Robert L. Lipnick
Correlative vs Mechanistic QSAR Models for Review of Industrial Chemicals Under TSCA.
Office of Pollution Prevention and Toxics (7403), U.S. Environmental Protection Agency, Washington, DC 20460, USA. E-mail: lipnick.robert@epamail.epa.gov
Prior to moving to EPA in 1979, I was at the Sloan-Kettering Institute for Cancer Research, working on the design and synthesis of nucleoside analogues as anti-cancer agents. Since such nucleoside analogues act by highly specific molecular mechanisms, I was studying the 3-dimensional conformational requirements for binding, activity, and how modifications in chemical structure of such molecules may enhance their activity and selectivity. This is in contrast to toxic responses to industrial chemicals, which I subsequently learned do not, at least for acute toxicity, usually involve highly specific interactions with receptors.
At EPA, I was assigned a task perhaps similar to what Wright-Patterson is undertaking in developing systematic methods for assessing potential adverse effects of chemicals of interest to the Air Force. Under the authority of Section 5 (Premanufacture Notification or PMN) of the 1976 Toxic Substances Control Act (TSCA), EPA's Office of Toxic Substances (now Office of Pollution Prevention and Toxics) is responsible for reviewing potential health and environmental hazards that may be associated with commercial production, release, and exposure to such substances. Since TSCA requires that the submitter be responsible for providing only data already existing within their files on health and safety, generally little or no toxicological test data have been available to reviewers. At that time the approach was to do a literature search to identify any additional literature data on the chemical, as well as to find other chemicals with test data that could be used as "analogues" in this assessment process.
This decision process based upon limited test data may be regarded as akin to a grant proposal to NIH to synthesize and study a new anti-cancer drug, based upon some structural analogy to related substances that have already been shown to be active. Thus, at this stage, EPA is not using structure-activity relationships to control a PMN chemical, but to determine its relative priority for additional testing. To support this effort, we sought to develop more analytic approaches for identifying appropriate analogues with respect to underlying biological mechanism, as well as what physicochemical and chemical properties could be used to quantify this process.
As the panelist for the SAT (Structure-Activity Team) that reviewed each of these PMNs, I was assigned the job of making the initial call with respect to ecotoxicology. When I joined the SAT, I was fortunate to be working with Charles Walker who was on loan from the U.S. Fish and Wildlife Service (USFWS) and who had spent a number of years in the laboratory testing chemicals for toxicity to fish by trying to relate chemical structure to fish toxicity. Mr. Walker had access to a number of fish toxicity screening studies performed under USFWS sponsorship for thousands of organic chemicals, and these studies became a goldmine for ecotoxicology analogue data. Realizing the significance of these data which were either unpublished or long out-of-print, I arranged for their publication as joint EPA-USFWS reports. For human health, we were fortunate to have Dr. Joseph Seifter, who served as a Presidential appointee to assist EPA in this work.
In 1975, Dr. Gilman Veith at the EPA laboratory in Duluth, MN organized a workshop with the Great Lakes International Joint Commission to assess the value of structure-activity relationships in studies of toxicity and bioconcentration with aquatic organisms. By the early 1980s, several papers had been published showing that the toxicity of simple non-electrolyte organic chemicals such as saturated monohydric alcohols can be fit to a simple regression model with log P (P=octanol/water partition coefficient) as the only parameter or molecular descriptor needed.
To get a better grasp of this field and how we should approach future work, I was able to arrange for Prof. Adrien Albert, Australian National University, and perhaps the foremost authority on structure-activity relationships, to spend a month at EPA Headquarters in November, 1982. During this time, Prof. Albert participated in SAT meetings and prepared a recommendation report on future directions, as well as delivering a series of 8 seminars on the history and theory of structure-activity relationships. These seminars were highly successful and overwhelmed the seating capacity of any EPA lecture room available at the time. Scientists who had been familiar with Prof. Albert's monograph Selective Toxicity (first published in 1951), were pleased to hear and meet him. We were also able to arrange for EPA's Audio-Visual Department to videotape these seminars.
My own interests led me to derive similar models using data already available in the toxicological literature, as well as to investigate the historical roots of these correlations at the turn of the century by Overton in Zürich and Meyer in Marburg, and later by Lazarev in St. Petersburg. The narcosis model serves as a reference for baseline or minimum toxicity for non-electrolyte organic chemicals and can be used as a probe for identifying more specific mechanisms. We found the screening data mentioned above a useful source for semi-quantitatively testing the scope of this narcosis correlation or QSAR for alcohols, and discovered that certain acetylenic alcohols (primary and secondary propargylic) were showing lethality at 5 or 10 mg/L screening levels, even though baseline toxicity predicted effects at above 100 mg/L.
In a paper describing the above work, we suggested the possibility of toxicity being mediated via metabolic activation to a chemical electrophile which could result in specific irreversible covalent binding. We termed these toxicants proelectrophiles. This finding was confirmed and extended based upon new test data from Duluth, and this class of chemicals has been used by others as a model for investigating a more specific mechanism and how to derive quantitative correlations. Thus, from a mechanistic standpoint, toxicity of such proelectrophiles should be related to uptake and distribution (log P), as well as one or more electronic properties that model rate of biotransformation and rate of covalent binding and inhibition of activity. This and other studies demonstrated overwhelmingly the importance of not only finding correlations between properties, structure, and toxicity, but also the need to understand the underlying mechanisms responsible for these events. Additional outliers were discovered by looking at two other classes (phenols and anilines) for which correlative QSARs had been reported. In each case, we found significant outliers between what the QSAR predicted and the low screening concentrations at which effects were observed. The outliers were all electrophiles or proelectrophiles. We also looked more generally at compounds containing chemically reactive functional groups (e.g., allyl and benzyl halides); their increased activity could also be accounted in this way with toxicity resulting from covalent binding to one or more critical enzymes and reduction in biochemical function. This also fit with what was learned during WWII about the mechanism of lachrymators and vesicants.
The above work involving correlations with log P and looking for outliers also identified literature studies in which simple nonelectrolyte chemicals appeared less toxic than predicted by the baseline or minimum toxicity narcosis model. In one such study, we were able to calculate Henry's Law constants for the test chemicals and found a correlation between the degree to which these less toxic chemicals were outliers and their expected rates of volatilization. In these literature reports, toxicity was measured in static tests; thus, if the substance was being lost by volatilization during the test duration, the apparent concentration required to produce the effect was greater than actual over the course of the experiment. Overton had reported a similar finding in his classic 1901 book Studies of Narcosis.
These efforts were considerably assisted over the years by the collaboration between Duluth and Office of Toxic Substances in the development of the AQUIRE database of aquatic toxicity data and the development of the CLOGP computer program via a cooperative agreement with Pomona College to estimate log P values of chemicals. The latter made it possible to easily assign values of this important parameter to many of the chemicals for which we had toxicity test data but no log P measurements available. In some cases, the studies also identified compounds with functional groups for which log P fragment constants were unavailable or needed to be measured more precisely in the specific chemical environment.
We were also able to extend these correlation studies for narcosis and outliers to various mammalian endpoints which were presented at various European QSAR Symposia and QSAR workshops. For rat oral LD50, mouse oral LD50, mouse intravenous LD50, rabbit dermal LD50, and eye irritation, we found a non-linear relationship that could be fit by a parabolic or better bilinear relationship which was derived from data on monohydric saturated alcohols. In each case, like the fish data, acetylenic alcohols (in particular, primary and secondary propargylic and homopropargyic alcohols) were outliers and this was attributed to a corresponding proelectrophile mechanism. The introduction of the terms electrophile and proelectrophile mechanisms into the toxicological literature also seems to have had a synergistic effect in attracting more interest to underlying toxicological mechanisms and what would be suitable mechanistic parameters (both measured and estimated) for QSAR model development. Of course, the increased mechanistic understanding has been valuable as well in probing the scope and limitations of these QSARs for predictive purposes.
A final word about correlative QSARs. These are always helpful as a first step and the most obvious parameter to start with is log P, even if it is likely that toxicity is governed by electronic factors as well. If data can be assembled on a sufficiently narrowly defined data set, these electronic differences may be minimal and a useful correlation can be found. Collections of such local QSAR models can be extremely valuable for assessment purposes, in the absence of more general mechanistically-based predictive models.
Back to Program PageBack to Main Page