The Comparison of Machine Learning and Mechanistic In Vitro-In Vivo Extrapolation Models for the Prediction of Human Intrinsic Clearance.

Accurate prediction of human pharmacokinetics (PK) remains one of the key objectives of drug metabolism and PK (DMPK) scientists in drug discovery projects. This is typically performed by using in vitro-in vivo extrapolation (IVIVE) based on mechanistic PK models. In recent years, machine learning (ML), with its ability to harness patterns from previous outcomes to predict future events, has gained increased popularity in application to absorption, distribution, metabolism, and excretion (ADME) sciences. This study compares the performance of various ML and mechanistic models for the prediction of human IV clearance for a large (645) set of diverse compounds with literature human IV PK data, as well as measured relevant in vitro end points. ML models were built using multiple approaches for the descriptors: (1) calculated physical properties and structural descriptors based on chemical structure alone (classical QSAR/QSPR); (2) in vitro measured inputs only with no structure-based descriptors (ML IVIVE); and (3) in silico ML IVIVE using in silico model predictions for the in vitro inputs. For the mechanistic models, well-stirred and parallel-tube liver models were considered with and without the use of empirical scaling factors and with and without renal clearance. The best ML model for the prediction of in vivo human intrinsic clearance (CLint) was an in vitro ML IVIVE model using only six in vitro inputs with an average absolute fold error (AAFE) of 2.5. The best mechanistic model used the parallel-tube liver model, with empirical scaling factors resulting in an AAFE of 2.8. The corresponding mechanistic model with full in silico inputs achieved an AAFE of 3.3. These relative performances of the models were confirmed with the prediction of 16 Pfizer drug candidates that were not part of the original data set. Results show that ML IVIVE models are comparable to or superior to their best mechanistic counterparts. We also show that ML IVIVE models can be used to derive insights into factors for the improvement of mechanistic PK prediction.

Prediction of Fraction Unbound in Microsomal and Hepatocyte Incubations: A Comparison of Methods across Industry Datasets.

The fraction unbound in the incubation, fu,inc, is an important parameter to consider in the evaluation of intrinsic clearance measurements performed in vitro in hepatocytes or microsomes. Reliable estimates of fu,inc based on a compound's structure have the potential to positively impact the screening timelines in drug discovery. Previous works suggested that fu,inc is primarily driven by passive processes and can be described using physicochemical properties such as lipophilicity and the protonation state of the molecule. While models based on these principles proved predictive in relatively small datasets that included marketed drugs, their applicability domain has not been extensively explored. The work presented here from the in silico ADME discussion group (part of the International Consortium for Innovation through Quality in Pharmaceutical Development, the IQ consortium) describes the accuracy of these models in large proprietary datasets that include several thousand of compounds across chemical space. Overall, the models do well for compounds with low lipophilicity. In other words, the equations correctly predict that fu,inc is, in general, above 0.5 for compounds with a calculated logP of less than 3. When applied to lipophilic compounds, the models failed to produce quantitatively accurate predictions of fu,inc, with a high risk of underestimating binding properties. These models can, therefore, be used quantitatively for less lipophilic compounds. On the other hand, internal machine-learning models using a company's own proprietary dataset also predict compounds with higher lipophilicity reasonably well. Additionally, the data shown indicate that microsomal binding is, in general, a good proxy for hepatocyte binding.

In-Silico Extraction of Design Ideas Using MMPA-by-QSAR and its Application on ADME Endpoints.

Matched molecular pair analysis (MMPA) has emerged as a powerful approach to mine and extract tacit knowledge from measured databases of small molecules. Extracted knowledge from past experimentation can assist future lead optimization as an idea generation tool and, hence, reduce the number of design-synthesis-test cycles. While attractive and intuitive, MMPA still presents several limitations. Analyses of internal absorption, distribution, metabolism, and excretion (ADME) databases of measured compounds show that chemical transformations with 10 pairs or more represent less than 1% of the total transforms identified by MMPA. A great wealth of design ideas remains effectively untapped and underutilized as the lack of measured data hinders extraction of robust trends. In this study we report the use of a quantitative structure-activity relationship (QSAR) model augmented MMPA approach (MMPA-by-QSAR) to infer the overall effect of chemical transformations on two essential ADME endpoints-lipophilicity and metabolic clearance. First, QSAR models are employed to predict compound activities, and subsequently, MMPA is used to identify and to extract virtual trends. Results obtained from retrospective analyses showed the ability to predict magnitudes of change close to experimental ones for the majority of transforms from each ADME data set. In the case of the lipophilicity endpoint (SFLogD) 73.7%, 87.85%, and 99% of transforms were predicted within 0.1, 0.15, and 0.3 units of the actual change. In the case of the clearance endpoint (HLM) 67.2%, 82.3%, and 99.5% of transforms were predicted within 0.08, 0.11, and 0.3 log units, respectively. Prospective application of MMPA-by-QSAR on untested compounds identified several novel transforms not observed in our measured data sets. When MMPs from these transforms were screened in our internal assays, it was found that the correct directionality of change was predicted for all but one of the tested transforms, and the predicted magnitudes of change have varying errors between predicted and measured mean changes ranging from 0.01 to 0.24 units for SFLogD and from 0.0 to 0.38 log units for HLM. This proposed MMPA-by-QSAR modeling approach has the potential to allow exploration of infrequent transforms or even identify completely novel transforms where no measured MMP is available.

A multi-endpoint matched molecular pair (MMP) analysis of 6-membered heterocycles.

Aromatic rings, ubiquitous in pharmaceutical compounds, are often exchanged with another ring during the optimization process of drug discovery. Inevitably, the preferred ring system for one endpoint may prove detrimental to another, thus necessitating a holistic, multiple endpoint optimization approach for finding the ideal replacement. Accordingly, we conducted an extensive matched molecular pair (MMP) analysis of common 6-membered aromatic rings across 4 endpoints critical for drug discovery (logD lipophilicity, microsomal metabolism, P-gp efflux and passive permeability). We also investigated the effect of context by considering the connecting atom. Heat maps were created as a simple yet comprehensive way to view and analyze the vast amount of interrelated data. Paired difference statistical tests were used to identify transforms with changes that were significantly different from zero. We conclude that the heat maps of transforms provide a unique and powerful approach for multiparameter optimization.

Reversion of somatic mutations of the respiratory syncytial virus-specific human monoclonal antibody Fab19 reveal a direct relationship between association rate and neutralizing potency.

The role of affinity in determining neutralizing potency of mAbs directed against viruses is not well understood. We investigated the kinetic, structural, and functional advantage conferred by individual naturally occurring somatic mutations in the Ab H chain V region of Fab19, a well-described neutralizing human mAb directed to respiratory syncytial virus. Comparison of the affinity-matured Ab Fab19 with recombinant Fab19 Abs that were variants containing reverted amino acids from the inferred unmutated ancestor sequence revealed the molecular basis for affinity maturation of this Ab. Enhanced binding was achieved through mutations in the third H chain CDR (HCDR3) that conferred a markedly faster on-rate and a desirable increase in antiviral neutralizing activity. In contrast, most somatic mutations in the HCDR1 and HCDR2 regions did not significantly enhance Ag binding or antiviral activity. We observed a direct relationship between the measured association rate (Kon) for F protein and antiviral activity. Modeling studies of the structure of the Ag-Ab complex suggested the HCDR3 loop interacts with the antigenic site A surface loop of the respiratory syncytial virus F protein, previously shown to contain the epitope for this Ab by experimentation. These studies define a direct relationship of affinity and neutralizing activity for a viral glycoprotein-specific human mAb.

Neutralizing antibodies derived from the B cells of 1918 influenza pandemic survivors.

Investigation of the human antibody response to influenza virus infection has been largely limited to serology, with relatively little analysis at the molecular level. The 1918 H1N1 influenza virus pandemic was the most severe of the modern era. Recent work has recovered the gene sequences of this unusual strain, so that the 1918 pandemic virus could be reconstituted to display its unique virulence phenotypes. However, little is known about adaptive immunity to this virus. We took advantage of the 1918 virus sequencing and the resultant production of recombinant 1918 haemagglutinin (HA) protein antigen to characterize at the clonal level neutralizing antibodies induced by natural exposure of survivors to the 1918 pandemic virus. Here we show that of the 32 individuals tested that were born in or before 1915, each showed seroreactivity with the 1918 virus, nearly 90 years after the pandemic. Seven of the eight donor samples tested had circulating B cells that secreted antibodies that bound the 1918 HA. We isolated B cells from subjects and generated five monoclonal antibodies that showed potent neutralizing activity against 1918 virus from three separate donors. These antibodies also cross-reacted with the genetically similar HA of a 1930 swine H1N1 influenza strain, but did not cross-react with HAs of more contemporary human influenza viruses. The antibody genes had an unusually high degree of somatic mutation. The antibodies bound to the 1918 HA protein with high affinity, had exceptional virus-neutralizing potency and protected mice from lethal infection. Isolation of viruses that escaped inhibition suggested that the antibodies recognize classical antigenic sites on the HA surface. Thus, these studies demonstrate that survivors of the 1918 influenza pandemic possess highly functional, virus-neutralizing antibodies to this uniquely virulent virus, and that humans can sustain circulating B memory cells to viruses for many decades after exposure-well into the tenth decade of life.

Characterization and Applications of Permeabilized Hepatocytes in Drug Discovery.

Hepatocytes are one of the most physiologically relevant in vitro liver systems for human translation of clearance and drug-drug interactions (DDI). However, the cell membranes of hepatocytes can limit the entry of certain compounds into the cells for metabolism and DDI. Passive permeability through hepatocytes can be different in vitro and in vivo, which complicates the human translation. Permeabilized hepatocytes offer a useful tool to probe mechanistic understanding of permeability-limited metabolism and DDI. Incubation with saponin of 0.01% at 0.5 million cells/mL and 0.05% at 5 million cells/mL for 5 min at 37°C completely permeabilized the plasma membrane of hepatocytes, while leaving the membranes of subcellular organelles intact. Permeabilized hepatocytes maintained similar enzymatic activity as intact unpermeabilized hepatocytes and can be stored at -80°C for at least 7 months. This approach reduces costs by preserving leftover hepatocytes. The relatively low levels of saponin in permeabilized hepatocytes had no significant impact on the enzymatic activity. As the cytosolic contents leak out from permeabilized hepatocytes, cofactors need to be added to enable metabolic reactions. Cytosolic enzymes will no longer be present if the media are removed after cells are permeabilized. Hence permeabilized hepatocytes with and without media removal may potentially enable reaction phenotyping of cytosolic enzymes. Although permeabilized hepatocytes work similarly as human liver microsomes and S9 fractions experimentally requiring addition of cofactors, they behave more like hepatocytes maintaining enzymatic activities for over 4 h. Permeabilized hepatocytes are a great addition to the drug metabolism toolbox to provide mechanistic insights.

Interpretable, probability-based confidence metric for continuous quantitative structure-activity relationship models.

A great deal of research has gone into the development of robust confidence in prediction and applicability domain (AD) measures for quantitative structure-activity relationship (QSAR) models in recent years. Much of the attention has historically focused on structural similarity, which can be defined in many forms and flavors. A concept that is frequently overlooked in the realm of the QSAR applicability domain is how the local activity landscape plays a role in how accurate a prediction is or is not. In this work, we describe an approach that pairs information about both the chemical similarity and activity landscape of a test compound's neighborhood into a single calculated confidence value. We also present an approach for converting this value into an interpretable confidence metric that has a simple and informative meaning across data sets. The approach will be introduced to the reader in the context of models built upon four diverse literature data sets. The steps we will outline include the definition of similarity used to determine nearest neighbors (NN), how we incorporate the NN activity landscape with a similarity-weighted root-mean-square distance (wRMSD) value, and how that value is then calibrated to generate an intuitive confidence metric for prospective application. Finally, we will illustrate the prospective performance of the approach on five proprietary models whose predictions and confidence metrics have been tracked for more than a year.

In Vitro-In Vivo Extrapolation and Scaling Factors for Clearance of Human and Preclinical Species with Liver Microsomes and Hepatocytes.

In vitro-in vivo extrapolation ((IVIVE) and empirical scaling factors (SF) of human intrinsic clearance (CLint) were developed using one of the largest dataset of 455 compounds with data from human liver microsomes (HLM) and human hepatocytes (HHEP). For extended clearance classification system (ECCS) class 2/4 compounds, linear SFs (SFlin) are approximately 1, suggesting enzyme activities in HLM and HHEP are similar to those in vivo under physiological conditions. For ECCS class 1A/1B compounds, a unified set of SFs was developed for CLint. These SFs contain both SFlin and an exponential SF (SFß) of fraction unbound in plasma (fu,p). The unified SFs for class 1A/1B eliminate the need to identify the transporters involved prior to clearance prediction. The underlying mechanisms of these SFs are not entirely clear at this point, but they serve practical purposes to reduce biases and increase prediction accuracy. Similar SFs have also been developed for preclinical species. For HLM-HHEP disconnect (HLM > HHEP) ECCS class 2/4 compounds that are mainly metabolized by cytochrome P450s/FMO, HLM significantly overpredicted in vivo CLint, while HHEP slightly underpredicted and geometric mean of HLM and HHEP slightly overpredicted in vivo CLint. This observation is different than in rats, where rat liver microsomal CLint correlates well with in vivo CLint for compounds demonstrating permeability-limited metabolism. The good CLint IVIVE developed using HLM and HHEP helps build confidence for prospective predictions of human clearance and supports the continued utilization of these assays to guide structure-activity relationships to improve metabolic stability.

Early Drug-Induced Liver Injury Risk Screening: "Free," as Good as It Gets.

For all the promise of and need for clinical drug-induced liver injury (DILI) risk screening systems, demonstrating the predictive value of these systems versus readily available physicochemical properties and inherent dosing information has not been thoroughly evaluated. Therefore, we utilized a systematic approach to evaluate the predictive value of in vitro safety assays including bile salt export pump transporter inhibition and cytotoxicity in HepG2 and transformed human liver epithelial along with physicochemical properties. We also evaluated the predictive value of in vitro ADME assays including hepatic partition coefficient (Kp) and its unbound counterpart because they provide insight on hepatic accumulation potential. The datasets comprised of 569 marketed drugs with FDA DILIrank annotation (most vs less/none), dose and physicochemical information, 384 drugs with Kp and plasma protein binding data, and 279 drugs with safety assay data. For each dataset and combination of input parameters, we developed random forest machine learning models and measured model performance using the receiver operator characteristic area under the curve (ROC AUC). The median ROC AUC across the various data and parameters sets ranged from 0.67 to 0.77 with little evidence of additive predictivity when including safety or ADME assay data. Subsequent machine learning models consistently demonstrated daily dose, fraction sp3 or ionization, and cLogP/D inputs produced the best, simplest model for predicting clinical DILI risk with an ROC AUC of 0.75. This systematic framework should be used for future assay predictive value assessments and highlights the need for continued improvements to clinical DILI risk annotation.

Use of high throughput ion channel profiling and statistical modeling to predict off-target arrhythmia risk - One pharma's experience and perspective.

INTRODUCTION: The use of high throughput patch clamp profiling to determine mixed ion channel-mediated arrhythmia risk was assessed using profiling data generated using proprietary internal and clinical reference compounds. We define the reproducibility of the platform and highlight inherent platform issues. The data generated was used to develop predictive models for cardiac arrhythmia risk, specifically Torsades de Pointes (TdP). METHODS: A retrospective analysis was performed using profiling data generated over a 3-year period, including patch clamp data from hERG, Cav1.2, and Nav1.5 (peak/late), together with hERG binding. RESULTS: Assay reproducibility was robust over the 3-year period examined. High throughput hERG patch IC50 values correlated well with GLP-hERG data (Pearson = 0.87). A disconnect between hERG binding and patch was observed for ∼10% compounds and trended with passive cellular permeability. hERG and Cav1.2 potency did not correlate for proprietary compounds, with more potent hERG compounds showing selectivity versus Cav1.2. For clinical compounds where hERG and Cav1.2 activity was more balanced, an analysis of TdP risk versus hERG/Cav1.2 ratio demonstrated low TdP probability when the hERG/Cav1.2 potency ratios were < 1. Modeling of clinical compound data revealed a lack of impact of the Nav1.5 (late) current in predicting TdP. Moreover, models using hERG binding data (ROC AUC = 0.876) showed an improved ability to predict TdP risk versus hERG patch clamp (ROC AUC = 0.787). DISCUSSION: The data highlight the value of high throughput patch clamp data in the prediction of TdP risk, as well as some potential limitations with this approach.

Extraction of tacit knowledge from large ADME data sets via pairwise analysis.

Pharmaceutical companies routinely collect data across multiple projects for common ADME endpoints. Although at the time of collection the data is intended for use in decision making within a specific project, knowledge can be gained by data mining the entire cross-project data set for patterns of structure-activity relationships (SAR) that may be applied to any project. One such data mining method is pairwise analysis. This method has the advantage of being able to identify small structural changes that lead to significant changes in activity. In this paper, we describe the process for full pairwise analysis of our high-throughput ADME assays routinely used for compound discovery efforts at Pfizer (microsomal clearance, passive membrane permeability, P-gp efflux, and lipophilicity). We also describe multiple strategies for the application of these transforms in a prospective manner during compound design. Finally, a detailed analysis of the activity patterns in pairs of compounds that share the same molecular transformation reveals multiple types of transforms from an SAR perspective. These include bioisosteres, additives, multiplicatives, and a type we call switches as they act to either turn on or turn off an activity.

The Impact of a Breastfeeding Promotion Intervention on Rapid Weight Gain among African American Infants.

BACKGROUND/OBJECTIVE: To evaluate the impact of exclusive breastfeeding (EBF) on rapid weight gain (RWG) among infants of African American women enrolled in a breastfeeding promotion intervention. METHODS: African American mothers in the 2nd or 3rd trimester who consented and attended four 30-minute breastfeeding promotion sessions prospectively provided breastfeeding and physical measurements at birth, four-, six-, and twelve-months. RESULTS: Mean age of mothers was 28.74±6.0 years, range 15-42 years, 62(38.8%) primiparous, 59 (36.9%) had ≤high school diploma, and 68 (42.5%) annual income <$15,000. Exclusive breastfeeding at birth, three, and six months were 104 (62.7%), 44 (34.4%), and 21 (17.9%). Rapid weight gain at four months and six months were 42 (36.2%) and 77 (74.8%). Difference in rapid weight gain at four months for babies breastfed up to three months vs. those who were not was significant, p<.04. Maternal demographics did not predict RWG in multiple regression modelling. The incidence of overweight at 12 months for babies who experienced RWG at four months vs. those who did not was significantly different, p<.001. CONCLUSION: Exclusive breastfeeding for six months was associated with reduced risk of RWG in early infancy.

Evaluation of Prediction Accuracy for Volume of Distribution in Rat and Human Using In Vitro, In Vivo, PBPK and QSAR Methods.

Volume of distribution at steady state (Vss) is an important pharmacokinetic parameter of a drug candidate. In this study, Vss prediction accuracy was evaluated by using: (1) seven methods for rat with 56 compounds, (2) four methods for human with 1276 compounds, and (3) four in vivo methods and three Kp (partition coefficient) scalar methods from scaling of three preclinical species with 125 compounds. The results showed that the global QSAR models outperformed the PBPK methods. Tissue fraction unbound (fu,t) method with adipose and muscle also provided high Vss prediction accuracy. Overall, the high performing methods for human Vss prediction are the global QSAR models, Øie-Tozer and equivalency methods from scaling of preclinical species, as well as PBPK methods with Kp scalar from preclinical species. Certain input parameter ranges rendered PBPK models inaccurate due to mass balance issues. These were addressed using appropriate theoretical limit checks. Prediction accuracy of tissue Kp were also examined. The fu,t method predicted Kp values more accurately than the PBPK methods for adipose, heart and muscle. All the methods overpredicted brain Kp and underpredicted liver Kp due to transporter effects. Successful Vss prediction involves strategic integration of in silico, in vitro and in vivo approaches.

A Physiologically Based in Silico Tool to Assess the Risk of Drug-Related Crystalluria.

Drug precipitation in the nephrons of the kidney can cause drug-induced crystal nephropathy (DICN). To aid mitigation of this risk in early drug discovery, we developed a physiologically based in silico model to predict DICN in rats, dogs, and humans. At a minimum, the likelihood of DICN is determined by the level of systemic exposure to the molecule, the molecule's physicochemical properties and the unique physiology of the kidney. Accordingly, the proposed model accounts for these properties in order to predict drug exposure relative to solubility along the nephron. Key physiological parameters of the kidney were codified in a manner consistent with previous reports. Quantitative structure-activity relationship models and in vitro assays were used to estimate drug-specific physicochemical inputs to the model. The proposed model was calibrated against urinary excretion data for 42 drugs, and the utility for DICN prediction is demonstrated through application to 20 additional drugs.

Structural attributes influencing unbound tissue distribution.

Unbound tissue-to-plasma partition coefficients (Kpuu) were determined for 56 structurally diverse compounds in rats following intravenous infusion. Five tissues were included in the study: white adipose, brain, heart, liver, and skeletal muscle. The rank ordering of the median tissue Kpuu values was: liver (4.5) > heart (1.8) > adipose (1.2) > skeletal muscle (0.6) > brain (0.05), with liver being most enriched and brain most impaired. The median Kpuu values of acids and zwitterions were lower than those of bases and neutrals in all tissues but liver. Selective tissue distribution was observed, dependent upon chemotype, which demonstrated the feasibility of targeting or restricting drug exposure in certain tissues through rational design. Physicochemical attributes for Kpuu were identified using recursive partitioning, which further classified compounds with enriched or impaired tissue distribution. The attributes identified provided valuable insight on design principles for asymmetric tissue distribution to improve efficacy or reduce toxicity.

Evaluation of Fraction Unbound Across 7 Tissues of 5 Species.

Binding to various tissues and species is frequently assessed in drug discovery and development to support safety and efficacy studies. To reduce time, cost, and labor requirements for binding experiments, we conducted a large comparison study to evaluate the correlation of fraction unbound (fu) across 7 tissues of 5 species, including white adipose, brain, heart, kidney, liver, lung, and skeletal muscle of mouse, rat, dog, monkey, and human. The results showed that there were no significant species differences of fu for tissue binding, and a single-species (e.g., rat) tissue fu can be used as a surrogate for binding in other species. Cross-tissue comparison indicated that brain, heart, liver, and muscle had quite similar fu values; rat liver binding can be used as a surrogate for binding of the other 3 tissues without any scaling factors. Binding to adipose, kidney, and lung can also be estimated with rat liver fu with scaling factors. This study suggests that a single tissue of a single species (e.g., rat liver) is a good predictor for fu of other tissues of various species with or without scaling factors. Molecular size, lipophilicity, pKa, and topological polar surface area are important physiochemical properties influencing tissue fu.

Mechanistic insights on clearance and inhibition discordance between liver microsomes and hepatocytes when clearance in liver microsomes is higher than in hepatocytes.

Human liver microsomes (HLM) and human hepatocytes (HHEP) are two common in vitro systems used in metabolic stability and inhibition studies. The comparison between the assays using the two systems can provide mechanistic insights on the interplay of metabolism, passive permeability and transporters. This study investigated the critical factors impacting the unbound intrinsic clearance (CLint,u) and IC50 of CYP3A inhibition between HLM and HHEP. The HLM/HHEP CLint,u ratio and HHEP/HLM IC50 ratio are inversely correlated to passive permeability, but have no correlation with P-gp efflux ratio. Cofactor-supplemented permeabilized HHEP (MetMax™) collapses the IC50 differences between HHEP and HLM. P-gp inhibitor, encequidar, shows minimal impact on CLint,u and IC50 in HHEP. This is the first study that is able to separately investigate the effects of passive permeability and efflux transport. These data collectively show that passive permeability plays a critical role in metabolism and enzyme inhibition in HHEP, while P-gp efflux has a minor role. This may be due to low functional P-gp activity in suspension HHEP under the assay conditions. Low passive permeability may limit metabolism and enzyme inhibition in HHEP, leading to lower CLint,u and higher IC50 in HHEP compared to HLM. When liver microsomes give higher CLint,u than hepatocytes, microsomes are more predictive of in vivo clearance than hepatocytes.

The use of matched molecular series networks for cross target structure activity relationship translation and potency prediction.

Matched molecular series (MMS) analysis is an extension of matched molecular pair (MMP) analysis where all of the MMPs belong to the same chemical series. An MMS within a biological assay is able to capture specific structure activity relationships resulting from chemical substitution at a single location in the molecule. Under this convention, an MMS has the ability to capture one specific interaction vector between the compounds in a series and their therapeutic target. MMS analysis has the potential to translate the SAR from one series to another even across different protein targets or assays. A significant limitation of this approach is the lack of chemical series with a sufficient number of overlapping fragments to establish a statistically strong SAR in most databases. This results in either an inability to perform MMS analysis altogether or a potentially high proportion of spurious matches from chance correlations when the MMS compound count is low. This paper presents the novel concept of an MMS Network, which captures the SAR relationships between a set of related MMSs and significantly enhances the performance of MMS analysis by reducing the number of spurious matches leading to the identification of unexpected and potentially transferable SAR across assays. The results of a full retrospective leave-one-out analysis and randomization simulation are provided, and examples of pharmaceutically relevant programs will be presented to demonstrate the potential of this method.