The companion site chemoprevention database proposes a ranking of the most potent agents detected in rodents' studies, and supports the notion that ACF may be used as a surrogate endpoint for tumors in rats. We will discuss here possible bias in the ranking (publication bias and selection process bias) before showing possible extensions of the work. We will then discuss correlation between ACF and tumor, also looking for bias, trying to explain the outlying point, and concluding that ACF may be used as surrogate endpoints. We will not discuss here the mechanisms of inhibition, already addressed in reviews.

By combining six carcinogenesis endpoints from ACF and tumor tables shown in a companion chemoprevention database this review suggests that the most potent agents to prevent colorectal cancer are
- PEG 8000 (polyethylene glycol, a water soluble laxative of high molecular weight),
- a protease inhibitor,
- DFMO alone or with piroxicam or aspirin,
- hesperidin,
- celecoxib,
- sulindac sulfone or sulfide, and
- Bifidobacteria strains. One may prefer to rely on a single endpoint (e.g., cancers), which would result in a slightly different list (see Tumor table). Our non-parametric ranking takes in account both the potency of the agent, and the number of studies published on it. This may be seen as a publication bias, but it is also a measure of the strength of the evidence. We think the most potent agents cited above are promising for cancer prevention, and should be tested in people at risk. However, agents with low potency can also be valuable, particularly those that are naturally present in foods. We like better the idea to prevent cancer by eating intact plant food with the multiplicity of agents that they contain, than by packaging potent anticarcinogenic constituent in a daily pill. However, up to now, we have no direct evidence that the first approach is efficient.

It is likely that the database missed some articles, particularly those published before 1989 and very recent ones. However, we think that no potent agents could be missed: the early comprehensive review of Angres and Beth points out the protective effect of wheat bran, cellulose, low fat, selenium and caloric restriction. Each of these factors is cited several times in table 2, but we did not report all early studies. For instance, although we reported many studies on the chemoprevention by fish oil or by n-3 fatty acids, only five studies on low-fat diet were included. One reason is that a low-fat diet is not exactly a "chemopreventive agent." Another reason is that the protection afforded by low-fat diets is often small. Indeed, a meticulous review of 14 studies of dietary fat and rat colon carcinoma shows that fat has no effect in one study out of two (298). Specifically, no association between colon cancer incidence and fat intake is seen for Sprague-Dawley rats, but a positive relationship is indicated for Fischer 344 rats. When the degree of saturation is taken into account, only n-6 polyunsaturated fat intake significantly promotes the cancers. We used the logistic regression analysis from the quantitative review (298), to calculate the "potency" of a low-fat diet to reduce the cancer incidence in Fisher 344 rats. Compared with a high-fat diet (20% fat), the median potency of a low-fat diet (5% fat) to reduce the cancer incidence was 1.3, which is in the bottom-ten of table 2.

Because our first aim was to find the most potent agents against colon carcinogenesis, we used a selection process to build the tables. The result could thus mislead the reader for some agents:
(i) Some potent agents are not shown in the tables because they were hidden by a more potent one in the same article. This is the case, for instance, of limonin. This citrus limonoid was reported in the same article as obacunone (54). Both were very potent to prevent ACF and tumors, but we dropped limonin since obacunone was more potent.
(ii) Some agents rank very high in the tables, but are less potent in duplicated studies. For instance, MMTS reduces 8-fold the tumor incidence in a first study (170), but has little effect (1.1-fold reduction) in a duplicated study (285).
(iii) Some agents might have been dropped from the tables, if negative studies had been taken in account too, and if mean potencies have been calculated for each agent. For instance, a specific nitric oxide synthase inhibitor prevents colon carcinogenesis (147), but can also promote carcinogenesis (299). Similarly, many agents shown here as preventive, did enhance carcinogenesis in other rat studies (e.g., beta-sitosterol, benzyl- and phenylhexyl- isothiocyanates, calcium, cellulose, diallyl sulfide, folic acid, genistein, germfree status, glucarate, pectin, quercetin, resistant starch, rutin, selenium, tea extracts, vegetables and fruits mixture, and vitamin D3). Thus, the tables may help to find the most potent agents, but cannot be used to calculate the average effect of a given agent.

The present work could be extended in four directions:
(i) The tables could be updated when new agents are published. We propose to do this on a slow access website with sorting abilities http://corpet.free.fr.
(ii) It could be useful to have a comprehensive view of all the studies on colon cancer chemoprevention in rodents. Thus, all results from each paper should be included in the tables, including less potent agents and those that promote carcinogenesis. This would allow the calculation of the mean potency of each agent. However, since most negative results are not published, a large publication bias would be inevitable. We decided not to do this in the present database, to produce tables of reasonable size.
(iii) Many agents have also been tested in the Min mouse model, or similar models (e.g., truncated Apc, Msh2, Mlh1). The potency of dietary agents to decrease the number of polyps in these mice could be gathered and ranked in a table. Indeed, NSAIDs decrease the number of polyps in the small intestine of mutant mice. But surprisingly, some potent chemopreventive agents against chemically-induced tumors, namely PEG, celecoxib, piroxicam, sulindac, and DFMO can increase the polyp number or size in the COLON of mutant mice. The reason for this puzzling discrepancy is not clearly understood. It may be a result of differences in key enzymes in the small and large intestines (305). Indeed, a Min mice database is proposed on a website with sorting abilities http://corpet.free.fr.
(iv) Any decision to test an agent in a human clinical trial should rely not only on the efficacy of the agent, but also on other considerations: lack of toxicity, the paucity of side-effects, acceptability (e.g., no taste), and price. The chemopreventive doses reported in this review are not toxic for rats. It would be useful however to review accurately the human safety of these agents, but this is out of this paper scope.

A significant correlation was found between the potencies in the ACF assay and in the tumor assay. This finding is weakened by a selection bias and a publication bias. This review gathers agents that inhibit ACF or tumors. It does not report null studies, or agents with promoting properties. Thus, the study could not show strong discrepancies between ACF and tumor data. For instance, an agent that reduces the number of ACF, but increases the tumor incidence could not appear in a correlation, e.g., 2-(carboxyphenyl) retinamide (73, 86). Moreover, many negative studies are never published, and this leads to a publication bias that we cannot overcome. However, out of 40 agents, Steele et al. showed that of the 30 agents tested as active in the ACF assay, 21 prevented colon cancer in rats (306). It is thus likely that correlation could remain true, even if discrepant agents were included in the study. Celecoxib is very potent against tumors, but not against ACF. It thus appeared as a major outlier in the tumor-ACF correlation (fig. 2). Three explanations may be given. Celecoxib could inhibit tumors at a late phase of carcinogenesis (182), maybe by reducing angiogenesis in tumors. It would not inhibit the growth of ACF, since they do not require extra-supply of blood. A second explanation is that celecoxib could inhibit the growth of intraepithelial neoplasia near lymphoid aggregates. These lesions may have a high potential to grow to cancers, but do not proturb in the gut lumen, and could not be counted as ACF (307). Last, the celecoxib potency on ACF may have been under-estimated by chance, or the potency on tumor over-estimated, since each value comes from a single study.

Conclusion: Finally, we propose that the potency of each new agent, or the result of each new chemoprevention study reported in the literature, should be compared with previously published agents. The present review shows that the median potency of effective agents published so far is two. A new agent that leads to a twofold reduction in the ACF number or in the tumor incidence is thus an average one. In contrast, an agent that reduces the ACF number or the tumor incidence more than fivefold is of outstanding potency.