OCTOBER 2006 VOLUME 19, NUMBER 10 © Copyright 2006 by the American Chemical Society
PerspectiVe Future of ToxicologysPredictive Toxicology: An Expanded View of “Chemical Toxicity” Ann M. Richard* National Center for Computational Toxicology, Mail Drop D343-03, U.S. EnVironmental Protection Agency, Research Triangle Park, North Carolina 27711 ReceiVed May 31, 2006
A chemistry approach to predictive toxicology relies on structure-activity relationship (SAR) modeling to predict biological activity from chemical structure. Such approaches have proven capabilities when applied to well-defined toxicity end points or regions of chemical space. These approaches are less wellsuited, however, to the challenges of global toxicity prediction, i.e., to predicting the potential toxicity of structurally diverse chemicals across a wide range of end points of regulatory and pharmaceutical concern. New approaches that have the potential to significantly improve capabilities in predictive toxicology are elaborating the “activity” portion of the SAR paradigm. Recent advances in two areas of endeavor are particularly promising. Toxicity data informatics relies on standardized data schema, developed for particular areas of toxicological study, to facilitate data integration and enable relational exploration and mining of data across both historical and new areas of toxicological investigation. Bioassay profiling refers to large-scale high-throughput screening approaches that use chemicals as probes to broadly characterize biological response space, extending the concept of chemical “properties” to the biological activity domain. The effective capture and representation of legacy and new toxicity data into mineable form and the large-scale generation of new bioassay data in relation to chemical toxicity, both employing chemical structure information to inform and integrate diverse biological data, are opening exciting new horizons in predictive toxicology. Introduction
Contents Introduction Local vs Global Bringing More Information To Bear on Toxicity Prediction Toxicity Data Informatics Bioassay Profiling Conclusion
1257 1258 1258 1258 1259 1261
* To whom correspondence should be addressed. Tel: 919-541-3934. Fax: 919-685-3263. E-mail:
[email protected].
Predictive toxicology, from a chemistry vantage point, uses a structure-activity relationship (SAR) built on available test data to predict the potential biological activity of a chemical based solely on its molecular structure and computed properties. Predictive toxicology approaches based on SAR modeling are entirely dependent on and limited by the conventions and reference data of toxicology studies, in terms of both the chemicals chosen for study and the experimental end points considered most informative for toxicological inferences. However, SAR modelers typically have had little prospective influence in defining either the chemical scope or the measured
10.1021/tx060116u CCC: $33.50 © 2006 American Chemical Society Published on Web 09/02/2006
1258 Chem. Res. Toxicol., Vol. 19, No. 10, 2006
parameters of toxicological studies. The entire enterprise of toxicology, historically organized around specific categories of responses in whole animals, is undergoing profound change. Trends effecting this change are three-fold: (i) large-scale data generation involving more fundamental, interdisciplinary technologies (driven by genomics and proteomics); (ii) the emergence of extensive public information resources to support these efforts; and (iii) greater investigative focus on chemical and biological mechanisms that underlie toxicological responses and disease states (1). Given its intimate reliance on toxicity reference data, the field of predictive toxicology is poised to undergo a correspondingly dramatic shift in focus and approach. The most significant drivers of this change are not the trade tools of chemists and SAR modelers, i.e., new chemical descriptors, new data analysis tools, or more sophisticated algorithms for defining chemical similarity, or domain of applicability and validity of models, although in all of these areas there have been significant advances in recent years. Rather, the greatest drivers for change will be in the elaboration and enrichment of the “activity” portion of the SAR paradigm.
Local vs Global Focused computational chemistry investigation has proven its ability to elucidate well-defined biochemical or mechanistic outcomes for a series of closely related chemical structures. Such mechanistically based SAR models are most successful when applied to well-defined categories of toxicities (2) or to narrowly defined subsets of chemicals and toxicity measures or target interactions (3, 4). Hence, SAR modeling performs best when applied locally with respect to either chemistry or biology. The most pressing challenges confronting predictive toxicology are more global in nature. In the environmental regulatory domain is the need to prioritize testing (5), reduce testing (6), and eliminate testing entirely in some cases (7) for a wide diversity of chemicals posing a broad range of potential toxicological concerns. In the pharmaceutical domain, the goal is to anticipate a spectrum of potential adverse effects of drug candidates in humans based largely on surrogate in vitro and animal test results. Some success has been achieved with the so-called “global SAR approaches” to predict broadly across chemical space for a particular toxicity end point. These often rely on a set of structural alerting features (i.e., chemical fragments) to grossly represent many regions of chemical space associated with a toxicity end point (8). However, the problem of predicting the potential toxicity of a never-before-seen chemical has largely defied accurate and reliable solutions and is considered by many to be beyond the reach of SAR approaches that rely exclusively on chemical structures and associated properties. Global models for a toxicity end point have proven of value in providing guidance and approximate prediction estimates; for example, if a known structural alerting feature for that end point is present, the chemical has a greater than chance probability of being associated with that toxicity. However, these models frequently fail to predict activity modulations within local structure-activity space (e.g., quite a few, but not all, aromatic amines are carcinogenic) or to predict outside the chemical space of the training set, where in both cases sufficient chemical test examples do not exist to adequately inform the model (9). A major challenge to be met by any successful global SAR predictive toxicity model is the sufficient characterization and coverage of local domains or “neighborhoods” of similar chemicals. When defined exclusively in chemical terms, this means that sufficient toxicity test data within chemical analogue
Richard
neighborhoods are needed to enable confident SAR inferences and sufficient test coverage across chemically diverse neighborhoods is needed to enable global SAR predictions. It is implicitly assumed that this broad chemical coverage, in turn, will span the full range of biological mode-of-action neighborhoods in relation to the ultimate adverse effect measured. Given that these requirements are with respect to each reference toxicological test end point of interest, these broad data requirements are a prescription either for endless chemical toxicity testing or for failure. Furthermore, given the diversity and extent of chemical space and the lack of full or even partial mechanistic definition for most toxicological end points of regulatory or human health concern, coverage requirements of all constituent local models are unlikely to be met except in the rarest of circumstances. A more fundamental problem with the concept of a global SAR model able to confidently predict the potential for toxicity in uncharted areas of chemical space is that chemical structure space is typically more rugged and discontinuous than corresponding biological activity space when the latter is represented by an integrated whole animal response (this may not be the case, however, when activity is represented at the biochemical or genomic response level). A basic tenet of SAR investigation that works well and often enough to be useful is that small changes in chemical structure lead to correspondingly small changes in biological activity. However, a biological organism is a complex system that frequently imposes discontinuity onto chemical space, such as when small structural changes trigger metabolic activation or a target interaction and genomic transcriptional response, with a correspondingly large impact on the ultimate biological outcome. In addition, the SAR paradigm is notoriously bad at projecting into areas of chemical structure space not sufficiently covered by the training set of compounds. One could argue that unrealistic expectations have been placed on SAR approaches to tackle these global toxicity prediction problems; there are simply not enough test data, and likely never will be, to support this SAR global toxicity prediction objective as a viable way forward.
Bringing More Information To Bear on Toxicity Prediction If predictive toxicology is to significantly advance in capability and utility, it must broaden its effective utilization of both legacy and new chemical toxicity test data across levels of biological organization and traditional toxicity study areas. In addition, to break free of the known limitations of chemistryexclusive SAR toxicity prediction models, a new paradigm is needed for defining more functionally meaningful chemical/ biological activity neighborhoods that are informed by integrated biological measures and that bear closer relation to the ultimate toxicity end point being predicted. Major advances that hold great promise for effecting these changes are occurring, largely in the public domain, in the areas of toxicity data informatics and bioassay profiling.
Toxicity Data Informatics Toxicity data informatics (Figure 1) refers to: (i) increasing standardization and digitization of legacy toxicity data and migration of these data into the public domain; (ii) development and population of database standards and models that facilitate data integration and enable relational exploration and mining of data across both historical and new areas of toxicological investigation; and (iii) association of these data with chemical structuressa universal metric spanning chemical toxicology
PerspectiVe
Chem. Res. Toxicol., Vol. 19, No. 10, 2006 1259
Figure 1. Sample construction of toxicity data model, including DSSTox structure annotation standards, ToxML toxicity data schema with hierarchical summary toxicity layers, and importation of these elements into a chemical relational database that supports data mining, modeling, and prediction (screenshot of Leadscope Toxicity Model).
domain that enables structure analogue searching and SAR hypothesis generation. A data model, or schema, consists of controlled vocabulary, standardized data descriptions, and a hierarchical field structure that reflects layers of study data and biological summarization and organization (10). Data models are being developed with the involvement of toxicology domain experts in areas such as genetic toxicity, chronic toxicity, and developmental toxicity (11, 12). A data model enables read-across exploration and mining of toxicity study. Such models are essential first steps for more effective data utilization and encourage both the efficient capture of newly generated data and the migration of legacy data archives to enrich public toxicity data resources [from the historical literature as well as from government regulatory agencies, such as the Food and Drug Administration and the Environmental Protection Agency (EPA)]. With the built-in flexibility to collapse and summarize study data to various quantitative and qualitative “end points”, data models offer a route to new testable hypotheses relating structure to biological activity (10) and potentially rich content input to nextgeneration prediction models (13). Populated toxicity data models, by virtue of being top-level chemically indexed, i.e., searchable by chemical name, CAS registry number, or structure, also can be easily linked to the larger world of chemically indexed public toxicity information. In contrast, many currently available biological data resources derived from chemical exposure testing are not yet chemically indexed. Prominent examples include the largest of the public microarray data repositories, GEO and ArrayExpress (14, 15), in which even minimal chemical annotation, i.e., a chemical name or CAS registry number, is not required in data submis-
sion, and neither chemical name nor CAS number is presented as a field search option for accessing microarray experiments within the databases at the time of this writing. As a result, these experimental results remain effectively isolated in data silos that serve limited study domain interests (16), and the chemical coverage and inventory within these databases are largely unknown. Chemical structure indexing is advancing onto these islands, however, and is proving to be a major and effective integrator of biological information across the Internet (17). This is perhaps best exemplified by the PubChem Project (18), a fully public data model that provides on-line structure searchability and full and open data access to millions of chemically indexed bioassay records from various public and user-deposited sources. The DSSTox database project is an example more specifically targeted to toxicology study areas, providing summary representations of activity potentially useful for SAR modeling applications, with standardized chemical structure annotation enabling read-across searchability over a wide range of chemical toxicity information (19). These and other public initiatives, such as the new InChI text structure representation (20), are moving us closer to the reality of a fully structure-searchable “chemical semantic web” (21). Such advances will enhance the ability to gather or mine a broad range of information on a chemical and its analogues from public Internet resources in support of chemical toxicity prediction efforts.
Bioassay Profiling Bioassay profiling refers to: (i) broad characterization of a chemical in terms of biological response space, i.e., extending
1260 Chem. Res. Toxicol., Vol. 19, No. 10, 2006
chemical “properties” to molecular functional and cellular levels in the biological domain; (ii) bioassay data typically generated in high-throughput and high-content assays for a wide spectrum of biological targets and cellular responses; and (iii) data that provide new information and have the potential, alone or in combination with other data, to serve as bioindicators of an integrated toxicological response in the whole animal of concern. A new paradigm for improving toxicity prediction is beginning to emerge that resembles the directed high-throughput screening (HTS) approaches more commonly employed in support of drug discovery in the pharmaceutical industry (22). In this approach, large numbers of chemicals representing chemical diversity space are employed as probes of biological activity. This is accomplished through the use of a wide variety of HTS bioassay screens measuring biochemical activity and cell function, effectively generating structure-activity information de novo across biological function space. Anchoring of these chemical-bioactivity profiles to appropriate reference toxicological data provides the interpretive context for developing predictive models. This new approach to data generation and prediction differs in two very fundamental respects from past efforts. First, SAR-based predictive toxicology approaches traditionally have exerted little to no influence on which chemicals or limited areas of chemical space were to undergo testing. In the new HTS approaches, chemicals are chosen to cover large areas of chemical diversity space and to broadly probe biological function space without a priori assumptions (23). Second, the nature of the data being generated at the chemical:biological interface represents a projection of chemical structure space into the realm of biological functional activity. As a result, there is potential for breaking free of the inherent limitations of a purely chemical similarity perspective and defining perhaps fewer or more coherent biological activity neighborhoods in relation to an integrated whole animal response. Covell and co-workers have considered the use of multidimensional bioassay profile information for growth inhibition activity across a series of 60 tumor cell lines in relation to the large NCI chemical structure space (24). In their words, “The considerable chemical and biological diversity inherent in these data offers an opportunity to establish a quantifiable connection between chemical structure and biological activity. We find that the connection between structure and biological response is not symmetric, with biological response better at predicting chemical structure than vice versa.” Similarly, Fliri and co-workers at Pfizer have examined the use of bioassay profiling as means for classifying chemicals according to common modes of activity and therapeutic biological response categories (25). Their findings provide evidence for coherent biological response patterns that overlap to some degree with chemical similarity considerations but that are not fully determined or anticipated by chemical similarity alone. These efforts to date have been largely proof-of-principle, focused on particular tumor cell lines, in the case of the NCI, or on limited numbers of commercially available HTS assays and pharmaceutical type chemicals. Some success has also been reported in deriving genomic signatures to predict selected toxicity end points of concern to the pharmaceutical industry (26). Although promising, these genomics approaches have limited applicability to global prediction when signatures are derived from proprietary microarray databases, when testing is limited mostly to pharmaceutical type compounds, when the chemical structure space of the training set and compound of interest is not considered in prediction, and when the application of the signature to prediction requires
Richard
use of either the commercial service or the microarray experiments with comparable protocols to this service. A very large public HTS initiative taking shape under the NIH Molecular Libraries & Imaging Roadmap Initiative (MLI) (27) has the potential to greatly enrich data resources and our basic understanding of chemical structure in relation to biofunctional activity space. To support the HTS component of this effort, the NIH Molecular Library Small Molecule Repository (28), consisting of upward of 70000 chemicals spanning large regions of chemical diversity space, is to undergo screening in several hundreds, perhaps ultimately thousands, of bioactivity and biofunctional HTS assays (29). This testing program is well underway, with HTS data currently being generated across the Molecular Libraries Screening Center Network (MLSCN) (27) for an ever-growing list of HTS assays. As they are produced, these data are being made fully publicly available in PubChem (18), the public data distribution arm of the NIH MLI effort. Given the tremendous data generation capability enabled by the new HTS technologies, what is the practical outlook for these new data impacting our ability to predict or screen for chemical toxicity? Success will depend on many additional factors. Are the chemical structures chosen for HTS sufficiently representative of environmental and industrial chemicals that comprise the bulk of reference toxicity data sets, to which the HTS data must be anchored? Are the HTS assays sufficiently representative of the bioactivity and biofunctional domains most relevant to the wide array of potential toxicity end points of concern? Are the achievable HTS concentrations appropriate to capture toxicologically relevant effects, i.e., to minimize false negatives? Will HTS generate sufficient numbers of actives and unique profiles across tested chemical space to be useful for activity discrimination? Also, will the chosen HTS assays sufficiently capture requirements for metabolic activation, either through sufficient sampling of chemical space (to include possible metabolites) or by explicit inclusion of metabolic capabilities in the assays? Despite these varied concerns, the opportunity to enhance predictive capabilities from HTS data generation is simply too great to ignore. As a result, both EPA and the NIEHS National Toxicology Program are embarking on their own directed HTS efforts, as well as actively engaging with the NIH Chemical Genomics Center (NCGC is the intramural screening center in the MLSCN) to expand and improve the value of the HTS results specifically in relation to chemical toxicity. The NTP has submitted 1408 chemicals (30, 31), a large portion of this set overlapping with the on-line NTP reference toxicity database (32), for testing by the NCGC in several toxicity-related HTS screens. This preliminary set of chemicals and assays has been run by the NCGC at expanded dose ranges (up to 15 dose dilutions), and a broader set of chemicals and HTS assays are under consideration for further testing. Within the EPA National Center for Computational Toxicology’s new ToxCast initiative (33), HTS and bioassay profiling are envisioned to be major components of toxicity “forecasting”. The EPA ToxCast program is collaborating with the NTP and NCGC efforts, as well as pursuing customized and commercial HTS testing of several hundred pesticidal actives, for which reference toxicity data across a number of end points of regulatory concern are available. Ultimately, the objective is to extract patterns of bioactivity profiles, possibly in combination with other types of data, that are predictive of integrated biological responses. Given the large amounts of HTS data to be generated in these efforts, there is additionally the potential to derive companion SAR models for particular HTS end points, augmenting HTS
PerspectiVe
Chem. Res. Toxicol., Vol. 19, No. 10, 2006 1261
Figure 2. Multidimensional and broadly integrated activities and information domains that will support improved predictive toxicology capabilities.
data or enabling the prediction of bioactivity profiles in the absence of HTS data (34).
Conclusion There are many exciting new developments in the world of predictive toxicology that should engage this community and give rise to significant optimism for the future (see Figure 2). First and foremost, and long overdue, are efforts to make better use of the reference toxicity data that we already have, to make it publicly available, structure-searchable, mineable, more suitable for modeling, and better integrated across domains of toxicological study. Second are large public efforts directed toward the generation of new bioassay profiling data of potential relevance to toxicology, where the active engagement of toxicologists, chemists, and modelers is essential if these data are to be effectively integrated with existing data and optimally utilized. Last, but not least, is recognition of the key role to be played by chemical structures in integrating diverse biological data, rationalizing such data through enhanced SAR models, and serving as effective probes of biological activity space. To comment on this or other Future of Toxicology perspectives, please visit our Perspectives Open Forum at http:// pubs.acs.org/journals/crtoec/openforum. Acknowledgment. This manuscript was reviewed by the U.S. EPA’s National Center for Computational Toxicology and approved for publication. Approval does not signify that the contents necessarily reflect the views and policies of the Agency, nor does mention of trade names or commercial products constitute endorsement or recommendation for use.
References (1) Liebler, D. C. (2006) The future of toxicologysThe poisons within: Application of toxicity mechanisms to fundamental disease processes. Chem. Res. Toxicol. 19 (5), 610-613. (2) Cronin, M. T. D., and Schultz, T. W. (2001) Development of quantitative structure-activity relationships for the toxicity of aromatic compounds to Tetrahymena pyriformis: Comparative assessment of the methodologies. Chem. Res. Toxicol. 14 (9), 1284-1295. (3) Yu, S. J., Keenan, S. M., Tong, W., and Welsh, W. J. (2002) Influence of the structural diversity of data sets on the statistical quality of threedimensional quantitative structure-activity relationship (3D-QSAR) models: Predicting the estrogenic activity of xenoestrogens. Chem. Res. Toxicol. 15 (10), 1229-1234. (4) Aptula, A. O., Patlewicz, G., and Roberts, D. W. (2005) Skin sensitization: Reaction mechanistic applicability domains for structureactivity relationships. Chem. Res. Toxicol. 18 (9), 1420-1426.
(5) DSL (Domestic Substance List): An initiative of Canadian Environmental Protection Act to categorize substances to categorize either manufactured in or imported chemicals (23000) for their toxicity or potential to become toxic by 2006; http://www.ec.gc.ca/substances/ ese/eng/dsl/dslprog.cfm. (6) EPA High Production Volume Challenge Program; http:// www.epa.gov/chemrtk/hpvchmlt.htm. (7) REACH (registration, evaluation, and authorization of chemicals); http://europa.eu.int/comm/environment/chemicals/reach.htm; http:// ecb.jrc.it/REACH/; 7th Amendment for cosmetics: A ban on the use of live animals for testing cosmetic products and ingredients by 2013; http://www.eceae.org/english/cosmetics.html. (8) Kazius, J., McGuire, R., and Bursi, R. (2005) Derivation and validation of toxicophores for mutagenicity prediction. J. Med. Chem. 48, 312320. (9) Benigni, R. (2005) Structure-activity relationship studies of chemical mutagens and carcinogens: Mechanistic investigations and prediction approaches. Chem. ReV. 105, 1767-1800. (10) Yang, C., Richard, A. M., and Cross, K. P. (2006) The art of data mining the minefields of toxicity databases to link chemistry to biology. Curr. Comput.-Aided Drug Des. 2 (2), 135-150. (11) ToxML: A publicly available XML standard and controlled vocabulary for representing toxicity data (2006); http://www.leadscope.com. (12) Using developmental toxicity data in SAR model development. Project of the ILSI Research Foundation, Risk-Sciences Institute (2006); http:// rsi.ilsi.org/Projects/DevToxSAR. (13) Yang, C., Benz, R. D., and Cheeseman, M. A. (2006) Landscape of current toxicity databases and database standards. Curr. Opin. Drug DiscoVery DeV. 9 (1), 124-133. (14) National Center for Biotechnology Information (NCBI) Gene Expression Omnibus (GEO) (2006); http://www.ncbi.nlm.nih.gov/geo/. (15) European Bioinformatics Institute (EBI) ArrayExpress (2006); http:// www.ebi.ac.uk/arrayexpress/. (16) Wishart, D. S., Knox, C., Guo, A. C., Shrivastava, S., Stothard, P., Chang, Z., and Woolsey, J. (2006) DrugBank: A comprehensive resource for in silico drug discovery and exploration. Nucleic Acids Res. 34, D668-D672. (17) Richard, A. M., Gold, L. S., and Nicklaus, M. C. (2006) Chemical structure indexing of toxicity data on the internet: Moving towards a flat world. Curr. Opin. Drug DiscoVery DeV. 9 (3), 314-325. (18) National Center for Biotechnology InformationsPubChem project structure-relational database and browser (2006); http:// pubchem.ncbi.nlm.nih.gov. (19) EPA DSSTox, Distributed Structure-Searchable Toxicity Database Network website (2006); http://www.epa.gov/ncct/dsstox/. (20) The IUPAC International Chemical Identifier (InChI): International Union of Pure and Applied Chemistry, Research Triangle Park, NC (2006); http://www.iupac.org/inchi/. (21) Coles, S. J., Day, N. E., Murray-Rust, P., Rzepa, H. S., and Zhang, Y. (2005) Enhancement of the chemical semantic web through the use of InChI identifiers. Org. Biomol. Chem. 3 (10), 1832-1834. (22) Whitebread, S., Hamon, J., Bojanic, D., and Urban, L. (2005) In vitro safety pharmacology profiling: an essential tool for successful drug development. Drug DiscoVery Today 10 (21), 1421-1433. (23) Strausberg, R. L., and Schreiber, S. L. (2003) From knowing to controlling: A path from genomics to drugs using small molecular probes. Science 300, 294-295. (24) Wallqvist, A., Huang, R., Thanki, N., and Covell, D. G. (2006) Evaluating chemical structure similarity as an Indicator of cellular growth inhibition. J. Chem. Inf. Model 46, 430-437. (25) Fliri, A. F., Loging, W. T., Thadelo, P. F., and Volkmann, R. A. (2005) Biological spectra analysis: Linking biological activity profiles to molecular structure. PNAS 102 (2), 261-266. (26) Fielden, M. R., Eynon, B. P., Natsoulis, G., Jarnagin, K., Banas, D., and Kolaja, K. L. (2005) A gene expression signature that predicts the future onset of drug-induced renal tubular toxicity. Toxicol. Pathol. 33, 675-683. (27) NIH Molecular Libraries & Imaging Roadmap Initiative (2005); http:// nihroadmap.nih.gov/molecularlibraries/. (28) NIH Molecular Libraries Small Molecule Repository; http:// mlsmr.discoverypartners.com/MLSMR_HomePage/. (29) Austin, C. P., Brady, L. S., Insel, T. R., and Collins, F. S. (2004) NIH Molecular Libraries Initiative. Science 306, 1138-1139. (30) National Toxicology Program High-Throughput Screening Program (2006); http://ntp.niehs.nih.gov/ntpweb/index.cfm?objectid) 05F80E15-F1F6-975E-77DDEDBDF3B941CD. (31) Smith, C., Collins, B., Tice, R., Wolf, M. A., and Richard, A. M. (2006) DSSTox National Toxicology Program High Throughput Screening Structure-Index File: SDF file and documentation, Launch version NTPHTS_v1a_1408_10Apr2006; http://www.epa.gov/ncct/ dsstox/sdf_ntphts.html.
1262 Chem. Res. Toxicol., Vol. 19, No. 10, 2006 (32) NTP On-line Bioassay Database, Study Data searches; http:// ntp-server.niehs.nih.gov/index.cfm?objectid)72016020-BDB7-CEBAF3E5A7965617C1C1. (33) Dix, D. J., Houck, K. A., Martin, M. T., Richard, A. M., Setzer, R. W., and Kavlock, R. J. (2006) The ToxCast program for prioritizing toxicity testing of environmental chemicals. Toxicol. Sci. (Forum), in press.
Richard (34) Poroikov, V. V., Filimonov, D. A., Ihlenfeldt, W. D., Gloriozova, T. A., Lagunin, A. A., Borodina, Y. V., Stepanchikova, A. V., and Nicklaus, M. C. (2003) PASS biological activity spectrum predictions in the Enhanced Open NCI Database Browser. J. Chem. Inf. Comput. Sci. 43, 228-236.
TX060116U
