The Cancer Drug Fraction of Metabolism Database

Abstract

This study aims to create a database for quantifying the fraction of metabolism of cytochrome P450 isozymes for cancer drugs approved by the US Food and Drug Administration. A reproducible data collection protocol was developed to extract essential information, including both substrate‐depletion and metabolite‐formation data from publicly available in vitro selective cytochrome P450 enzyme inhibition studies. We estimated the fraction of metabolism from the curated data. To demonstrate the utility of this database, we conducted an in vitro drug interaction prediction for the 42 cancer drugs. In the drug–drug interaction prediction, we identified 31 drug pairs with at least one cancer drug in each pair that had predicted area under concentration ratios > 2. We further found clinical drug interaction pieces of evidence in the literature to support 20 of these 31 drug–drug interaction pairs.

Study Highlights.

• WHAT IS THE CURRENT KNOWLEDGE ON THE TOPIC?

☑ The fractions of drug metabolism frequently were investigated, but currently there was a lack of fraction of metabolism database.

• WHAT QUESTION DID THIS STUDY ADDRESS?

☑ The high‐quality fraction of metabolism database of cancer drugs was created from published literature based on a well‐characterized data‐curating procedure for data sharing.

• WHAT DOES THIS STUDY ADD TO OUR KNOWLEDGE?

☑ This study demonstrated that there were significant amounts of literature that contributed to fraction of metabolism estimation. In this article, we focused on the drug metabolism related to cytochrome P450 and established a well‐characterized data‐curating procedure for constructing a pharmacology database.

• HOW MIGHT THIS CHANGE DRUG DISCOVERY, DEVELOPMENT, AND/OR THERAPEUTICS?

☑ This database was mainly designed for public data sharing that would facilitate academic research.

Drugs are eliminated by excretion or metabolism after entering the human body.1 Drug metabolism refers to the process in which the chemical structure changes under the action of various drug metabolism enzymes (especially liver drug enzymes) in the body. The majority of small‐molecule drugs are metabolized by cytochrome P450 (CYP450) enzymes, which are located in the hepatic endoplasmic reticulum.2, 3 Many factors can alter hepatic drug metabolism, including genetic polymorphisms, disease, concomitant medications, and foods.4, 5, 6 Among these factors, concomitant medications are vital because of poly‐pharmacy.7, 8, 9, 10, 11

Many enzymatic routes of elimination, including almost all of those via the CYP450 enzymes, can be inhibited or induced by concomitant medications. Notably, when the primary metabolic pathways of a drug are inhibited or induced by strong inhibitors or inducers, drug and metabolite concentrations in the blood and tissue can be significantly changed.12, 13 The dramatically changed drug exposure may result in unwanted adverse reactions or reduced efficacy.14 A randomized, open‐label, parallel‐group study indicated that after coadministration of ketoconazole for 12 days, the AUC_0–∞ of midazolam was about 6.56 times higher than midazolam alone (1,280 ng • hour/mL vs. 195 ng • hour /mL).15 Also, as a pharmacodynamic index, Symbol Digit Modalities Test score of midazolam was reduced to 34.7 from 48.3 when midazolam was coadministrated with ketoconazole, which might indicate a significant increase in midazolam‐related cognition impairment caused by drug interactions.16 These pieces of evidence showed that midazolam and ketoconazole have strong drug–drug interactions (DDIs) both in pharmacodynamics and pharmacokinetics (PK). As another example, in a randomized study, the coadministration of irbesartan (IRB) with hydrochlorothiazide (HCT; IRB/HCT 300 mg/25 mg) significantly decreased the HCT area under the curve (AUC) by 26.3% (1,373 ng • hour/mL vs. 1,087 ng • hour/mL HCT alone). At the same time, the effect of IRB on systolic blood pressure when administered with HCT was significantly different from those when IRB was administered alone. The average maximal reductions of systolic blood pressure and diastolic blood pressure when IRB was administrated alone were 9.7 ± 5.1 mmHg and 6.1 ± 2.7 mmHg, respectively. Whereas in combination with HCT, these reductions of systolic blood pressure and diastolic blood pressure reached 11.1 ± 4.3 mmHg and 7.2 ± 3.9 mmHg, which suggested a synergistic blood pressure–lowering effect for the combination that was indicated as a result of pharmacodynamic DDI.17

It is well recognized that both the precipitant's inhibition potency (e.g., K _i or IC50 for inhibition, E_max and EC50 for induction) and the fraction of metabolism (Fm) of the victim drug for each CYP450 enzyme play critical roles18, 19, 20, 21 for the effective prediction of PK DDI caused by the inhibition and induction of drug metabolism, especially through CYP450 enzymes.

The present study aims to focus on Fm data, which quantify the contribution of different metabolic pathways. There are multiple ways for Fm estimation through clinical PK studies or in vitro PK experiments. First, the contribution of a specific enzyme for a drug's metabolism can be determined as the change in AUC or clearance (CL) in the absence and presence of a coadministered selective inhibitor according to an in vivo approach.22, 23 For example, Yeung et al.24 used clinical drug interaction studies in which ketoconazole was used as the CYP3A4 index inhibitor and calculated a drug's Fm in the CYP3A4 pathway using the following equation:

$${\text{Fm}}_{3\mathrm{A}4}=1-\frac{{\text{AUC}}_{\mathrm{control}}}{{\text{AUC}}_{\mathrm{inhibited}}}$$

where AUC is the area under the concentration‐time curve of the victim drug. For example, bosutinib is metabolized primarily by CYP3A4. When coadministered with ketoconazole, its AUC ratio in plasma (AUC_inhibited/AUC_control) is 8.15. Thus, the Fm_3A4 of cinacalcet is 0.88 according to the above equation.25

Second, Fm can be estimated via a pharmacogenetics study where it can be calculated from the fold change in the exposure of a victim drug in extensive metabolizers (EMs) when compared with poor metabolizers (PMs).26 In the study by Silas et al.27 metoprolol metabolism was studied in a large population of patients. After a single 200 mg oral dose of metoprolol, the average AUC of metoprolol in the blood for the 24 hours of six CYP2D6 PMs was 7,250 ng • mL⁻¹ • hour. In the CYP2D6 EM population, the average AUC was much lower at 1,246 ng • mL⁻¹ • hour. Given that CL = Dose/AUC, thus, metoprolol's Fm_2D6 can be calculated from the following formula: Fm_2D6  = 1 − CL_PM/CL_EM = 1 − AUC_EM/AUC_PM = 0.828.

Third, in vitro ADME studies in hepatocytes, liver microsomes, or cytosol using radiolabeled substrate (¹⁴C or ³H), which measure the concentration of the unchanged radiolabeled drug and its metabolites in plasma, urine, and feces, also were regarded as a valuable clinical PK study to estimate the metabolic pathways of a drug.28, 29 For example, faldaprevir was used in an ADME study to measure the formation rates of its metabolites by various recombinant human CYP450 isoforms or CYP450‐selective chemical inhibitor by human liver microsomes (HLMs). The contribution of each CYP450 to the fraction of metabolites was determined by the rates of metabolite formation after normalization by relative liver content of each CYP450: Fm = (the rates of metabolite formation)/(relative liver content of CYP450). The results showed that the normalized contributions by rCYP3A4 were 0.94 and 0.97 for two kinds of metabolites.30

Several in vitro methods have been developed to determine the contribution of the individual enzyme in a drug's metabolism. For example, substrate depletion in HLMs is one method that the drug was incubated with or without specific CYP450 selective inhibitors.31 Comparing the metabolism rate, V_max /K _m, of the substrate without any inhibitor, the percent inhibition of a specific CYP450 pathway by the CYP450‐selective chemical inhibitor reflects the contribution of this CYP450 toward the substrate's metabolism. The chemical inhibitors should be potent, selective, and metabolically stable. In addition, a substrate‐depletion experiment can be conducted with individual hepatic recombinant human enzymes isoforms.32 This approach estimates the metabolism rate of the substrate in recombinant human CYP450 and scales the recombinant human CYP450 V_max /K _m to HLM CL_int via an relative activity factors (RAF)/ intersystem extrapolation factors (ISEF) approach,28 assuming that the CYP450s are the only metabolism enzymes. The Fm was estimated from the percentage contribution of each CYP450 enzyme toward the total HLM CL_int.

A metabolite formation study is another in vitro option. For example, after incubation, the mixtures of HLM and carbamazepine were analyzed by high‐performance liquid chromatography/mass spectrometry. The rates of carbamazepine metabolites (two‐hydroxycarbazepine and three‐hydroxycarbazepine) formation were determined in microsomes and then compared with typical CYP450 enzyme activities. The data were analyzed by nonlinear regression (GraFit 5; Erithacus Software Ltd., Surrey, UK) and linear transformation (Eadie‐Hofstee plots) to estimate the apparent PK values and enzyme models, respectively. The formation of two‐hydroxylated and three‐hydroxylated carbamazepine metabolites was evaluated in the presence or absence of known P450 inhibitors.33 HLMs from high/low CYP450 activity donors were used to estimate the inhibition percentage of carbamazepine metabolite formation. In the meantime, cDNA‐expressed isoforms were examined for the affinity of different metabolite formation.

Recently, as a result of the success of the cryopreservation of human hepatocytes,34 a hepatocyte suspension model35 became a new method to estimate Fm. Physiologically, the cryopreserved human hepatocyte is closer to human hepatic metabolism than the other in vitro system. Desbans36 used cryopreserved human hepatocytes from 12 donors to estimate Fm_3A for five prototypical CYP3A substrates with varying degrees of CYP3A‐dependent in vivo CL using intrinsic metabolic stability measurements in the presence and absence of the CYP3A probe inhibitor ketoconazole. After hepatocytes were incubated with test compounds and/or the inhibitor, the intrinsic CL was estimated from the parent compound depletion profile. Then Fm_3A was calculated from the ratio between CL_int in the absence and presence of ketoconazole as:

$${\text{Fm}}_{3\mathrm{A}}=1-\frac{{\text{CL}}_{\mathrm{int}}\text{with ketoconazole}}{{\text{CL}}_{\mathrm{int}}\text{without ketoconazole}}$$

Although there are some widely used databases for drug metabolism—such as (i) DrugBank,37 which is a comprehensive database that combines detailed drug (e.g., chemical, pharmacological, and pharmaceutical) data with comprehensive drug target information (e.g., sequence, structure, and pathway); (ii) Transformer (the former Super CYP450),38 which integrates CYP450 enzyme interactions and some pharmacological information; and (iii) DIDB,39 which can evaluate the impact of DDI in the clinic by in vitro and in vivo DDI data—limited Fm data can be found in these databases.

In this article, we present our initial effort in developing the Fm database on CYP450. As a demonstration example, we primarily focus on small‐molecule cancer drugs because of their broad applications. We further focus on the hepatic CYP450‐based Fm calculation, as hepatic CYP450 enzymes are the most widely investigated. Therefore, in our data‐curating study, Fm data are collected and estimated from in vitro inhibition studies of HLM, including metabolite‐formation and substrate‐depletion studies. The selection for this in vitro system was mainly because of its rich data sources. Other data types from clinical studies (e.g., pharmacogenetics (PGx) PK studies, or clinical PK drug interaction studies) or in vitro experiments (i.e., cytosol or hepatocyte studies) for Fm estimation were also investigated in our data‐curation process. However, these data sources turned out to be highly sparse. They will be further explored in the future.

Materials and Methods

The Fm database was curated from published articles in the PubMed index (http://www.ncbi.nlm.nih.gov/pubme​d/). An overview of the data collection is available in Figure  1 . In cancer drug selection, 237 cancer drugs approved by the US Food and Drug Administration (FDA; Table S1 ) were identified in DrugBank and the National Cancer Institute. We mainly focused on the 188 small‐molecule cancer drugs. The next stage is the keyword search, including cancer drug names, “CYP450,” “HLMs,” and/or “metabolism.” If the drug has specific CYP450 metabolism information in DrugBank, the keyword may be composed of the drug name and the specific CYP450. Cancer drug generic names and their synonyms and brand names published in DrugBank were included in the search. Similarly, CYP450 enzyme names include their synonymous names in the HUGO Gene Nomenclature Committee.

Figure 1.

The flowchart of data‐collection procedure. Fm, fraction of metabolism.

All of the identified PubMed papers were then filtered. In each abstract, we checked whether this was a drug metabolism study, whether the contribution of CYP450 to the metabolism were investigated, or whether the CYP450 enzyme inhibitors was discussed in the abstracts. If neither of these pieces of evidence was reported in the abstract, this paper was then removed. If an abstract passed the filtering step, its full text received a further examination. In the article Method section, substrate depletion and CYP450 enzyme inhibition in HLM study information or the information in a metabolite‐formation study were examined. In particular, in the metabolite‐formation studies, several experiment details were required. All incubations were performed at 37°C; HLMs were incubated with sodium phosphate buffer and NADPH (usually NADP + glucose‐6‐phosphate dehydrogenase) before adding the substrates, and the drug was incubated with pooled HLMs in the presence or absence of P450 selective chemical inhibitors for several minutes; control samples containing no NADPH or substrates were also included. Incubations were carried out for a defined time, and at the final stage, the drug metabolites were evaluated using high‐performance liquid chromatography/LC‐MS/MS in the presence or absence of known CYP450 enzyme selective inhibitors. In the substrate‐depletion studies, several experiment details were required to report. Preincubation and incubation procedures were the same as the metabolite‐formation studies, and those that had no NADPH or substrate samples for control groups were also needed. After being vortexed, the concentration of the remaining parent drugs in the supernatants were measured using the high‐performance liquid chromatography assay in the presence or absence of selective inhibitors. If this information was not reported in the Method section for either a substrate‐depletion study or a metabolite‐formation study, the paper was then removed.

The Results section was then reviewed carefully, including the narrative, figures and tables (including their legends), and the supplemental materials if available. We collected data from the paper and calculated the corresponding Fms. The following are Fm calculation steps for the metabolite‐formation data:

1. The following data elements were extracted from the results: the metabolites for each drug and/or their relative contribution of the metabolisms, the percentage of inhibitions for each CYP450 enzyme, and their related CYP450 inhibitor. (Sometimes this percentage was directly reported. Otherwise, the metabolism rates under inhibition and control were reported.)

2. The formation of metabolites by HLMs estimated in the presence or absence (i.e., a control sample) of known P450 isoform‐selective inhibitors. The relative proportion of the metabolite formation in the reaction mixtures with no inhibitor was set as 100%. If the inhibition percentage of the drug metabolite could not be collected directly from the paper, the proportion was estimated from the changes in the portion of control metabolite that was calculated relative to incubations with no inhibitor.

3. If multiple metabolites were found to be the predominant drug metabolic pathways (such as exemestane, tamoxifen, toremifene, lapatinib), the contribution of each metabolite in the total drug metabolism was calculated.

4. If there existed several competitive inhibitors for a specific CYP450 enzyme, the percent of inhibition for the inhibitor was calculated. The mean value was calculated and taken as the inhibition percentage when there were multiple inhibitors.

5. If the substrate concentration of the inhibition experiment varied in the paper, the rate of inhibition at each concentration was calculated. Then their mean value was calculated.

6. If there were several HLM samples in the experiment, the percentage of inhibition at each sample was calculated and the mean value was taken (such as fluorouracil).

7. The total inhibition percentage of one metabolite was normalized. The fraction of metabolite for the ith CYP450 enzyme in the jth metabolite is:

$${\text{Fm}}_{m_{i\mathrm{j}}}=\frac{{\text{inhibition}}_i}{{\sum }_i{\text{inhibition}}_i}\times \text{percentage of metabolite}j$$

where inhibition_i refers to the percentage of inhibition for the ith enzyme. The sum of Fm_ij over all metabolites is regarded as the fraction of metabolized for enzyme i.

For example, two metabolite pathways of tramadol were reported in 40 with CYP450 information. According to the data collection pipeline we proposed previously, we can collect a proportion of the contribution of each pathway and normalize:

Pathway	% Contribution	Normalized
M1	32.881	33.532
M2	65.18	66.471

Then CYP450 inhibition information can be collected from each metabolite pathway and normalized by metabolite pathway contribution:

Inhibitor	CYP450	Pathway	% Inhibition	Normalized
Furafylline	1A2	M1	37	7.381
Diethyldithiocarbamate	2E1	M1	13	2.592
Troleandomycin	3A4	M1	16	3.193
Sulfaphenazole	2C9	M1	10	2
Quinidine	2D6	M1	81	16.171
S‐Mephenytoin	2C19	M1	11	2.2
Furafylline	1A2	M2	13	6.701
Diethyldithiocarbamate	2E1	M2	15	7.732
Troleandomycin	3A4	M2	61	31.431
Sulfaphenazole	2C9	M2	12	6.18
Quinidine	2D6	M2	15	7.731
S‐Mephenytoin	2C19	M2	13	6.7

Finally, Fm is estimated as:

Drug	Pathway	1A2	2C9	2C19	2D6	2E1	3A4
Tramadol	M1	7.381	2	2.2	16.171	2.592	3.193
	M2	6.701	6.18	6.7	7.732	7.732	31.431
	Total	14.082	8.18	8.9	23.903	10.324	34.634

On the other hand, Fm can be calculated from the substrate‐depletion study.

1. The inhibition percentage or the remaining proportions of substrates was evaluated in the presence or absence (i.e., a control sample) of known CYP450 isoform‐selective inhibitors. It was assumed that the sum of inhibition percentage of a specific substrate and its remaining proportion equaled to 100%. If the inhibition percentage of the drug was not collected directly in the paper, the ratio was estimated from the changes of a control sample or the remaining percentage of the substrate.

2. If a CYP450 enzyme was inhibited by several inhibitors simultaneously, the mean percent of inhibition for all inhibitors was chosen.

3. If there were two or more substrate concentrations of the inhibition experiment, we took the mean percentage of inhibition at all concentrations for estimation.

4. The metabolized fraction for the ith CYP450 enzyme is:

$${\text{Fm}}_i=\frac{{\text{inhibition}}_i}{{\sum }_i{\text{inhibition}}_i}$$

where inhibition_i refers to the percentage of inhibition for the ith enzyme.

To ensure data integrity, two curators with biology backgrounds conducted the data‐curating process. Dr Sara K. Quinney, who has an extensive pharmacology training background further checked any differentially annotated abstracts. Then the data extraction from the full text was carried out by those two annotators again. Among the disagreed data collection between these two annotators, a group review was conducted by Drs Wang, Quinney, and Li to reach the final agreement.

The data‐validation step was conducted, and two sets of drugs were used for validation. The part of the drugs with Fm data was selected randomly as the positive set, and their Fm data were reevaluated. Some cancer‐related drugs without Fm data were selected randomly as the negative set and reevaluated. During this validation process, two independent annotators went through the entire Fm data‐curating process for these 16 drugs. These two validation annotators have masters or doctorate degrees in computational biology and/or biology. Drs Quinney and Li, who have pharmacology backgrounds, further evaluated the concordance among these three sets of annotations. The consistency of the Fm data is reported later.

Results

Among 237 cancer‐related drugs (Table S1 ), there are 42 drugs that have identified and curated Fm data (Table S2 ) from the PubMed literature. During the validation process, among the eight negative cancer‐related drugs (i.e., no Fm data detected from the first annotator), the other two validation annotators also did not find their Fm data. Therefore, the validation accuracy rate is 100% for the negative drug set. Among those eight positive cancer drugs, the first validation annotator managed to find Fm data for seven of eight drugs, and the second validation annotator found the Fm data for all eight drugs. The overlap rate is 93.8% for the positive drug set. However, the consistency of the Fm value was not good. The first validation annotator had consistent Fm calculations for six of eight drugs, 75%. The inconsistency was primarily attributed to one completely missed drug and another one that was miscalculated. The second validation annotator had only 62.5% of Fm values that were in concordance with the original annotator because he only considered the top one or two major CYP450 metabolic pathways and ignored the other minor pathways. Finally, among the discordant Fm values for the positive cancer drugs, their full‐text papers and Fm values were further evaluated by two additional pharmacologists. They found the original annotator had the highest accuracy in calculating the Fm. She was right on 90% of the discordant values among the three annotators. The validation results are shown in Figure  2 .

Figure 2.

The validation results: (a) The consistency evaluation results between original annotator and validation annotator1; (b) The consistency evaluation results between original annotator and validation annotator2; (c) The validation results between original annotator and finalized annotator. the different colors represent different cytochrome P450 enzymes, and the different shapes represent different drugs. The flash line represents the equality of the horizontal and vertical axes. The horizontal axis represents original annotator, and the vertical axes represent validation annotator 1, validation annotator 2, and finalized annotation.

All of the extracted and validated Fm data are presented in Table S2 . In this data set, the key data elements included the drug names, PubMed unique identifier, and Fms for various CYP450 and their corresponding inhibitors. Especially note that column C annotates whether the data were from the substrate‐depletion study or the metabolite formation study, and columns D–P display Fm data for various CYP450 pathways. The numbers represent the percentiles, and a blank cell indicates that the corresponding CYP450 pathway was not being investigated. Please keep in mind that in some experiments, there existed more than one inhibitor for one enzyme, and we annotate them all in the database.

One important utility of the Fm database was to predict drug interactions. These data characterized all of the hepatic CYP450 metabolic pathways and their contributions in predicting drug interactions. We therefore further explored and predicted the drug interactions between these 42 cancer drugs (i.e., substrates) and 178 drugs (i.e., inhibitors), in which their inhibitions on CYP450 were known, in the following case study.

AUC ratio (AUCR) is the key parameter to measure drug interaction. The predicted AUCR estimate as follows41:

$$\text{AUCR}=\frac{{\text{AUC}}_i}{\text{AUC}}=\frac{1}{(1-f_e){\sum }_n^{i=1}{\text{Fm}}_i\times \frac{1}{1+{\sum }_J^{j=1}\frac{[I_j]}{K_{i,j}}}+f_e}$$

where AUC_i/AUC is the ratio of the area under the plasma concentration‐time profile of the substrate drug in the presence (AUC_i) and absence (AUC) of the inhibitor drug; Fm is the fraction of the total hepatic metabolism mediated through a CYP450 enzyme (from our Fm database), f _e is the fraction of the unchanged drug via urine. I _u is the unbound inhibitor concentration, K _i,u is the unbound inhibition constant. In this paper, I _u equals to C_max  × f _u, where C_max is the maximum concentration, and f _u is the fraction of unbound drug in plasma. In this paper, we have 42 cancer drugs with Fm and f _e 42 there are 178 drugs (16 cancer drugs and 162 noncancer drugs where there was nonoverlap with 237 cancer drugs) with K _i and I _u, with K _i obtained from published literature and I _u collected from published clinical studies, and comprehensive pharmacoinformatics databases such as DIDB and Goodman and Gilman's The Pharmacological Basis of Therapeutics 43 where most of the data were collected through the PubMed literature review.

Following the FDA DDI guideline44 and expert experience, an AUCR > 1.25 is regarded as PK DDI evidence. Balancing recommendations provided by regulatory agencies and the scientific community, we define AUCR > 2 as evidence of significant PK DDI. Based on our steady‐state model‐based DDI predictions, we find 31 drug pairs (the substrate is a cancer drug, and the inhibitor is a noncancer drug) with an AUCR > 2. After having been validated in DrugBank, Drugs.com and PubMed, 15 drug pairs have supporting clear clinical PK DDI evidence in these databases. There are only four inhibitors (acetaminophen, amprenavir, ritonavir, and indinavir) in Table 1 because the K _i on the CYP3A4 of these four drugs (1.855 μM, 0.79 μM, 0.18 μM, and 3.51 μM) were less than 10 μM, a value considered in this study as a cutoff that indicated potential inhibition to the CYP3A4 substrate.

Table 1.

The validation results of drug–drug interaction prediction based on steady‐state model for AUC

Substrate	Inhibitor	AUCR	DrugBank	Drugs.com	Pubmed (clinical)
Fenretinide	Acetaminophen	11.19	—	—	The fenretinide plasma level increased sharply after ceftriaxone and acetaminophen were initiated. Kang et al.45,
Aprepitant	Acetaminophen	5.75	The serum concentration of aprepitant can be increased when it is combined with acetaminophen.	—	—
Bosutinib	Acetaminophen	4.68	The serum concentration of bosutinib can be increased when it is combined with acetaminophen.	—	—
Vinblastine	Amprenavir	3.06		Combining these medications may significantly increase the blood levels and effects of vinblastine.	—
Trabectedin	Acetaminophen	3	The metabolism of trabectedin can be increased when combined with acetaminophen.	Trabectedin may cause liver problems, and using it with other medications that can also affect the liver such as acetaminophen may increase that risk.	—
Everolimus	Acetaminophen	2.39	The serum concentration of everolimus can be increased when it is combined with acetaminophen.	—	—
Aprepitant	Amprenavir	2.3	Amprenavir may increase the blood levels and effects of aprepitant.	Amprenavir may increase the blood levels and effects of aprepitant.	—
Vinblastine	Ritonavir	2.24	The serum concentration of vinblastine can be increased when it is combined with ritonavir.	Combining these medications may significantly increase the blood levels and effects of vinblastine.	There are multiple case reports describing serious infectious complications in patients receiving vinblastine and ritonavir. Ezzat et al.46
Gefitinib	Amprenavir	2.21		Amprenavir may increase the blood levels and effects of gefitinib.	—
Bosutinib	Amprenavir	2.17	The metabolism of bosutinib can be decreased when combined with amprenavir.	Amprenavir may significantly increase the blood levels of bosutinib.	—
Vincristine	Acetaminophen	2.14	The excretion of vincristine can be decreased when combined with acetaminophen.	—	—
Axitinib	Amprenavir	2.11	The metabolism of trabectedin can be increased when combined with acetaminophen.	Amprenavir may significantly increase the blood levels of axitinib. This may increase side effects such as high blood pressure, diarrhea, nausea, vomiting, constipation, decreased appetite, weight loss, and rash, itching, or peeling of skin on the hands and feet.	—
Vinblastine	Indinavir	2.09	The metabolism of vinblastine can be decreased when combined with indinavir.	Combining these medications may significantly increase the blood levels and effects of vinblastine.	—
Tamoxifen	Acetaminophen	2.07	The serum concentration of tamoxifen can be increased when it is combined with acetaminophen.	—	—
Vinorelbine	Amprenavir	2.06	The metabolism of vinorelbine can be decreased when combined with amprenavir.	Combining these medications may significantly increase the blood levels and effects of vinorelbine.	—

AUCR, AUC ratio.

Information is from a case report.

Discussion

The primary goal of this paper is to provide a database of Fm. In this cancer drug Fm database, the relative contributions of the CYP450 isozyme were curated and calculated from drug metabolism studies using the HLMs. First, relevant papers were identified through the keyword search in PubMed. Second, only the articles with relevant in vitro experiment information were selected. These in vitro experiments included both metabolite formation and substrate‐depletion studies. In this database, there are four drugs (exemestane, tamoxifen, toremifene, lapatinib) whose Fms were calculated from their metabolite‐formation experiments, and the other 38 drugs were estimated from substrate‐depletion studies. A predefined data‐curating protocol was established to assure data quality and data reproducibility. Multiple annotators were employed in the data‐filtering, data‐curating, and data‐validation stages. Two independent validation annotators reevaluated the randomly selected drugs from positive and negative sets. The overlap rates were 93.8% and 100% for the positive and negative sets, respectively. After being evaluated by two additional pharmacologists, the original annotation has the highest accuracy in collecting and calculating the Fm (Figure  2 ).

This paper, for the first time, demonstrates a well‐characterized data‐curating procedure in constructing a pharmacology database. Although there are many pharmacology databases, there is limited published evidence on how data annotation and validation are conducted, and most importantly, where the potential errors come from and how these errors are identified and corrected during the data‐curating process. To our knowledge, the best practice is a well predefined data‐curating guideline. The more accurate and detailed it is, the less likely it is the annotators will make mistakes. A double‐annotation process is also a great way to identify potential errors. We believe this is one of the primary contributions in this paper.

In our initial effort to curate and calculate Fm data from the literature, we chose to focus on primarily the Fm data from the HLM inhibition experiments. These experiments are suitable for studying the relative contribution of CYP450 isozymes to drug metabolism. If we want to evaluate the contribution of other enzymes (e.g., uridine 5’‐diphospho‐glucuronosyltransferase) to drug metabolism, we will need to explore hepatocyte or cytosol studies. Although we focused on HLM studies, which can provide us with abundant data, their Fms are limited to CYP450 metabolism enzymes. In the future, we will further expand Fm data collections to the hepatocyte or cytosol studies.

Publication bias may influence the quantity and quality of the available data. We only search literature from PubMed, not considering other resources. All of the drugs did not have in vitro data. We also recognize that there is no selective inhibitor available in vitro for CYP2C19 and CYP2B6. Also, there is a lack of discussion for the fraction of the renal CL with the unchanged drug and the inducer. Our Fm estimates may not be entirely accurate. Furthermore, in our Fm estimation, we assumed a total inhibition of the in vitro experiment, which is also questionable in reality. It may lead to biased Fm estimates.

There are additional data, such as pharmacogenetics and drug interaction clinical PK studies, for Fm estimation. These studies could provide more accurate data than in vitro studies. According to PharmGKB, there are two drugs among 42 cancer drugs—docetaxel and paclitaxel—where Fm can be estimated from genetic polymorphisms. The related literature about the CYP450 variant can find in PharmGKB, and we can calculate the contribution of the CYP450 enzyme via the pharmacogenetics research from the fold change in the exposure of a victim drug in the EMs when compared with PMs. The contribution of CYP2C8 in paclitaxel is more consistent when compared with the Fm in vitro (Table  2 ). This result indicates in vitro Fm is reliable and further indicates that the Fm estimated from PGx studies are sparse.

Table 2.

The comparative data between in vitro and pharmacogenetics

Drug name	CYP	In vitro calculated		Pharmacogenetics calculated
		Fm, %	PMID	Fm, %	PMID
Docetaxel	3A4	70.7	8640817	42.1	18509327
Paclitaxel	2C8	58.2	903909	50.0	17092739

CYP, cytochrome P450; Fm, fraction of metabolism; PMID, PubMed unique identifier.

This Fm database is designed mainly for public data sharing and specifically for academic research on FDA‐approved cancer drugs. It is not necessarily the best resource for the pharmaceutical industry. Many drug companies, especially large companies, usually have their experiments and/or own data sources for Fm calculation. However, their data are generally for internal usage and drug development but are not accessible to academic investigators. Our effort in developing the Fm database is a significant first step for sharing data. We hope others contribute their findings to enhance our database.

In summary, our paper demonstrates that there are significant amounts of literature that contribute to Fm estimation. Using the informatics approach, we have successfully identified Fm data for 42 of 237 cancer drugs. In this paper, we focus on the drug metabolism data related to CYP450, and these data are generated from either metabolite formation studies or substrate‐depletion studies. To create a more comprehensive Fm database, we shall further consider alternative metabolism pathways, such as uridine 5’‐diphospho‐glucuronosyltransferases; alternative in vitro experiments, such as hepatocytes; and alternative clinical data, such as pharmacogenetic PK studies or clinical DDI studies. In addition, we will develop annotation guidelines for each task and conduct literature reviews and validation analyses afterward.

Funding

This work has been supported by several National Institutes of Health grants (DK102694, GM10448301‐A1, RO1 GM117206, and R01LM011945) and National Science Foundation Grant NSF1622526. It has also been supported by the Natural Science Foundation of Heilongjiang Province, China (QC2016086) and the Fundamental Research Funds for the Central Universities, China (HEUCF160420).

Conflict of Interest

The authors declared no competing interests for this work. As an Associate Editor for CPT: Pharmacometrics & Systems Pharmacology, Lang Li was not involved in the review or decision process for this article.

Author Contributions

L.H., S.K.Q., L.C. and L.W. wrote the manuscript. L.L., W.C. and C.‐W.C. designed the research. L.H., F.W. and C.‐W. C. performed the research. L.H., X.W., J.L. and L.W. analyzed the data.

Supporting information

Table S1. Drug list for 237 cancer drugs.

Table S2. Fraction of drug metabolism.

Application: AUCR.xlsx: The predicted AUCR results, fe.txt: The fraction of unchanged drug via urine, .txt: The fraction of drug metabolism, kui.txt: The unbound inhibition constant of inhibitors, Model code.docx: The code was used for AUCR prediction.