In 1996, transgenic crops (cotton, corn, and potatoes) containing the gene from Bacillus thuringiensis (Bt) that encodes for an insecticidal toxin were first commercially produced in the United States. Because we can expect the commercialization of numerous crops with Bt insecticidal genes, environmentalists have cautioned regulators to explore the risks of this new technology. In response, industry has argued that this really is not a new technology, since plant breeders have a long history of artificially selecting for herbivore-resistant crop varieties using conventional methods. Both sides of this debate have merit—herbivore resistance is in fact a selected trait in many crops, but on the other hand these particular Bt genes are entirely novel to the plant species into which they are introduced. For instance, a Bt gene product that acted as an insecticide against cabbage worms (Pieris rapae) inserted into a Brassica crop would be expressed as an entirely novel trait—Brassicas that are immune to attack from cabbage worms have never been produced by artificial selection. Rather than splitting hairs about what is "novel," the more pertinent question is: what are possible effects of Bt resistance genes entering wild populations of plants? The assumption here is that the genes will escape into wild populations; even if gene flow is extremely infrequent, extensive commercialization makes such a scenario highly likely. To answer this question we examine quantitative field experiments that have manipulated the number of herbivores attacking plants in a manner that well simulates the action of a Bt gene. Second, we use the effects detected by these field experiments to simulate a "risk assessment" and compare how well various risk assessment strategies would fare in predicting long-term population trends. Finally, we discuss results from previous studies that add insight into the challenge of conducting a risk assessment for transgenic crops containing resistance genes.

To assess the potential risks associated with the escape of herbivore resistance genes into wild populations of plants, we reviewed the strength of herbivore effects documented in the recent ecological literature. Specifically, we compiled and reviewed a collection of studies that compare the reproductive performance of plants protected from herbivores versus plants exposed to herbivores. The increased success of plants in the absence of herbivores provides a hint of how the performance of a non-crop plant might be affected if it were to obtain herbivore resistance genes from a transgenic crop. We used a formal statistical approach called meta-analysis (Hedges and Olkin 1985) to quantify the overall effects that insect herbivores have on plant reproductive success. Meta-analysis involves standardizing the difference between treatments in each experiment and using these standardized differences as individual data points. We searched the contents of eight ecological journals (volumes from January 1983-June 1997 of American Naturalist, Ecological Applications, Ecological Monographs, Ecology, Journal of Applied Ecology, Journal of Ecology, Oecologia, and Oikos) and selected terrestrial field studies that measured plant reproductive responses to manipulated densities of insect and/or mollusc herbivores. We excluded agricultural and greenhouse studies because we were interested in assessing the effects of real herbivores on natural plants and maximizing the relevance of our findings to our scenario of a transgenic weed that has obtained herbivore resistance genes. To minimize the number of non-independent contrasts, we used only one measure of plant reproduction for each experiment, according to the following ranking (starting with most preferred): viable seeds per plant, total seeds per plant, fruits per plant, inflorescences or flowers per plant, seeds per inflorescence, fruits per inflorescence, and fruits per initiated bud.

We located 18 publications that satisfied our criteria (Table 1).

Table 1. References used for meta-analysis. These studies measured the effects of manipulated insect and/or mollusc densities on plant reproductive performance.

Several of these papers reported data from multiple experiments, multiple plant species, or multiple sites, resulting in a total of 52 comparisons. This data set includes responses of 28 plant species in 19 genera (Table 1). To compare the strength of herbivore effects, we used the weighted standardized mean difference, Hedges' d, as the measure of effect size for each study. Hedges' d is calculated as the difference between the means of the experimental and control treatments divided by the pooled standard deviation, weighted by sample size (Hedges and Olkin 1985). This metric compares the difference between treatments to the difference within treatments: the numerator is the difference between the means of the two groups and the denominator is the presumably random variation within groups. Thus, using this metric we can determine whether the difference in reproductive performance between treatments is large enough that it probably did not occur by chance alone. We defined as controls the treatment that exposed plants to a natural abundance of herbivores. The sign of d was reversed for studies that augmented herbivores. Thus, a positive value of d indicates an increase in reproductive performance in the absence of herbivores. For example, a study might report that plants protected from herbivores produced 165 seeds on average, whereas those exposed to herbivores produced 150 seeds on average, with a standard deviation within each group of 25 seeds. For this study, the average difference between treatments is d = (165 - 150) / 25 = 0.6 standard deviations. Assuming that seed production is normally distributed, d = 0.6 means that the average protected plant produced more seeds than 73% of the exposed plants. By convention, d = 0.2 is considered a small effect, d = 0.5 a medium effect, and d ≥ 0.8 a large effect (Cohen 1969). We calculated d for each comparison and then combined these values across comparisons. The statistical significance of d can be simply assessed by examining whether its 95% confidence interval overlaps with zero. We also compared the strength of herbivore effects between studies that excluded herbivores versus those that augmented herbivores using mixed model homogeneity analysis, roughly analogous to mixed model analysis of variance.

The meta-analysis of our collection of herbivore studies demonstrated that, on average, herbivory caused a statistically significant and large reduction in plant reproductive performance (Figure 1; overall effect size = 0.86; 95% CI= 0.55-1.17; n = 52).

One caveat from this analysis concerns its relevance to the insertion of a Bt gene. In particular, do the levels of herbivore protection used as treatments in the experiments listed in Table 1 reflect the likely levels of herbivore protection afforded by a Bt gene, which typically works against only a subset of herbivore species? Most of the reduced herbivore treatments in Table 1 targeted only a subset of herbivores (just as Bt does) and not all herbivore species. Of course, the treatments tended to focus on dominant herbivores, but this is exactly the case with Bt (one does not insert Bt genes into a group of plants to target minor herbivores). Indeed, 10 of 18 studies in Table 1 removed five or fewer species of herbivores and 8 removed only one herbivore species. Secondly, the main lesson from Figure 1 is that herbivory is clearly a potent demographic factor in natural populations of plants. Thus, the generality evident from field experiments is that any trait that protects a plant against a major herbivore will likely enhance that plant’s rate of reproduction.

One reason environmentalists are concerned with plant resistance genes is the worry that these genes may exacerbate weed problems. It is important to note that "weed" is used here in a broad sense, not simply to include agronomic problems—indeed, exotic invasive plants are the major threat to maintaining native vegetation and biodiversity within the nature reserves of North America (US Congress, 1993). The question is whether a plant like a weedy Brassica might acquire a Bt gene and, because of reduced herbivore pressure, become more invasive in natural communities. Such an increase in invasiveness might actually be more likely in natural communities, which, unlike agricultural settings, harbor large populations of insect herbivores because they are not treated with insecticides. Clearly, any change that increases rates of population growth will cause a plant to become more invasive. However, our simulation was aimed at a slightly different question: given the variability in "effect size" uncovered by our meta-analysis, how likely are short-term field experiments to detect the "true" effect of an herbivore-resistance gene in a wild plant? This is a pertinent question because in the absence of a "similarity argument" (which cannot be applied to a novel trait like production of Bt toxin) direct experiments are a major tool for risk assessment.

First, we compared rates of population growth obtained by using effect sizes either from herbivore augmentation studies or from herbivore removal studies. Second, we asked how the number of sites used in a field trial affects our conclusions about the risks posed by herbivore resistant weeds. We were particularly interested in how these methodological differences affect the frequency with which we would erroneously conclude that gene escape would not be a significant problem. The model of plant population growth was built on a few simplifying assumptions. First, we assumed that herbivore susceptible (control group) plant populations are stable. We set mean production of seeds to 20 per individual with the standard deviation in seed production equal to 10. This rate of seed production combined with the probability of survival from seed to adult = 0.05 gave stable populations of herbivore susceptible plants. We assumed that plant populations grow exponentially and that population growth is seed limited, so an increase in seed production due to herbivore resistance translates directly into a higher rate of population growth.

For our first set of simulations we used the results from all herbivore removal studies as one pool of effect sizes and results from all herbivore augments as a second pool of effect sizes. Assuming equal variation in the control and herbivore resistant groups, the mean performance of herbivore resistant (HR) individuals in any given year is: (mean seeds per HR individual)_t = sd * (effect size)_t + mean seeds per control individual. We then drew the population size for the next year from a normal distribution with mean = (mean performance per HR individual * N_t), and standard deviation = (sd² * N_t)^0.5, where N_t is the number of individuals in the population at time t. We grew the population for 50 generations, calculated l, the population growth rate for each run and found the average and standard deviation of l across 100 replicate runs of the model. Populations of herbivore resistant weeds are predicted to grow quickly when effect sizes are drawn from removal studies and even more quickly when effect sizes are drawn from augment studies (Figure 2).

Figure 2. Population growth rate (l) for herbivore susceptible plants (control group) versus herbivore resistant plants (removal and augment groups). Mean l’s are for 100 replicates. Error bars are one standard deviation.

We also used our simulations to examine how variation in herbivore impacts affects our ability to predict the risk associated with herbivore resistance genes. For this set of simulations we assembled lists of effect sizes reported from multiple sites but for a single plant species. From the set of papers used in the meta-analysis, we found three papers that reported data for more than two unique sites (Table 2). For each iteration, the model selects a subset of the effect sizes reported for any one plant species with 1, 2, ... n sites per subset where n equals the total number of effect sizes reported. Growth of the plant population is then projected using effect sizes randomly selected each year from the subset. We calculated l for each iteration of the model and found the average and standard deviation for l across 500 replicate runs. We also calculated an error rate for each set of simulations by recording how many times we would erroneously conclude that the herbivore resistant plant was not as significant a pest as indicated by the simulations using data from all possible sites. For example, we recorded how often l was predicted to be less than 1.0, 1.05, and 1.1.

Table 2. References that tested herbivore effects at more than two unique sites.

Reference	Plant species	Herbivore	Comparison
Jordano et al. 1990	Astragulus lusitanicus	Lepidopteran seed predators	5 sites
Louda 1983	Haplopappus venetus	Mixed insect herbivores	3 zones
Brown et al. 1987	Vicia hirsuta	Folivorous and sucking insects	3 successional stages
Brown et al. 1987	Vicia sativa	Folivorous and sucking insects	3 successional stages

Figure 3. Risk posed by herbivore resistance genes. Scatter points are mean l’s for 500 replicate populations of herbivore resistant plants. Larger values of l indicate increased potential weediness of resistant genotypes. The number of sites used to estimate l was varied up to the total number of sites reported by Jordano et al. (1990). Error bars are one standard deviation.

Measuring the impacts of herbivores at multiple sites gives a more consistent estimate of l (Figure 3: standard deviation decreases as the number of sites increases), and the magnitude of herbivore effects is less likely to be underestimated as more sites are sampled (Figure 4: error rate is dramatically reduced as more sites are sampled). The results presented in Figures 3 and 4 are for simulations based only on the results of Jordano et al. (1990), but our findings were qualitatively consistent across sets of simulations tailored to the additional studies listed in Table 2. Clearly, the potential risks associated with herbivore resistance genes can only be accurately assessed when trials are performed at multiple sites that offer potentially different environments for plant growth as well as different background densities of herbivores.

Figure 4. Rate of underestimation of the risk posed by herbivore resistant genes. When all five of the sites from Jordano et al. (1990) are used, the meanl is 1.15 (Figure 3). Error rate is the proportion of the 500 replicates where the predicted l was less than either 1.0, 1.05, or 1.1.

Two points emerge from the analyses above: (i) protecting a plant against herbivory, even against just one species of herbivore, is likely to markedly increase that plant’s rate of reproduction, and (ii) field tests for assessing a plant’s enhanced invasibility are prone to mistakenly assure safety unless they are repeated at multiple sites or under multiple conditions. Both of these points have been anticipated previously (e.g., Williamson 1993; Parker and Kareiva 1996; Kareiva et al. 1996) with the admonition that caution and tenacious monitoring are warranted for certain transgenic crops. Although it will be hard to exercise that caution given the current pressure to ease regulations on the basis of the safe record to date, caution should clearly be maintained. A survey of historical records for past invasions by weeds in the northwestern United States indicated that the median timelag between the first record of a weed and the onset of widespread infestation was on the order of 30-50 years (Marvier et al. 1999). Timelags between the introduction of ornamental woody plants and their escape into the wild in Germany are on the order of 150 years (Kowarik 1995)! Moreover, for most monitoring efforts, early detection of weed problems will remain unlikely and timelags are still expected on the order of decades (Marvier et al. 1999). Examples from the "exotic species" literature are often rejected in the biotechnology arena because it is pointed out that exotic species contain thousands of "novel genes" whereas a transgenic plant contains only a few novel genes. However, the point of learning lessons from exotic species is NOT to claim that such introductions are one and the same as those of transgenic plants, but to point out some inherent features of biological invasions. Those features include extensive timelags and the observation that most invasions require the chance concordance of a suite of favorable conditions before taking off. It is entirely reasonable to expect that invasions of transgenes will share these features.

WHAT DOES THIS MEAN FOR RISK ASSESSMENT?

Like most biological invasions, the majority of transgenic plants will not become pests, and most will have minimal impact. However, it would be arrogant to assume that transgenic plants represent a fail-safe technology. We have previously witnessed such brazen overconfidence. For example, in 1947, the President of the Entomological Society of America gave a presidential address commenting on the implication of pesticides for the control of pest insects. He said,

We are in danger of repeating this blind arrogance with transgenic plants. Fortunately, biotechnology truly has the potential to promote sustainability in ways that chemical insecticides never could. Nonetheless, because engineered resistance genes, such as Bt endotoxin, are entirely novel to their recipient plants (indeed to the entire plant kingdom), they do entail some risk. In addition, we know from ecological field experiments that protection of plants against even small subsets of herbivore species generally causes dramatically enhanced seed production. The information that remains lacking is a broader understanding of what regulates weedy plant populations in nature. Thus, the way to enhance risk assessment of transgenic plants is to explore weed population dynamics and the role of biotic stresses (herbivores and pathogens) in governing those population dynamics. Until we have that understanding, the most sensible strategy is a triage approach that recognizes that most transgenes are likely to be safe, but that resistance genes introduced into crops with close, weedy relatives bear extra scrutiny. A few experiments under a narrow range of conditions should not be accepted as proof of safety; neither should comfort be drawn from any argument of "similarity"—production of Bt toxin is an entirely novel plant trait. Finally, we should not allow a prior record without ecological ill-effects to lull us into complacency—large time lags between introduction and the onset of weed spread are the norm, and we have not been looking very long or all that closely for problems involving transgenes. Herbivory is a potent demographic force in plant populations, and the implications of resistance for plant invasiveness cannot be easily extrapolated across sites or over long periods of time.