With the standard agent database, the COMPARE algorithm had been particularly effective with antimitotic agents directed against tubulin. The top 10 matches with each known antimitotic agent in the database included all other such agents in the database. Consequently, in 1990 when only a few thousand compounds had passed through the screen, we began a study to determine whether COMPARE could select other antimitotic agents that had been examined for cytotoxicity against the 60 cell lines and whether entirely novel antimitotic agents could be identified. We knew that such agents included colchicine, dolastatin 10 (46, 53), dolastatin 15 (70), combretastatin A-4 (71, 72), podophyllotoxin, carbamates (73), and benzylbenzodioxole derivatives (74). The entire database was probed at that time with colchicine, vinblastine, vincristine, podophyllotoxin, and paclitaxel, and we arbitrarily examined the best 100 matches with each. We found that no matter what antimitotic agent was used as the seed, whether one of those in the standard agent database (vinblastine, vincristine, or paclitaxel) or one of those that had been separately screened (colchicine or podophyllotoxin), most known antimitotic compounds in the overall database appeared among the top 100 matches.
In this initial work, we noted two potently cytotoxic marine natural products that appeared repeatedly. These were the complex macrolide polyethers halichondrin B and homohalichondrin B (75, 76). These natural products were specifically examined for interactions with tubulin and for antimitotic activity, and such properties were confirmed (6). This work demonstrated that the halichondrin B noncompetitively inhibited the binding of vinblastine to tubulin, and thus these agents may bind at a unique site on the protein.
This success led us to examine compounds on the five lists with structures we considered novel for antimitotic agents for effects on in vitro tubulin polymerization (7). Positive and negative compounds in this assay were then evaluated in terms of their cytotoxicity with the human tumor cell lines and in terms of their PCC values relative to the five seeds. This analysis led us to conclude that the COMPARE algorithm would yield optimal results with antimitotic agents if we imposed two restrictions on the compounds selected with any seed. First, the PCC should be at least 0.6. Second, these initial results indicated that compounds with low cytotoxicity generally did not greatly affect tubulin polymerization. Therefore, we imposed as a second criterion that selected compounds have a GI50 value of 1 µ M or less in the original screen with HL-60 (TB) human leukemia cells.
We then performed our ``definitive'' searches of the overall database in late 1990 at a time when over 7000 compounds had been screened. We used nine seeds (vinblastine, colchicine, podophyllotoxin, vincristine, paclitaxel, maytansine, dolastatin 10, rhizoxin, and combretastatin A-4). Beside the seeds themselves, 73 compounds in the database met the criteria summarized previously. Among these compounds were 13 analogues of podophyllotoxin, 3 of colchicine, 9 of dolastatin 10, 7 of combretastatin A-4, 3 of paclitaxel, 3 of carbamates, and 2 of benzylbenzodioxole derivatives. There were 32 structurally novel compounds representing 19 distinct chemical species. Two of these were halichondrin B and homohalichondrin B, which had been identified by seven of the seeds. Because halichondrin B may have a unique binding site on tubulin (6), we used it as well as a seed for COMPARE but no additional compounds were selected from the database.
The antitubulin and antimitotic properties of the other 30 compounds, including their structures, have been presented in detail elsewhere (7). In summary, 20 compounds (representing 11 distinct chemical species) were effective inhibitors of tubulin polymerization and caused the accumulation of cells arrested in mitosis in tissue culture. These were all synthetic compounds with relatively simple chemical structures, and they all interfered with the binding of colchicine to tubulin. All but one of these compounds were identified by at least six seeds, and the remaining compound was identified by a single seed. One compound, identified by a single seed, weakly inhibited tubulin polymerization and colchicine binding, but we could not demonstrate accumulation of mitotic cells with this agent even though it was cytotoxic. Seven compounds, representing four chemical groups, had no affect on tubulin polymerization or accumulation of cells arrested in mitosis. Only one of these agents was identified by multiple seeds, and we were unable to confirm the cytotoxicity observed in the original screening studies. This finding suggests that this compound was chemically unstable. Finally, two compounds were particularly interesting. These two agents, one of which was tritylcysteine, were identified by multiple seeds, and they both caused the accumulation of cells arrested in mitosis in tissue culture. Neither compound interacted with tubulin in vitro. Further studies are required to determine whether microtubules are the cellular target of these two compounds or whether their target is another cellular component involved in the mitotic process.
Therefore, we conclude that we have established a reasonable scheme of using COMPARE to identify new antimitotic compounds that have a high probability of interacting with tubulin. Our overall data indicate that we could reduce the number of false positives found with COMPARE by adding a third criterion, identification by more than four seeds. We hesitate to do this, however, because the potent antimitotic natural product dolastatin 15 (70) was found with only two seeds.