vivo data set.

The Proximal Lilly Collection: Mapping, Exploring and Exploiting Feasible Chemical Space.

Venturing into the immensity of the small molecule universe to identify novel chemical structure is a much discussed objective of many methods proposed by the chemoinformatics community. To this end, numerous approaches using techniques from the fields of computational de novo design, virtual screening and reaction informatics, among others, have been proposed. Although in principle this objective is commendable, in practice there are several obstacles to useful exploitation of the chemical space. Prime among them are the sheer number of theoretically feasible compounds and the practical concern regarding the synthesizability of the chemical structures conceived using in silico methods. We present the Proximal Lilly Collection initiative implemented at Eli Lilly and Co. with the aims to (i) define the chemical space of small, drug-like compounds that could be synthesized using in-house resources and (ii) facilitate access to compounds in this large space for the purposes of ongoing drug discovery efforts. The implementation of PLC relies on coupling access to available synthetic knowledge and resources with chemo/reaction informatics techniques and tools developed for this purpose. We describe in detail the computational framework supporting this initiative and elaborate on the characteristics of the PLC virtual collection of compounds. As an example of the opportunities provided to drug discovery researchers by easy access to a large, realistically feasible virtual collection such as the PLC, we describe a recent application of the technology that led to the discovery of selective kinase inhibitors.

Selectivity data: assessment, predictions, concordance, and implications.

Could high-quality in silico predictions in drug discovery eventually replace part or most of experimental testing? To evaluate the agreement of selectivity data from different experimental or predictive sources, we introduce the new metric concordance minimum significant ratio (cMSR). Empowered by cMSR, we find the overall level of agreement between predicted and experimental data to be comparable to that found between experimental results from different sources. However, for molecules that are either highly selective or potent, the concordance between different experimental sources is significantly higher than the concordance between experimental and predicted values. We also show that computational models built from one data set are less predictive for other data sources and highlight the importance of bias correction for assessing selectivity data. Finally, we show that small-molecule target space relationships derived from different data sources and predictive models share overall similarity but can significantly differ in details.

Integration of in silico and in vitro tools for scaffold optimization during drug discovery: predicting P-glycoprotein efflux.

In silico tools are regularly utilized for designing and prioritizing compounds to address challenges related to drug metabolism and pharmacokinetics (DMPK) during the process of drug discovery. P-Glycoprotein (P-gp) is a member of the ATP-binding cassette (ABC) transporters with broad substrate specificity that plays a significant role in absorption and distribution of drugs that are P-gp substrates. As a result, screening for P-gp transport has now become routine in the drug discovery process. Typically, bidirectional permeability assays are employed to assess in vitro P-gp efflux. In this article, we use P-gp as an example to illustrate a well-validated methodology to effectively integrate in silico and in vitro tools to identify and resolve key barriers during the early stages of drug discovery. A detailed account of development and application of in silico tools such as simple guidelines based on physicochemical properties and more complex quantitative structure-activity relationship (QSAR) models is provided. The tools were developed based on structurally diverse data for more than 2000 compounds generated using a robust P-gp substrate assay over the past several years. Analysis of physicochemical properties revealed a significantly lower proportion (<10%) of P-gp substrates among the compounds with topological polar surface area (TPSA) <60 Å(2) and the most basic cpKa <8. In contrast, this proportion of substrates was greater than 75% for compounds with TPSA >60 Å(2) and the most basic cpKa >8. Among the various QSAR models evaluated to predict P-gp efflux, the Bagging model provided optimum prediction performance for prospective validation based on chronological test sets. Four sequential versions of the model were built with increasing numbers of compounds to train the models as new data became available. Except for the first version with the smallest training set, the QSAR models exhibited robust prediction profiles with positive prediction values (PPV) and negative prediction values (NPV) exceeding 80%. The QSAR model demonstrated better concordance with the manual P-gp substrate assay than an automated P-gp substrate screen. The in silico and the in vitro tools have been effectively integrated during early stages of drug discovery to resolve P-gp-related challenges exemplified by several case studies. Key learning based on our experience with P-gp can be widely applicable across other DMPK-related challenges.

Rules for identifying potentially reactive or promiscuous compounds.

This article describes a set of 275 rules, developed over an 18-year period, used to identify compounds that may interfere with biological assays, allowing their removal from screening sets. Reasons for rejection include reactivity (e.g., acyl halides), interference with assay measurements (fluorescence, absorbance, quenching), activities that damage proteins (oxidizers, detergents), instability (e.g., latent aldehydes), and lack of druggability (e.g., compounds lacking both oxygen and nitrogen). The structural queries were profiled for frequency of occurrence in druglike and nondruglike compound sets and were extensively reviewed by a panel of experienced medicinal chemists. As a means of profiling the rules and as a filter in its own right, an index of biological promiscuity was developed. The 584 gene targets with screening data at Lilly were assigned to 17 subfamilies, and the number of subfamilies at which a compound was active was used as a promiscuity index. For certain compounds, promiscuous activity disappeared after sample repurification, indicating interference from occult contaminants. Because this type of interference is not amenable to substructure search, a "nuisance list" was developed to flag interfering compounds that passed the substructure rules.

Accelerating two algorithms for large-scale compound selection on GPUs.

Compound selection procedures based on molecular similarity and diversity are widely used in drug discovery. Current algorithms are often time consuming when applied to very large compound sets. This paper describes the acceleration of two selection algorithms (the leader and the spread algorithms) on graphical processing units (GPUs). We first parallelized the molecular similarity calculation based on Daylight fingerprints and the Tanimoto index and then implemented the two algorithms on GPU hardware using the open source Thrust library. Experiments show that the GPU leader algorithm is 73-120 times faster than the CPU version, and the GPU spread algorithm is 78-143 times faster than the CPU version.

Structure-guided expansion of kinase fragment libraries driven by support vector machine models.

This work outlines a new de novo design process for the creation of novel kinase inhibitor libraries. It relies on a profiling paradigm that generates a substantial amount of kinase inhibitor data from which highly predictive QSAR models can be constructed. In addition, a broad diversity of X-ray structure information is needed for binding mode prediction. This is important for scaffold and substituent site selection. Borrowing from FBDD, the process involves fragmentation of known actives, proposition of binding mode hypotheses for the fragments, and model-driven recombination using a pharmacophore derived from known kinase inhibitor structures. The support vector machine method, using Merck atom pair derived fingerprint descriptors, was used to build models from activity from 6 kinase assays. These models were qualified prospectively by selecting and testing compounds from the internal compound collection. Overall hit and enrichment rates of 82% and 2.5%, respectively, qualified the models for use in library design. Using the process, 7 novel libraries were designed, synthesized and tested against these same 6 kinases. The results showed excellent results, yielding a 92% hit rate for the 179 compounds that made up the 7 libraries. The results of one library designed to include known literature compounds, as well as an analysis of overall substituent frequency, are discussed.

GPU accelerated support vector machines for mining high-throughput screening data.

Support Vector Machine (SVM), one of the most promising tools in chemical informatics, is time-consuming for mining large high-throughput screening (HTS) data sets. Here, we describe a parallelization of SVM-light algorithm on a graphic processor unit (GPU), using molecular fingerprints as descriptors and the Tanimoto index as kernel function. Comparison experiments based on six PubChem Bioassay data sets show that the GPU version is 43-104x faster than SVM-light for building classification models and 112-212x over SVM-light for building regression models.

Rationalizing lead optimization by associating quantitative relevance with molecular structure modification.

Historically, one of the characteristic activities of the medicinal chemist has been the iterative improvement of lead compounds until a suitable therapeutic entity is achieved. Often referred to as lead optimization, this process typically takes the form of minor structural modifications to an existing lead in an attempt to ameliorate deleterious attributes while simultaneously trying to maintain or improve desirable properties. The cumulative effect of this exercise performed over the course of several decades of pharmaceutical research by thousands of trained researchers has resulted in large collections of pharmaceutically relevant chemical structures. As far as the authors are aware, this work represents the first attempt to use that data to define a framework to quantifiably catalogue and summate this information into a medicinal chemistry expert system. A method is proposed that first comprehensively mines a compendium of chemical structures compiling the structural modifications, abridges them to rectify artificially inflated support levels, and then performs an association rule mining experiment to ascribe relative confidences to each transformation. The result is a catalogue of statistically relevant structural modifications that can potentially be used in a number of pharmaceutical applications.

Evaluating the performances of quantitative structure-retention relationship models with different sets of molecular descriptors and databases for high-performance liquid chromatography predictions.

Quantitative structure-retention relationship (QSRR) models were studied for two databases: one with 151 compounds and the other with 1719 compounds. In both cases, the three modeling methods employed (multiple linear regression, partial least squares, and random forests) provided similar prediction results with regard to root-mean-square error of prediction. The reversed-phase retention related seven molecular descriptors provided better models for the smaller dataset, while the use of over 2000 molecular descriptors generated better models for the larger dataset. The QSRR models were then validated with a mixture of an active pharmaceutical ingredient and its four process/degradation impurities. Finally, classification of compounds based on similar logD profiles before QSRR modeling improved chromatographic predictability for the models used. The results showed that database composition had a desirable effect on prediction accuracy for certain input molecules.

Indirect similarity based methods for effective scaffold-hopping in chemical compounds.

Methods that can screen large databases to retrieve a structurally diverse set of compounds with desirable bioactivity properties are critical in the drug discovery and development process. This paper presents a set of such methods that are designed to find compounds that are structurally different to a certain query compound while retaining its bioactivity properties (scaffold hops). These methods utilize various indirect ways of measuring the similarity between the query and a compound that take into account additional information beyond their structure-based similarities. The set of techniques that are presented capture these indirect similarities using approaches based on analyzing the similarity network formed by the query and the database compounds. Experimental evaluation shows that most of these methods substantially outperform previously developed approaches both in terms of their ability to identify structurally diverse active compounds as well as active compounds in general.

Method for effective virtual screening and scaffold-hopping in chemical compounds.

Methods that can screen large databases to retrieve a structurally diverse set of compounds with desirable bioactivity properties are critical in the drug discovery and development process. This paper presents a set of such methods, which are designed to find compounds that are structurally different to a certain query compound while retaining its bioactivity properties (scaffold hops). These methods utilize various indirect ways of measuring the similarity between the query and a compound that take into account additional information beyond their structure-based similarities. Two sets of techniques are presented that capture these indirect similarities using approaches based on automatic relevance feedback and on analyzing the similarity network formed by the query and the database compounds. Experimental evaluation shows that many of these methods substantially outperform previously developed approaches both in terms of their ability to identify structurally diverse active compounds as well as active compounds in general.