Prediction of intracellular exposure bridges the gap between target- and cell-based drug discovery

Significance

Exposure at the site of action has been identified as one of the three most important factors for success in drug discovery and the design of chemical probes. Modern drug discovery programs have, to a great extent, shifted to intracellular targets, but methods to determine intracellular drug concentrations have been lacking. Here, we use a methodology for predicting intracellular exposure of small-molecule drugs to understand their potency toward intracellular targets. We show that our approach is generally applicable to multiple targets, cell types, and therapeutic areas. We expect that routine measurements of intracellular drug concentration will contribute to reducing the high attrition observed in drug discovery and the design of both better chemical probes and medicines.

Abstract

Inadequate target exposure is a major cause of high attrition in drug discovery. Here, we show that a label-free method for quantifying the intracellular bioavailability (F_ic) of drug molecules predicts drug access to intracellular targets and hence, pharmacological effect. We determined F_ic in multiple cellular assays and cell types representing different targets from a number of therapeutic areas, including cancer, inflammation, and dementia. Both cytosolic targets and targets localized in subcellular compartments were investigated. F_ic gives insights on membrane-permeable compounds in terms of cellular potency and intracellular target engagement, compared with biochemical potency measurements alone. Knowledge of the amount of drug that is locally available to bind intracellular targets provides a powerful tool for compound selection in early drug discovery.

Most known drug targets are located in the cell interior (1, 2). Inadequate cellular drug exposure is hypothesized to lead to a lower “biochemical efficiency” (3), typified by compounds that bind to the isolated target protein with high affinity but fail to perform in cellular assays, a phenomenon termed “cell drop off” (2). Furthermore, insufficient exposure at the target is an important contributor to failure in clinical drug development and the high attrition rate in drug discovery programs (4–6). In a recent analysis of the drug candidate pipeline of a major drug company, all programs in which target exposure was uncertain (18 of 44 programs) resulted in failure of progression to phase III clinical trials (4). Importantly, routine measurement of compound levels at intracellular sites in target tissues is hindered by sampling constraints in human subjects (5).

Currently, indirect estimates of intracellular drug levels are extrapolated from transcellular permeability experiments (7) or cellular target engagement data (8–17). Some of these approaches are limited to certain target classes, because they require probes that are known to bind the target (13–16, 18) or target properties to allow detection [e.g., the stability of the target needs to be affected by the compound (8, 9)]. Furthermore, exposure can only be detected in cell types that express the target, limiting their use when studying off-target effects and drug toxicity. Qualitative measurements of the cellular distribution of compounds can be obtained with various imaging modalities (19–25); however, most of these techniques require labeling of the compound of interest. Labeling often alters the molecular properties of the compound, thereby perturbing its distribution and target affinity. Tissue-based methodologies allow accurate determinations of intracellular drug concentrations using unlabeled compounds, but they are based on animal tissues and have low throughput (26, 27). Therefore, we developed a cell-based methodology for determination of intracellular drug concentrations in a high-throughput format, which is applicable to a wide variety of cell systems, not only those directly relevant for the target pharmacology (28–30). Our method does not require chemical labeling, has a high sensitivity, and measures the unbound drug concentration. Importantly, it is this unbound fraction of the dosed drug that is available for interactions in the intracellular environment [i.e., which is intracellularly bioavailable (F_ic)] (Fig. 1A).

Fig. 1.

A comparison of biochemical and cellular assays used in drug discovery and measurement of intracellular bioavailability (F_ic). (A) In biochemical assays, the compound is directly available to engage the target. In cellular assays, only the fraction of compound that is not bound to proteins in cell culture medium (f_u,medium) or cellular components (f_u,cell) is available for binding to an intracellular target; this fraction is the intracellularly bioavailable fraction (F_ic). After engaging the target, the compound elicits a functional response, leading, in turn, to a phenotypic response. (B) General protocol for measurement of intracellular bioavailability (F_ic). Intracellular fraction of unbound compound (f_u,cell) measured after dialysis and intracellular compound accumulation (Kp) are quantified using LC-MS/MS and combined to give F_ic.

In a preliminary study with a single cytosolic target (thymidylate synthase), we showed that intracellular concentrations of active hits and reference drugs correlate with intracellular target engagement (31). We, therefore, hypothesized that F_ic might be a practical metric for predicting compound access to intracellular targets. Herein, we tested this hypothesis with multiple compound sets and in different cell types. The compounds were screened in biochemical assays, where the compound can directly access and inhibit the target protein, and cellular assays, where the compound has to enter the cell before binding the target and eliciting the cellular response (Fig. 1A).

We first investigated inhibitors of Mitogen-activated protein kinase 14 (MAPK14 or p38α), an intracellular kinase involved in autoimmune and inflammatory diseases (32). Next, we used the published kinase inhibitor set (PKIS) (33, 34) to select compounds with the potential to inhibit cell growth (a far from target end point) in a collection of 60 cancer cell lines from the National Cancer Institute, the NCI-60 cell panel (35). The selected compounds were inhibitors of cyclin-dependent kinase 2 (CDK2) (36) and polo-like kinase 1 (PLK1) (37), cytosolic kinases that are translocated to the nucleus where they have important roles in the cell cycle, and therefore, have been considered as targets in cancer therapy (38, 39). Finally, we investigated inhibitors of β-secretase 1 (BACE-1), a protease involved in the cleavage of the amyloid precursor protein (APP) into amyloid-β peptides (Aβs) (40). Modulation of this protease is proposed as a potential treatment of Alzheimer’s disease (41). This target posed an additional challenge for prediction of exposure, because it is only active in early endosomes (42–44). Compound concentration in these organelles often differs from that in the cytosol, complicating predictions of subcellular pharmacology (45, 46).

We found that compounds displaying high cellular potency generally exhibited high F_ic, whereas those with a lower potency in the cell than toward the isolated target (i.e., compounds with a marked cell drop off) had a low F_ic (Fig. 1A). Knowledge of F_ic consistently improved predictions of cellular drug potency compared with using biochemical target binding assays alone. We conclude that F_ic is a generally applicable parameter for predicting target exposure and ranking of compounds acting on intracellular targets. With our technique (Fig. 1B), F_ic can be measured in a high-throughput manner in a variety of cell types, including those directly relevant for the disease (e.g., patient cells).

Results

Intracellular Bioavailability (F_ic) of p38α (MAPK14) Inhibitors.

We investigated the impact of intracellular bioavailability (F_ic) (below and Fig. 1B) on the cellular potency of a set of 35 inhibitors of MAPK14 (p38α). These compounds were screened for inhibition of TNFα expression in human peripheral blood mononuclear cells (PBMCs) (Fig. 2A and Dataset S1) (32, 47, 48). Although there was a good correlation (r_S = 0.83) between the biochemical negative logarithm of the half-maximal inhibitory concentration (pIC₅₀) and the cellular pIC₅₀ for these compounds (Fig. 2B), the compounds were, on average, one order of magnitude more potent in the biochemical assay (where all added compound can interact directly with the target) (Fig. 1A). For kinase inhibitors, this drop in potency is commonly attributed to the high cellular ATP concentrations (compared with those used in biochemical assays) that can compete with the compound (49, 50), leading to lower biochemical efficiency (3). Here, we studied if, in addition to this mechanism, the access of these inhibitors to the intracellular target in the cellular assay was limited. We therefore determined the F_ic of these compounds by measuring the intracellular fraction of unbound compound (f_u,cell) and the cellular compound accumulation (Kp) (details are in Methods and Fig. 1B) in PBMCs (i.e., the same cell type as in the cellular pharmacological assay) (SI Appendix, Table S3).

Fig. 2.

Intracellular bioavailability (F_ic) of MAPK14 (p38α) inhibitors. (A) p38α Inhibition leads to a reduction in TNFα expression via p38α-activated MAPK-activated protein kinase 2 (MK2) and MK3, which are responsible for the phosphorylation of tristetraprolin (TTP). TTP inhibits TNFα mRNA translation by binding the AU-rich element in the 3′ UTR (32, 48, 70). (B) p38α Inhibitors were less potent in the cellular assay than in the biochemical assay (linear fit: y = 1.6x − 5.6; P < 0.0001). Biochemical potency was determined using an assay based on inhibition of ATP consumption, and cellular potency was determined as the capacity to inhibit production of TNFα in PBMCs. Both measurements were performed with technical triplicates at each concentration. (C) Frequency distribution of intracellular bioavailability (F_ic) of p38α inhibitors measured in PBMCs (blue line) and HL60 cells (pink line) in triplicate on two independent occasions. (D) Prediction of cellular pIC₅₀ of p38α inhibitors by combining their F_ic in PBMCs with their biochemical pIC₅₀ (linear fit: y = 1.2x − 1.4; P < 0.0001). (E) Structures of two enantiomers (compounds 1 and 2) that displayed similar biochemical pIC₅₀ against p38α. (F) Compound 2 (purple) showed a higher F_ic in PBMCs than compound 1 (green), which was accompanied by an increase of similar magnitude in cellular pIC₅₀ (Lower). Bars represent geometrical mean ± SEM. (G) Prediction of intracellular target engagement pEC₅₀ of p38α inhibitors by combining their F_ic in HL60 cells with their target affinity (biochemical pIC₅₀; linear fit: y = 0.9x + 0.21; P < 0.0001). CETSA measurements were performed with technical triplicates at each concentration.

In agreement with the observation of marked cell drop off, most compounds displayed low F_ic (median = 0.088, interquartile range = 0.069–0.19) (Fig. 2C). Because F_ic represents the fraction of extracellularly added compound that is able to reach intracellular targets, we set out to predict the cellular pIC₅₀ by using the F_ic to correct the biochemically determined potencies of the compounds toward the isolated target (log F_ic + biochemical pIC₅₀). A good correlation was still observed between the predicted and measured cellular potencies (r_S = 0.79), and importantly, the predicted cellular potencies were now at the same level as the experimental ones (Fig. 2D)—indicating that the use of F_ic can explain the observed cell drop off.

A low F_ic, commensurate with a significant cell drop off, has been historically attributed to low membrane permeability. However, we observed a poor correlation for these compounds between F_ic and membrane permeability as determined in a typical artificial membrane system [parallel artificial membrane permeability assay (PAMPA)] routinely used in drug discovery (r_S = 0.03) (SI Appendix, Fig. S1A and Table S5). In contrast to F_ic, the permeability data did not explain the cell drop off (r_S = 0.21) (SI Appendix, Fig. S2A). These results are in line with our previous observations that F_ic is the net result of membrane permeability and other mechanisms, including carrier-mediated transport, metabolism, and nonspecific cellular binding (51). It should also be remembered that the widely used artificial membrane permeability data are reported as a rate, whereas F_ic is directly related to the amount of unbound drug in the cells and not related to how fast it gets there.

Interestingly, our dataset included one pair of enantiomers (compounds 1 and 2) (Fig. 2E) with identical biochemical potencies (IC₅₀ = 12 nM) but different cellular potencies (12 and 4.5 nM, respectively) (Fig. 2F). The 2.5-fold higher potency of compound 2 accompanied a 2.3-fold higher F_ic, meaning that F_ic accurately detected the increase in cellular potency caused by structural differences (Fig. 2F). We speculated that the difference in cellular potency of these enantiomers was caused by selective active transport across the cellular membrane (51). We, therefore, investigated the F_ic of these compounds in the presence of cyclosporine A, a paninhibitor of active transport processes (52, 53). The F_ic of compound 1 was threefold higher in the presence of cyclosporine A (P = 0.05, Mann–Whitney U test), whereas there was no change for compound 2 (P = 0.95) (SI Appendix, Fig. S3). This result suggested the involvement of carrier-mediated efflux in the cellular disposition of compound 1 (but not of 2), which decreased its bioavailability in the cell interior and therefore, limited its cellular potency. These observations illustrate how F_ic, when measured in the relevant cell type, can provide information about target exposure without prior knowledge of the underlying mechanisms of compound disposition (51).

To further explore the versatility of F_ic as a metric of compound access to intracellular targets, we investigated a different cell type (HL60) that also expresses p38α. As in PBMC, the F_ic values of these compounds in this cell type were also low (median = 0.11, interquartile range = 0.062–0.45) (Fig. 2C). Subsequently, we determined intracellular target engagement using the cellular thermal shift assay (CETSA) in the isothermal dose–response fingerprint (ITDRF) mode (Dataset S1) (8, 9). After correcting the target affinity (biochemical pIC₅₀) for F_ic in the relevant cell type, the extent of intracellular target engagement was successfully predicted (r_S = 0.76) (Fig. 2G). Thermal shift assays depend on not only the affinity of the ligand and its concentration at the target but also, the thermodynamics of its binding and protein unfolding (54, 55). The latter two were not considered in our prediction and may, therefore, explain the minor offset observed (on average, 0.5 log units).

Intracellular Bioavailability (F_ic) of Compounds from the PKIS.

Next, we applied F_ic on molecules with more complex cellular pharmacology by investigating the impact of F_ic on the cellular potency of compounds from the PKIS (33, 34). We selected kinase inhibitors that affect the growth of cell lines in the NCI-60 cell panel (Fig. 3A) (34, 35). We grouped the compounds according to their biochemical affinity profiles in the Nanosyn kinase panel (www.nanosyn.com/) using hierarchical clustering with complete linkage (56) (SI Appendix, Fig. S4). This procedure provided clusters of compounds that inhibit the same kinases and therefore, should have similar effects on the cell. To enable the study of compounds with potential to impact cell growth, we then selected only the clusters that contained at least one compound that was active [negative logarithm of half-maximal growth inhibition concentration (pGI₅₀) > 6] in the majority of the NCI-60 cell lines (SI Appendix, Fig. S4). The selection resulted in three clusters of compounds (Fig. 3A): one cluster (n = 39) containing compounds originally targeting CDK2 or glycogen synthase kinase 3 (GSK3) and two clusters (n = 9 and n = 8) that included compounds initially developed as inhibitors of PLK1 (Dataset S2). Importantly, the potency of these compounds toward their respective original intracellular targets was a poor predictor of their pGI₅₀ (r_S = 0.33) (SI Appendix, Fig. S5). To assess if the differences in pGI₅₀ were a result of different concentrations of compound available at the target, we determined their F_ic (SI Appendix, Table S2).

Fig. 3.

Intracellular bioavailability of compounds from the PKIS. (A) Adapted with permission from Macmillan Publishers Ltd: Nature Biotechnology, ref. 34, copyright 2015. Each row represents one compound, and each column represents one kinase of the Nanosyn kinase panel (www.nanosyn.com/) or one cell line of the NCI-60 panel. Colors represent the potency of the compounds (pIC₅₀ for kinases and pGI₅₀ for cell lines). In Left, compounds, kinases, and cell lines are each sorted alphabetically. In Right, compounds were clustered based on their biochemical data to identify groups of compounds with similar target interaction patterns (SI Appendix, Fig. S4). For a better visualization of the cell growth inhibition potential of each compound, cell lines were sorted by pGI₅₀ (from the most affected to the least affected cell line). Selected compounds are highlighted in blue, and their primary targets are in gray. (B) Relationship between cellular potency (pGI₅₀) and F_ic. Cellular potency represents an average of the pGI₅₀ in the NCI-60 panel; F_ic was measured in HEK293 cells in triplicate and on two independent occasions (geometrical mean is shown). (C) Classification model based on pGI₅₀ and F_ic. Classes are defined by thresholds of intracellular bioavailability (F_ic = 1, which corresponds to a concentration at the target equal to the concentration added to the cells) and cellular potency (in this example, pGI₅₀ = 6).

The F_ic of the selected compounds followed the cellular potency (r_s = 0.64) (Fig. 3B). In other words, compounds with a high F_ic were also highly active in cells. As for the p38α inhibitors, we predicted the cellular pGI₅₀ by correcting the biochemical potency with the specific F_ic of each compound. The predicted pGI₅₀ values were, on average, one log unit higher than the measured ones (r_s = 0.65) (SI Appendix, Fig. S6). This difference is perhaps not surprising given that cell growth inhibition is a readout far downstream of the kinase inhibition event, with redundant pathways in the cell attenuating the effect of the compound.

For such cases where the relation between target inhibition and pharmacological effect is not clearly defined (e.g., in phenotypic screening assays), we established a classification model to visualize the impact of the compound F_ic on the cellular activity. The model comprises four classes (Fig. 3C). Class 1 compounds are active in cellular assays, and their concentrations at the target are equal or higher than the nominal concentration added to the cells. Class 2 compounds are active in cells, despite restricted access to the target [compounds with high affinity to the target(s)]. Class 3 compounds are inactive in cells, although they have high F_ic [compounds with low affinity to the target(s)], and class 4 compounds are not active in cells and have low F_ic. For the PKIS dataset, we defined compounds as active in the cellular assay if their pGI₅₀ was ≥6 (the same cutoff used for compound selection as described above). Most compounds (71%) in the dataset belonged to either class 1 (23%) or class 4 (48%) (Fig. 3C), where the cellular potency is explained by F_ic. Only 18% of compounds were in class 2 and 11% in class 3, with all compounds being close to the cutoffs of each category (Fig. 3C). Our observations thus suggested that F_ic is a useful predictor of compound activity in cellular assays even when there is only a modest pharmacological connection between the cellular readout and target inhibition [or if the target is completely unknown (e.g., in phenotypic screening)].

Endosomal Bioavailability (F_endo) of BACE-1 Inhibitors.

To explore the limits of the F_ic methodology, we selected an intracellular target (BACE-1) that is only active in the early endosomal subcellular compartment (42–44). In other words, the compounds must reach this low pH compartment to exert their effects (Fig. 4A). Because F_ic reflects an average bioavailability in the cell interior, we expected a lower predictive power than for targets available in the cytoplasm (e.g., p38α).

Fig. 4.

Endosomal bioavailability (F_endo) of BACE-1 inhibitors. (A) BACE-1 inhibition leads to a reduction of Aβ_1–42 formation from APP. Proteolysis of APP by BACE-1 occurs in early endosomes. (B) BACE-1 inhibitors are less potent in a cellular assay than in a biochemical assay (linear fit: y = 0.75x − 0.96; P = 0.004). Biochemical potency was determined as the capacity to inhibit the cleavage of a peptide based on the Swedish mutant APP sequence, and cellular potency was determined as the capacity to inhibit formation of Aβ_1–42 in SH-SY5Y cells. Both measurements were performed with technical triplicates at each concentration. (C) Prediction of cellular pIC₅₀ of BACE-1 inhibitors by combining their endosomal bioavailability (F_endo) in SH-SY5Y cells with their biochemical pIC₅₀ (linear fit: y = 0.94x − 1.2; P < 0.0001). F_endo was calculated from F_ic, which was measured in triplicate on two independent occasions (geometrical mean is shown).

Thirty BACE-1 inhibitors were assayed both biochemically and in SH-SY5Y cells for their capacity to inhibit the cleavage of an APP peptide containing the Swedish mutation (57) (Dataset S3). On average, potencies were one order of magnitude lower in the cellular screen than in the biochemical screen (r_S = 0.49) (Fig. 4B). In accordance, these inhibitors displayed low F_ic in the SH-SY5Y cells (median = 0.27, interquartile range = 0.16–0.72) (SI Appendix, Table S4).

As expected from the specific localization required, the biochemical pIC₅₀ corrected for the F_ic was not a good predictor of the cellular pIC₅₀ of these compounds (SI Appendix, Fig. S9). In contrast, a much improved correlation with the cellular pIC₅₀ (r_S = 0.82) was observed when the biochemical potencies were instead corrected using the predicted accumulation in early endosomes [F_endo; calculated from F_ic assuming pH-dependent compound partitioning into subcellular compartments (45, 46)] (details are in Methods, SI Appendix, Table S4, and Dataset S3). Interestingly, all of these compounds were now predicted to be more potent than what was actually observed in the cellular assay (Fig. 4C). A contributing factor to this discrepancy is likely that BACE-1 is not present in all early endosomes [as evident from double-staining experiments with BACE-1 and endosomal markers (42, 58, 59)]. In contrast, the positively charged BACE-1 inhibitors will be evenly distributed in the entire endolysosomal space, and consequently, only a fraction of the estimated F_endo is colocalized with the target and can exert its effect. Additional studies are warranted to confirm this hypothesis, but the overprediction of cellular potency by 1.7 orders of magnitude suggests that the fraction of active and target-accessible compound is only 2% of the concentration added to the cells.

Discussion

Knowing the amount of compound available at the target has been identified as a key factor for reducing attrition rates in the drug discovery process (4, 5). Lack of high-throughput methodologies has precluded routine measurement of this parameter for targets in the cell interior. Instead, drug discovery efforts are commonly directed at increasing membrane permeability in the hope that it translates to increased intracellular compound levels. However, a high permeability across a membrane layer does not always result in high compound concentrations in the cell interior (as evident by the lack of correlation between membrane permeability and F_ic) (SI Appendix, Fig. S1). This discrepancy is because membrane permeability is only one of many factors that can affect intracellular compound concentrations (51). Thus, although cell permeability assays are important tools in predicting oral bioavailability, they are not ideal for the prediction of intracellular drug exposure. In contrast, the intracellular bioavailability concept (F_ic) (Fig. 1B) (28) provides a direct measure of intracellular drug exposure and considerably improves predictions of cellular drug response for targets in a variety of subcellular compartments (including the cytosol, nucleus, and endosomes).

By predicting F_ic in different cell types and for compound sets with diverse biological mechanisms, we have shown the general applicability of our approach. The F_ic value compiles all ongoing processes that influence target exposure, including cell-specific transport and elimination mechanisms (51). The aggregation of multiple parallel processes explains why F_ic is in agreement with not only responses closely related to the target (such as p38α target engagement and the close-downstream production of TNFα) but also, far from target responses that are only indirectly associated with target inhibition (such as cell growth inhibition elicited by kinase inhibitors in the PKIS). Furthermore, by accounting for pH-dependent distribution, predictions of BACE-1 inhibitory activity in SH-SY5Y cells were considerably improved for the majority of the compounds tested; such predictions were not possible with a biochemical assay alone for this target, which is located in a specific subcellular compartment.

In our example of a close to target cellular response (p38α inhibition of TNFα production) (Fig. 2A), the inhibitors had a low F_ic in both PBMCs and HL60 cells (Fig. 2C), which translated to a lower potency in the cellular assay than in the biochemical assay (Figs. 1A and 2B). By simply correcting the biochemical pIC₅₀ of these compounds with the bioavailable concentration at the target (i.e., F_ic), we obtained good predictions of their cellular potency (Fig. 2D) and intracellular target engagement (Fig. 2G). For the cellular potency, 74% of the compounds were predicted within one order of magnitude of their true potency. The high predictivity is remarkable considering that kinase inhibitors are generally promiscuous (Fig. 3A) (60), with inhibition of other kinases possibly affecting TNFα expression.

Our PKIS kinase inhibition data showed that a high F_ic is required for pharmacological effect of compounds that act on intracellular targets (Fig. 3B). Where the connection between the cellular activity and target inhibition is unknown (e.g., in phenotypic screening), we implemented a simple and flexible classification model for the interpretation of the cellular potency results in relation to the cellular drug exposure (Fig. 3C). In this classification system, the cellular potency cutoff (i.e., the potency at which a compound is considered active in the cells) can be adjusted to reflect the desired cellular potency. For classes 1 and 4, the F_ic explains why some compounds are active (class 1; high F_ic, high cellular potency) and why some have low cellular potency (class 4; low F_ic, low cellular potency). The class 2 compounds (low F_ic, high cellular potency) will appear when the biochemical potency is very high (e.g., as a consequence of high lipophilicity) (61). These compounds may have safety issues, because F_ic is cell type-specific (51); cells other than the one tested might be exposed to higher than anticipated levels of drug. In contrast, class 3 compounds (high F_ic, low cellular potency) will appear when the biochemical potency is low or when the compound targets proteins that are irrelevant to the desired cellular response [i.e., off-target effects (62)] (SI Appendix, SI Results). Thus, by knowing F_ic, any discrepancies between biochemical and cellular potencies can be rationalized, and optimization efforts can be focused (e.g., on increasing target affinity vs. intracellular exposure).

Our studies with BACE-1 inhibitors showed that F_ic can provide estimates of subcellular distribution of compounds after taking into consideration the organelle pH and volume, and the local compound charge. Importantly, biochemical assays do not provide such information. Indeed, the biochemical BACE-1 assay did not predict the cellular potency (Fig. 4B). When the biochemical pIC₅₀ was corrected for the endosomal bioavailability (F_endo), a much better prediction of the cellular potency of these compounds was obtained (r_S = 0.82). Still, the cellular pIC₅₀ was lower than predicted (Fig. 4C). We attribute this to the fact that the compound is evenly distributed across all acidic organelles, whereas only 2% is located in endosomes that also contain the target and its substrate (APP) (42, 58, 59). Additional studies are needed to confirm this hypothesis.

Although F_ic is broadly applicable for predicting intracellular target exposure, its potential limitations should be identified and scrutinized. First, it relies on sensitive analytical MS for compound quantification, which is considered a low-throughput technology. However, we recently showed that extensions of our methodology can process >1,000 compounds per week (29, 30), which should be sufficient for secondary screening of hits from biochemical assays. Second, for compounds that need to be activated in the cell (prodrugs), the F_ic of the dosed parent molecule might not be relevant. However, metabolically activated species can easily be monitored as well, such as we previously did for compounds targeting thymidylate synthase (31). Third, the method is designed to give information on exposure in the whole cell (i.e., an average of cytosolic and organelle exposure), which limits the resolution for targets located in specific subcellular compartments. The example with BACE-1 inhibitors shows that mathematical modeling can be used to deconvolute the accumulation of the compounds in early endosomes. This approach can be applied to other subcellular compartments if their volumes, pH, and other relevant physical properties are considered (63). A powerful expansion of our technology would be to combine it with high-content imaging techniques (20–25) or proteome-wide target engagement approaches (10, 11), allowing for a quantitative assessment of intracellular compound distribution.

In conclusion, we show that our label-free F_ic technology accurately predicts cellular and subcellular drug exposure across multiple different drug targets and cell types. Our methodology provides direct measurements of intracellularly available drug concentrations in cells relevant to the pharmacological effect in contrast to indirect extrapolations from more demanding cell permeability experiments in generic cell lines. Compared with imaging-based studies, which typically require labeling of compounds, our label-free approach ensures that compound localization and target affinity are not altered by chemical modification. We believe that insight into the amount of drug that is locally available to bind intracellular targets is a powerful tool for improving success rates in early drug discovery. Our method provides essential information about drug exposure at the site of action, the first of the “three pillars” for successful clinical drug development (4), and we believe that it will contribute to improved translation of target potency. Thus, when combined with methodologies to measure target engagement and pharmacological response (the other pillars), it has great potential for contributing to reducing drug attrition. Beyond drug discovery, F_ic can be universally applied in other fields where information of intracellular exposure is desired, such as in toxicological and metabolomics research and any applied science where a clearer understanding of local intracellular concentrations would be beneficial.

Methods

Materials.

Reagents for cell culture were purchased from Thermo Fisher Scientific or Sigma-Aldrich. Cell lines used in this study were acquired from ATCC (HL60: catalog no. CCL-240; SH-SY5Y: catalog no. CRL-2266) or Thermo Fisher Scientific (HEK293: catalog no. R75007) and regularly tested for mycoplasma contamination with the MycoAlert mycoplasma detection kit (Lonza) according to the manufacturer’s instructions. Drug-like compounds were synthesized at GlaxoSmithKline or made available from their screening collections. Compound purity was confirmed, and only compounds with ≥95% purity were used. Compounds were dissolved in DMSO at a concentration of 10 mM and stored at −80 °C.

Cell Culture.

Cell cultures were kept at 37 °C in a humidified 5% CO₂ atmosphere. HL60 cells were cultured in Roswell Park Memorial Institute medium 1640 (RPMI 1640) with GlutaMAX, 10% FBS, penicillin (100 units/mL), and streptomycin (100 µg/mL). HEK293 cells were maintained in DMEM supplemented with 10% FBS, 2 mM l-glutamine, and 75 µg/mL Hygromycin B. SH-SY5Y cells were grown in a 1:1 mixture of DMEM and nutrient F12 medium with GlutaMAX and 10% FBS.

PBMCs were isolated from a buffy coat purchased from the Uppsala University Hospital, which has ethical approval for blood collection from the Uppsala Regional Ethics Committee. Volunteers gave informed consent, and all samples were anonymized immediately after collection of blood samples. The buffy coat was layered on top of Histopaque-1077 (Sigma-Aldrich) and centrifuged for 30 min at 400 × g. Cells were collected from the interface and washed three times by centrifugation (200 × g for 5 min) in CO₂-independent medium (Thermo Fisher Scientific) supplemented with 5% heat-inactivated FBS, penicillin (100 units/mL), and streptomycin (100 µg/mL). Cells were resuspended in RPMI 1640 with GlutaMAX; supplemented with 10% FBS, penicillin (100 units/mL), and streptomycin (100 µg/mL); and used immediately after isolation.

Biochemical and Cellular Screens of p38α Inhibition.

Inhibition potency toward purified p38α (recombinantly expressed in Escherichia coli) was measured for a series of 35 compounds. Compounds were incubated for 30 min with 5 mU p38α and 50 µM [³³P-γ-ATP] in 25 mM Tris⋅HCl, pH 7.5, 100 µM EGTA, 330 µg/mL myelin basic protein, and 10 mM magnesium acetate. Assays were stopped by addition of 5 µL of 500 mM orthophosphoric acid and then harvested onto P81 Unifilter plates (Sigma-Aldrich) with a wash buffer of 50 mM orthophosphoric acid. After scintillation counting, pIC₅₀ was calculated based on the intensity of the signal at different compound concentrations.

Inhibition of TNFα production (a proximal downstream target of p38α) was measured for the same series of compounds. PMBCs (10,000 cells per well in 384-well plates in RPMI 1640 supplemented with 10% FBS, 100 units/mL penicillin G, 100 µg/mL streptomycin sulfate, 250 ng/mL amphotericin B) were incubated with the compound and LPS (final concentration: 100 pg/mL) for 4 h at 37 °C and 5% CO₂. At the end of the incubation, TNFα levels were determined using a pair of antibodies, one of which was labeled with europium cryptate and the other was conjugated with XL665 (catalog no. 62TNFPEC; Cisbio). After 2 h of incubation in the dark, fluorescence was measured (320-nm excitation; 615- and 665-nm emission) on an EnVision plate reader (PerkinElmer). pIC₅₀ was calculated based on the fluorescence intensity ratio (665/615 nm) at different concentrations of compound.

Target engagement was measured using a high-throughput version of the CETSA (9). In a preliminary experiment, we determined the apparent temperatures of aggregation (T_aggs) of p38α with and without 10 µM AMG-548 (Tocris Biosciences), a known inhibitor of p38α (64), to be 58 °C and 47 °C, respectively (SI Appendix, Fig. S10). The p38α inhibitors were then screened in ITDRF mode at 52 °C to maximize the response window in the presence of inhibitors. Positive (10 µM AMG-548) and negative controls (DMSO) were included on each assay plate. Briefly, cells (16,000 cells per well in 384-well plates) were incubated at 37 °C with the compounds diluted in RPMI 1640 (supplemented with 10% FBS, 2 mM l-glutamine, 100 units/mL penicillin, 100 µg/mL streptomycin) for 30 min before heating to 52 °C for 3 min and cooling to 20 °C with a ProFlex PCR System (Thermo Fisher Scientific). Cells were lysed with Alpha SureFire Ultra lysis buffer (PerkinElmer) and kept at −80 °C until detection. For detection of the remaining folded p38α, we added a detection mix consisting of 10 µg/mL rabbit acceptor beads (catalog no. AL104C; PerkinElmer), 10 µg/mL mouse donor beads (catalog no. AS104D; PerkinElmer), 0.2 nM rabbit anti-p38α antibody (catalog no. ab170099; Abcam), 0.8 nM mouse anti-p38α antibody (catalog no. ab31828; Abcam), and 0.05% SDS in immunoassay buffer (PerkinElmer). After overnight incubation at room temperature, luminescence was read on an EnVision plate reader, and the negative logarithm of the half-maximal stabilization concentration (pEC₅₀) was calculated based on the fraction of stabilized protein [(signal – negative control)/positive control] at different compound concentrations.

PKIS Data.

Data from the PKIS screening for inhibition of kinases were collected from ChEMBL (https://www.ebi.ac.uk/chembldb/extra/PKIS/). These data included percentages of inhibition of 200 kinases (Nanosyn kinase panel; www.nanosyn.com/) at two concentrations (0.1 and 1 µM), named biochemical data herein. The pIC₅₀ for each compound–kinase pair was estimated (SI Appendix, Fig. S11 and Dataset S2). Because of the sigmoidal nature of the IC₅₀ curve, we limited the predictions of pIC₅₀ to values of inhibition between 20% and 80%, because outside this range, experimental errors would considerably affect pIC₅₀ predictions. Hence, the following rules were applied to the data in this order: (i) If inhibition was ≥80% at 0.1 µM, pIC₅₀ was set to eight. (ii) If inhibition was between 20% and 80% at 0.1 µM, pIC₅₀ was estimated according to Eq. 1 ([I] = 10⁻⁷ M, and percentage inhibition [%inhibition] was the value at that concentration). (iii) If inhibition was between 20% and 80% at 1 µM, pIC₅₀ was estimated according to Eq. 1 ([I] = 10⁻⁶ M, and percentage inhibition [%inhibition] was the value at that concentration). (iv) If inhibition was ≤20% at 1 µM, pIC₅₀ was set to five:

$${\text{pIC}}_{50}=-\log \left[\left[\text{I}\right]\cdot \left(\frac{100}{\text{\%inhibition}}-1\right)\right].$$ [1]

Data from the PKIS screening for growth inhibition of cell lines in the NCI-60 panel were collected from the work by Elkins et al. (34), named cellular data herein.

Compounds with potential to inhibit cell growth were selected from the PKIS and grouped according to their biochemical affinity profiles using hierarchical clustering with complete linkage (56) (SI Appendix, Fig. S4). Clusters of compounds were selected that contained at least one compound that was active (pGI₅₀ > 6) in more than one-half of the NCI-60 cell lines (SI Appendix, Fig. S4). This procedure resulted in three clusters containing a total of 51 compounds (Dataset S2).

Biochemical and Cellular Screens of BACE-1 Inhibition.

BACE-1 biochemical and cellular data were generated as part of a legacy drug discovery program undertaken within GlaxoSmithKline. Inhibition potency of a series of 30 compounds toward purified BACE-1 was measured using a quenched fluorescent peptide based on the Swedish mutant APP sequence (FAM-SEVNLDAEFK-TAMRA). Typically, compounds were incubated with 5 µM substrate and ∼1 nM purified BACE-1 at pH 4.5 for 240 min as described by Hussain et al. (57). Quantification of cleaved substrate was measured using an Analyst fluorimeter (LJL Biosystems; 485-nm excitation, 535-nm emission). The pIC₅₀ was calculated based on the intensity of the signal at different compound concentrations.

For the same series of compounds, inhibition of BACE-1 in SH-SY5Y cells expressing the Swedish mutant of APP was determined with an AlphaLISA assay (PerkinElmer) or bioequivalent MSD assay (catalog no. K151FUE; Meso Scale Discovery). Typically, 10,000 cells per well in 384-well plates were incubated with compounds (in a 1:1 mixture of DMEM and nutrient F12 medium supplemented with 10% FBS) for 48 h. For the AlphaLISA assay, Aβ_1–42 was quantified in the supernatant using a pair of antibodies, of which one was biotinylated (catalog no. 6E10; Signet Laboratories) and the other was conjugated to acceptor beads (produced in house). Chemiluminescence (680-nm excitation, 615-nm emission) was measured after addition of streptavidin-coated donor beads on an EnVision plate reader. For the MSD assay, Aβ_1–42 was quantified in the supernatant using a pair of antibodies, of which one was adsorbed to an MSD plate and the other was conjugated to a ruthenium MSD tag. Electrochemiluminescence was measured using an MSD sector Imager 6000 reader (Meso Scale Discovery). pIC₅₀ was calculated based on the fluorescence intensities at different concentrations of compound.

Measurement of Intracellular Compound Bioavailability (F_ic).

Intracellular compound bioavailability (F_ic) was determined using the technique described in the work by Mateus et al. (28, 29). This technique is made up of parallel measurement of the intracellular fraction of unbound compound (f_u,cell) and the steady-state cellular compound accumulation (Kp) as described below (Fig. 1B).

For p38α inhibitors, F_ic was first evaluated in PBMCs for comparison with the TNFα cellular screen and then evaluated in HL60 cells for comparison with the CETSA results. For compounds from the PKIS, F_ic was determined in HEK293 cells. This cell line was not included in the NCI-60 panel, but its gene expression profile was close to that of an average cell line from this panel (SI Appendix, Fig. S7), allowing the use of the average of pGI₅₀ across 60 cell lines as a measure of compound potency. For BACE-1 inhibitors, F_ic was measured in SH-SY5Y cells, the same as in the cellular potency screen.

Measurement of f_u,cell.

The f_u,cell was measured in cell homogenates at a concentration of 10 × 10⁶ cells per 1 mL using a Rapid Equilibrium Dialysis device (Thermo Fisher Scientific). Groups of six substances, together with high and low binding controls (atorvastatin and lopinavir, respectively), were added to cell homogenates at a concentration of 0.5 µM (29). Dialysis was performed against HBSS at 37 °C for 4 h; f_u,cell was calculated as

$${\text{f}}_{\text{u,cell}}=\frac{1}{\text{D}\cdot \left(\frac{1}{{\text{f}}_{\text{u,hom}}}-1\right)+1},$$ [2]

where f_u,hom is the ratio of compound concentrations in the buffer chamber and the cell homogenate chamber, and D is used to correct for homogenate dilution. D is calculated as 1/(P_hom.⋅V_cell), where P_hom. is the protein concentration of the cell homogenate (in milligrams per microliter), and V_cell is the cellular volume [6.5 µL/mg protein (28)]. Each compound was measured in at least three different randomly assigned groups. Compounds with f_u,cell < 0.01% were excluded because of large variability in the determination of f_u,cell below this value.

Measurement of Kp.

Before Kp measurements, freshly isolated PBMCs were transferred to a 96-well plate (500,000 cells per well) immediately before the experiment. HL60 cells in exponential growth phase (<1 × 10⁶ cells per 1 mL in the culture flask) were resuspended in RPMI 1640 with GlutaMAX and 10% FBS and transferred to a 96-well plate (500,000 cells per well) immediately before the experiment. HEK293 cells were seeded for 48 h in 24-well plates (600,000 cells per well) in DMEM with 10% FBS and 2 mM l-glutamine. SH-SY5Y cells were seeded in 24-well plates (500,000 cells per well) in a 1:1 mixture of DMEM and nutrient F12 medium with GlutaMAX and 10% FBS immediately before the experiment.

Measurement of Kp was then performed by incubating cells at 37 °C with 0.5 µM compound solutions (in HBSS for HEK293 cells or the culture medium described above for each of the other cell types). The incubation was stopped after 45 min for PBMCs, HL60 cells, and HEK293 cells and after 48 h for SH-SY5Y cells. The longer incubation time for SH-SY5Y reflected that used in the corresponding cellular potency assay. At the end of the experiment, a sample of the incubated solutions was collected, and cells were rapidly washed with ice-cold PBS before compound extraction. Kp was calculated as

$$\text{Kp}=\frac{\frac{{\text{A}}_{\text{cell}}}{({\text{V}}_{\text{cell}}\cdot {\text{P}}_{\text{cell}})}}{{\text{C}}_{\text{extracellular}}},$$ [3]

where A_cell is the amount of substance in the cell fraction, P_cell is the protein amount of the cell fraction (in milligrams), and C_{extracellular} is the compound concentration in the incubated solution at the end of the experiment. Each compound was measured in triplicate in at least two independent occasions.

Calculation of F_ic.

Intracellular bioavailability (F_ic), which represents the ratio of intracellular unbound (C_u,cell) concentration to extracellular (C_{extracellular}) compound concentration, was determined as

$${\text{F}}_{\text{ic}}=\frac{{\text{C}}_{\text{u,cell}}}{{\text{C}}_{\text{extracellular}}}={\text{f}}_{\text{u,cell}}\cdot \text{Kp}.$$ [4]

Estimation of Endosomal Bioavailability (F_endo).

In analogy to intracellular bioavailability (F_ic), we defined endosomal bioavailability (F_endo) as the ratio between endosomal unbound compound concentration (C_u,endo) and extracellular compound concentration (C_{extracellular}) (Eq. 9). F_endo was estimated using a model based on pH partitioning theory (65). The model consisted of three compartments: the extracellular space (pH 7.4), the cytosol [pH 7.2 (66)], and early endosomes, where BACE-1 is active (42–44) [i.e., a pH of 6 (66) and volume equal to 1% of the cytosolic volume (67)]. Accumulation of basic compounds in the endolysosomal compartment produces significant increases in the volume of these organelles (68, 69). However, F_endo estimates were insensitive to changes in the volume of the endosomal compartment (up to 5% of the cytosolic volume). Therefore, we used the baseline value of endosomal volume, 1% of cytosolic volume (67). In the model, transfer between compartments was limited to the unbound uncharged molecular species (C_u,uncharged; i.e., at equilibrium, C_u,uncharged was assumed to be equal in three compartments). C_u,uncharged was estimated from F_ic and the Henderson–Hasselbalch equation for basic compounds in the following manner:

$${\text{F}}_{\text{ic}}=\frac{{\text{C}}_{\text{u,cell}}}{{\text{C}}_{\text{extracellular}}}=\frac{({\text{A}}_{\text{u,cyto}}+{\text{A}}_{\text{u,endo}})\cdot {\text{V}}_{\text{cell}}}{{\text{C}}_{\text{extracellular}}},$$ [5]

where A_u,cyto and A_u,endo are the amounts of compound unbound in the cytosol and the endosomes, respectively, the sum of which was assumed to be the total amount of unbound drug in the cell. V_cell is the volume of the cell, considered to be equal to the cytosolic volume (V_cyto). The amount of the unbound species in each compartment (A_u,comp) was the sum of the amounts of uncharged (A_u,uncharged) and charged (A_u,charged) species:

$${\text{A}}_{\text{u,comp}}=({\text{C}}_{\text{u,uncharged}}+{\text{C}}_{\text{u,charged}})\cdot {\text{V}}_{\text{comp}},$$ [6]

where V_comp is the volume of the relevant compartment. C_u,charged was calculated with the Henderson–Hasselbalch equation for basic compounds:

$${\text{C}}_{\text{u,charged}}=\frac{{\text{C}}_{\text{u,uncharged}}}{{10}^{{\text{pH}}_{\text{comp}}-\text{pKa}}},$$ [7]

where C_u,charged is the concentration of the unbound charged molecular species, pH_comp is the pH of the relevant compartment, and pK_a was the estimated basic pK_a from ADMET Predictor v7.0 (SimulationsPlus). Replacing Eq. 7 in Eq. 6 and the subsequent equation into Eq. 5, C_u,uncharged was estimated to be

$${\text{C}}_{\text{u,uncharged}}=\frac{{\text{C}}_{\text{extracellular}}\cdot {\text{F}}_{\text{ic}}}{1+\frac{1}{{10}^{{\text{pH}}_{\text{cyto}}-\text{pKa}}}+\frac{{\text{V}}_{\text{endo}}}{{\text{V}}_{\text{cyto}}}\left(1+\frac{1}{{10}^{{\text{pH}}_{\text{endo}}-\text{pKa}}}\right)}.$$ [8]

F_endo was then calculated as

$${\text{F}}_{\text{endo}}=\frac{\frac{{\text{A}}_{\text{u,endo}}}{{\text{V}}_{\text{endo}}}}{{\text{C}}_{\text{extracellular}}}=\frac{{\text{F}}_{\text{ic}}\cdot \left(1+\frac{1}{{10}^{{\text{pH}}_{\text{endo}}-\text{pKa}}}\right)}{1+\frac{1}{{10}^{{\text{pH}}_{\text{cyto}}-\text{pKa}}}+\frac{{\text{V}}_{\text{endo}}}{{\text{V}}_{\text{cyto}}}\left(1+\frac{1}{{10}^{{\text{pH}}_{\text{endo}}-\text{pKa}}}\right)}.$$ [9]

PAMPA.

Permeability of p38α and BACE-1 inhibitors was measured in an artificial phosphatidylcholine/cholesterol membrane assay. The membranes were prepared by adding 3.5 µL of a solution of 1.8% (wt/vol) phosphatidylcholine and 1% (wt/vol) cholesterol in decane to a Millicell 96-well culture plate (0.4 µm; MilliPore Corp.). The plate was then briefly shaken, and 250 µL of 50 mM phosphate buffer with 0.5% encapsin was added to the donor side, and 100 µL of the same buffer was added to the receiver side. After shaking the plate for 45 min, compounds were added to a final concentration of 100 µM. Samples from the receiver and donor sides were collected after 3 h, and permeability was calculated as

$${\text{P}}_{\text{app}}=-\text{ln}\left(1-\frac{{\mathrm{C}}_{\text{A}}}{{\mathrm{C}}_{\text{D}}}\right)\cdot \frac{{\mathrm{V}}_{\text{A}}\cdot {\mathrm{V}}_{\text{D}}}{({\mathrm{V}}_{\text{A}}+{\mathrm{V}}_{\text{D}})\cdot \text{a}\cdot \text{t}},$$ [10]

where C_A and C_D are the compound concentrations at time t in the acceptor and donor compartments, respectively; V_A and V_D are the volumes of the acceptor and donor compartments, respectively; and a is the area of the membrane. Compounds were assayed in duplicate.

Analytical Procedures.

Compound quantification was performed using liquid chromatography coupled to tandem MS (LC-MS/MS). The system consisted of a Waters Xevo TQ MS with electrospray ionization coupled to a Waters Acquity UPLC system. Compounds were separated on a reversed phase Waters BEH C18 column (2.1 × 50 mm; 1.7 µm) at 60 °C. Mobile phase A consisted of 5% acetonitrile and 0.1% formic acid in water, and mobile phase B consisted of 0.1% formic acid in acetonitrile. Chromatographic separation comprised a 2-min gradient with a flow rate of 0.5 mL/min: (i) 0–0.5 min, 5% mobile phase B; (ii) 0.5–1.2 min, linear gradient from 5 to 90% mobile phase B; (iii) 1.2–1.6 min, 90% mobile phase B; and (iv) 1.6–1.7 min, return to initial conditions. Samples were kept at 10 °C until analysis. Sample injection volume was 5 µL. Internal standard (warfarin) was added to all samples during sample preparation. Detection of the compounds was performed in a Waters Xevo TQ MS instrument with electrospray ionization. Mass transitions and mass spectrometric conditions (ionization mode, cone voltage, and collision energy) can be found in SI Appendix, Table S1.

Protein quantification was performed with the BCA assay (Thermo Fisher Scientific) according to the manufacturer’s instructions.