- Open Access
Structural similarity assessment for drug sensitivity prediction in cancer
© Shivakumar and Krauthammer; licensee BioMed Central Ltd. 2009
- Published: 17 September 2009
The ability to predict drug sensitivity in cancer is one of the exciting promises of pharmacogenomic research. Several groups have demonstrated the ability to predict drug sensitivity by integrating chemo-sensitivity data and associated gene expression measurements from large anti-cancer drug screens such as NCI-60. The general approach is based on comparing gene expression measurements from sensitive and resistant cancer cell lines and deriving drug sensitivity profiles consisting of lists of genes whose expression is predictive of response to a drug. Importantly, it has been shown that such profiles are generic and can be applied to cancer cell lines that are not part of the anti-cancer screen. However, one limitation is that the profiles can not be generated for untested drugs (i.e., drugs that are not part of an anti-cancer drug screen). In this work, we propose using an existing drug sensitivity profile for drug A as a substitute for an untested drug B given high structural similarities between drugs A and B.
We first show that structural similarity between pairs of compounds in the NCI-60 dataset highly correlates with the similarity between their activities across the cancer cell lines. This result shows that structurally similar drugs can be expected to have a similar effect on cancer cell lines. We next set out to test our hypothesis that we can use existing drug sensitivity profiles as substitute profiles for untested drugs. In a cross-validation experiment, we found that the use of substitute profiles is possible without a significant loss of prediction accuracy if the substitute profile was generated from a compound with high structural similarity to the untested compound.
Anti-cancer drug screens are a valuable resource for generating omics-based drug sensitivity profiles. We show that it is possible to extend the usefulness of existing screens to untested drugs by deriving substitute sensitivity profiles from structurally similar drugs part of the screen.
- Cancer Cell Line
- Prediction Accuracy
- Response Prediction
- Similar Drug
- Sensitivity Profile
In the last decade, cancer treatment has seen a shift from a "one size fits all" philosophy to a more personalized approach. Technical advances allow for the assessment of complex genetic defects and pathway aberrations, enabling refined cancer classification and treatment prediction based on molecular rather than histological features. Gene expression profiling has been used to classify patient samples as being benign or malignant or to classify them into cancer subclasses [1–3]. Lately, molecular profiling of cancer samples has been applied to the prediction of drug sensitivity . Drug-sensitive samples have been compared to drug-resistant samples to build "drug sensitivity profiles", statistical models built from a defined set of genes whose differential expression in a cell line may confer sensitivity to a drug. Thus, cancer cell lines whose gene expression patterns are similar to the genes in the sensitivity profile will have a higher probability of responding to the drug. It has been demonstrated that gene expression profiling can refine the prediction of drug response to targeted drug therapies. For example, Harris et al. used gene expression profiling to identify lists of genes that are potential predictors of response to Herceptin and Vinorelbine in HER2-positive breast cancers . Another study used the COXEN (CoExpression Extrapolation) algorithm to build sensitivity profiles for the drugs cisplatin and paclitaxel . The profiles were subsequently used to predict drug response in bladder and breast cancer samples. Of particular interest is the fact that these samples were not part of the original anti-cancer cancer panel, showing that drug sensitivity profiles generated from the screen can be generalized to any tumor sample.
Approaches such as COXEN make use of comprehensive pharmacogenomic resources to derive drug sensitivity profiles. COXEN was applied to the NCI-60 (National Cancer Institute) anti-cancer screen, which tested >40,000 compounds on 60 cancer cell lines, all of which have been profiled (in a separate effort) for genome-wide gene expression [6, 7]. Sensitivity profiles for the compounds in the screen can be generated by comparing gene expression values of sensitive and resistant cell lines.
The COXEN algorithm (and similar approaches) can only generate profiles for compounds that are part of an anti-cancer drug screen . If a drug was not tested as part of the NCI-60 dataset, its sensitivity profile cannot be generated. Updating an existing cancer screen with the latest available or experimental drugs is a non-trivial issue, and requires the same expertise, infrastructure and conditions as when the screen was established the first time around.
In this work, we present a method to predict responses to drugs that are not part of available anti-cancer screens. We propose to find substitute sensitivity profiles for these untested drugs by using structural similarity.
It has been known that the structure of a compound is related to its activity (QSAR – Quantitative Structure Activity Relationship). Shi et al. performed an analysis where 131 compounds (whose mechanisms of action were known) were clustered based on activity across cancer cell lines . They found that compounds with similar structures clustered together. Several other studies have been performed to show that specific classes of compounds (eg. Taxols) have similar activity patterns across cancer cell lines . Therefore, it can be hypothesized that structurally similar compounds have similar activities and might affect the same pathways. In other words, the sensitivity pattern for drug A (not part of the screen) might be similar to that of drug B (part of the screen) if drug A and drug B are structurally similar. We are interested in extending this idea to test whether the drug sensitivity profiles of a structural analogue of a drug can be used as a substitute when predicting its response in cancer cell lines.
To analyze whether a structural analogue of a drug can be used to predict its response, two questions need to be answered. The first question is whether two structurally similar drugs have the same sensitivity pattern across the NCI-60 cancer cell lines. Secondly, even if we can show that there is a strong association between structural similarity and drug sensitivity, we need to test whether the correlation is strong enough to be of any practicality. The second question we would like to ask is whether the sensitivity profile of drug A be used to predict activity of drug B without a significant loss of accuracy, if drug A and drug B are structurally similar. We investigated whether this proposition holds only for compounds in the same chemical family, or also for compounds that are similar but from different families.
Structural similarity and response similarity in NCI-60 cancer cell lines
The first question we set out to address was whether two structurally similar drugs have the same sensitivity pattern across cancer cell lines in an anti-cancer screen, even when the mechanisms of action or the classes of the compounds might not be known. We used the Tanimoto coefficient between pairs of compounds as the structural similarity measure (see Methods for more details).
Predictive power of sensitivity profiles of structurally similar drugs
The second question we wanted to ask is whether we can use one drug as a substitute to perform gene-expression-based sensitivity prediction for another drug. In this scenario, we take advantage of an existing anti-cancer screen, such as the NCI-60, to construct a sensitivity profile for a drug A, and use it as a substitute profile for another drug B. Specifically, we wanted to see what the loss of accuracy in response prediction would be if the sensitivity profile was constructed from a drug with high structural similarity. Using a cross-validation setup within the NCI-60 data set, we could measure the change in prediction accuracy when using drug B's sensitivity profile compared to using a substitute profile generated from drugs A with various levels of structural similarity to B.
Families of compounds in our dataset
Cancer patients need access to the latest decision criteria given an increasing number of highly specific and targeted anti-cancer drugs. It is believed that the pre-treatment cellular state is predictive of whether a cancer cell is going to respond to a particular drug. Functional anti-cancer drug screens give access to pre-treatment gene expression states from sensitive and resistant cancer cells, and are useful in deriving drug sensitivity profiles that can be applied to independent cancer samples. One major limitation of this approach is that profiles can only be generated for drugs that have been tested in anti-cancer screens. In our work, we addressed the issue of deriving sensitivity profiles for untested drugs for which we do not (yet) have response information across a number of cell lines. For those drugs, we examined the possibility of using substitute drug sensitivity profiles from structurally similar drugs that have been tested in the screen.
We established that it is indeed true that structural similarity is highly correlated with sensitivity across cancer cell lines. We have also shown that for structurally related drugs, the use of substitute sensitivity profiles is likely to result in response predictions similar to using a drug's own profile in the first place. Our study may thus be helpful in increasing the relevance of existing screens, by expanding their use to other drugs (not part of the screen) with high structural similarity. Since not all new cancer drugs are immediately tested in anti-cancer screens, our approach ultimately increases the number of drugs for which patient response can be predicted.
A drawback of the current study is the small number of drugs that we could include in our analysis, i.e. drugs that showed sufficient variation across cell lines. This introduces some bias that may favor pairs of compounds in the same class and with high similarity in activity patterns. Nonetheless, the above experiments demonstrate that there is a strong correlation between structural similarity and drug sensitivity in the NCI-60 cancer cell lines.
In terms of future work, we want to explore the following issues. First, we want to fine-tune our sensitivity profile generating algorithm, in order to increase the prediction accuracy when using a compound's own profile (which currently averages at 74%). Second, we want to explore other structural similarity measures to ensure that we indeed identify the closest structural pairs among the NCI-60 compounds. Third, we want to investigate other similarity measures (based on function or mechanism) that can identify an analogue for sensitivity profile substitution.
The NCI-60 is an anti-cancer screen composed of over 40,000 compounds tested on 60 cancer cell lines from 9 different tissue types. We downloaded the compounds' log(GI50) values (concentration of compound needed for 50% growth inhibition) from http://dtp.nci.nih.gov. Gene expression values for the 60 untreated cancer cell lines were obtained from Gene Expression Omnibus (Accession GDS1761) .
Standard deviation of the log(GI50) values of the compound across the 60 cancer cell lines is at least 0.625
At least 10 of the 60 cell lines are sensitive
At least 10 of the 60 cell lines are resistant
At least 30 of the 60 cell lines are either sensitive or resistant (do not fall in the intermediate category)
compounds matched the aforementioned criteria, and were used in our analysis.
Structural similarity and response similarity in NCI-60 cancer cell lines
Structural similarity between compounds
where P is the set of molecular fragments taken into account, and ϕ G (p) is a bit value representing the presence of a molecular fragment p in the graph representation of a compound, G. Chemcpp outputs the normalized Tanimoto coefficients between all pairs of compounds in the dataset, thus giving a relative measure of similarity among them.
Response similarity between compounds
For each pair of compounds in our dataset, the similarity between their responses across the 60 cell lines was calculated as the percentage of cell lines that responded similarly to (either both sensitive or both resistant) both the compounds. Cell lines classified as intermediate for either of the compounds were not included in the calculation. Figure 1 shows Tanimoto coefficients plotted along response similarities, to demonstrate the correlation between structural similarity and response similarity in the NCI-60 cancer cell lines.
Predictive power of sensitivity profiles of structurally similar drugs
If a drug has not been tested as part of a screen, we wanted to see how much accuracy would have to be compromised if a structurally similar drug was used as a substitute to predict response. This second question involved generating the sensitivity profiles for each of the 244 compounds in our dataset. We then used the sensitivity profile of a drug A as a substitute for drug B, recorded the structural similarity between A and B, and measured the drop in prediction accuracy compared to using drug B's sensitivity profile.
Sensitivity profile generating algorithm
For each of the 244 compounds, the log(GI50) data and gene expression data were integrated to generate sensitivity profiles (a statistical model based on a list of differentially expressed genes) which were then used to predict the response of a cancer cell line to one of the compounds. For each compound, the response of each of the cell lines in the NCI-60 panel was predicted in a leave-one-out cross-validation setting. Specifically, the remaining 59 cell lines were used as a training set to predict the response of the compound to the 60th cell line using a generalized linear model. The steps of our algorithm were as follows
Using each one of the 60 cell lines as the testing set, the remaining 59 cell lines were assigned as the training set.
The 59 cell lines were separated into those sensitive and resistant to the compound.
Using the expression values of 9706 genes across the cell lines, a t-test was used to identify the 10 most differentially expressed genes between the sensitive and resistant sets. These 10 genes were identified as the 'sensitivity profile' genes for the compound.
A generalized linear model was used in Matlab to predict the response of the 60th cell line, using the expression values of the 10 genes that were part of the sensitivity profile .
Structural similarity and sensitivity prediction
The sensitivity profile was generated for all 244 compounds using NCI-60 data.
For each drug A, 60 responses were predicted for the NCI-60 cell lines using the drug's profile. The predicted responses were compared to drug A's experimental responses from NCI-60 to calculate an overall accuracy for drug A, which we termed AccuracyA using A's profile
The profiles of the remaining 243 drugs (drug B) were used to predict responses of drug A to the 60 cancer cell lines. These predicted responses were compared to drug A's experimental responses to calculate AccuracyA using B's profile
The loss in prediction accuracy was calculated as the absolute difference between AccuracyA using A's profile and AccuracyA using B's profile. This difference can be thought to represent the loss of accuracy when using another drug's profile as a substitute in response prediction
Figure 2 shows Tanimoto coefficients plotted along loss in prediction accuracies, to demonstrate the relationship between structural similarity and predictive value of substitutive sensitivity profiles in the NCI-60 dataset.
Families of compounds in our dataset
To analyze the families of the compounds in our dataset, we first annotated all 244 compounds with their generic or IUPAC names, downloaded from Pubchem . Using these annotations, we were able to identify broad families for most of the compounds. Using the Tanimoto structural similarity matrix calculated by Chemcpp, we calculated the first two principal components of each of the compounds, using Matlab's princomp function. We used these two components as the x and y coordinates for visualizing the clustering of the compounds in Figure 3.
This study was supported by the Yale SPORE in skin cancer. We would also like to thank Nam Tran and Joshua Swamidass for his valuable input.
This article has been published as part of BMC Bioinformatics Volume 10 Supplement 9, 2009: Proceedings of the 2009 AMIA Summit on Translational Bioinformatics. The full contents of the supplement are available online at http://www.biomedcentral.com/1471-2105/10?issue=S9.
- Golub TR, Slonim DK, Tamayo P, Huard C, Gaasenbeek M, Mesirov JP, et al.: Molecular Classification of Cancer: Class Discovery and Class Prediction by Gene Expression Monitoring. Science 1999, 286(5439):531–537. 10.1126/science.286.5439.531View ArticlePubMedGoogle Scholar
- Khan J, Wei JS, Ringner M, Saal LH, Ladanyi M, Westermann F, et al.: Classification and diagnostic prediction of cancers using gene expression profiling and artificial neural networks. Nat Med 2001, 7(6):673–679. 10.1038/89044PubMed CentralView ArticlePubMedGoogle Scholar
- Sørlie T, Perou CM, Tibshirani R, Aas T, Geisler S, Johnsen H, et al.: Gene expression patterns of breast carcinomas distinguish tumor subclasses with clinical implications. Proc Natl Acad Sci USA 2001, 98(19):10869–74. 10.1073/pnas.191367098PubMed CentralView ArticlePubMedGoogle Scholar
- Lee JK, Havaleshko DM, Cho H, Weinstein JN, Kaldjian EP, Karpovich J, et al.: A strategy for predicting the chemosensitivity of human cancers and its application to drug discovery. Proc Natl Acad Sci USA 2007, 104(32):13086–13091. 10.1073/pnas.0610292104PubMed CentralView ArticlePubMedGoogle Scholar
- Harris LN, You F, Schnitt SJ, Witkiewicz A, Lu X, Sgroi D, et al.: Predictors of Resistance to Preoperative Trastuzumab and Vinorelbine for HER2-Positive Early Breast Cancer. Clin Cancer Res 2007, 13(4):1198–1207. 10.1158/1078-0432.CCR-06-1304View ArticlePubMedGoogle Scholar
- Monks A, Scudiero DA, Johnson GS, Paull KD, Sausville EA: The NCI anti-cancer drug screen: a smart screen to identify effectors of novel targets. Anticancer Drug Des 1997, 12(7):533–41.PubMedGoogle Scholar
- Ross DT, Scherf U, Eisen MB, Perou CM, Rees C, Spellman P, et al.: Systematic variation in gene expression patterns in human cancer cell lines. Nat Genet 2000, 24(3):227–35. 10.1038/73432View ArticlePubMedGoogle Scholar
- Staunton JE, Slonim DK, Coller HA, Tamayo P, Angelo MJ, Park J, et al.: Chemosensitivity prediction by transcriptional profiling. Proc Natl Acad Sci USA 2001, 98(19):10787–92. 10.1073/pnas.191368598PubMed CentralView ArticlePubMedGoogle Scholar
- Shi LM, Fan Y, Lee JK, Waltham M, Andrews DT, Scherf U, et al.: Mining and visualizing large anticancer drug discovery databases. J Chem Inf Comput Sci 2000, 40(2):367–79.View ArticlePubMedGoogle Scholar
- Shi LM, Myers TG, Fan Y, O'Connor PM, Paull KD, Friend SH, et al.: Mining the National Cancer Institute Anticancer Drug Discovery Database: cluster analysis of ellipticine analogs with p53-inverse and central nervous system-selective patterns of activity. Mol Pharmacol 1998, 53(2):241–51.PubMedGoogle Scholar
- Gilmour DS, Elgin SC: Localization of specific topoisomerase I interactions within the transcribed region of active heat shock genes by using the inhibitor camptothecin. Mol Cell Biol 1987, 7(1):141–148.PubMed CentralPubMedGoogle Scholar
- Swamidass SJ, Chen J, Bruand J, Phung P, Ralaivola L, Baldi P: Kernels for small molecules and the prediction of mutagenicity, toxicity and anti-cancer activity. Bioinformatics 2005, 21(Suppl 1):i359–68. 10.1093/bioinformatics/bti1055View ArticlePubMedGoogle Scholar
- Perret J, Mahe P, Vert J: Chemcpp: an open source c++ toolbox for kernel functions on chemical compounds.2007. [http://chemcpp.sourceforge.net]Google Scholar
- Sayers E: PubChem: An Entrez Database of Small Molecules. NLM Tech Bull 2005, 342(Jan–Feb):e2.Google Scholar
This article is published under license to BioMed Central Ltd. This is an open access article distributed under the terms of the Creative Commons Attribution License (http://creativecommons.org/licenses/by/2.0), which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.