Background: Many natural and synthetic compounds might prove to be effective in cancer chemotherapy. To identify potentially useful agents, the National Cancer Institute screens over 10,000 compounds annually against a panel of 60 distinct human tumor cell lines in vitro. This screening program generates large amounts of data that are organized into relational databases. Important questions concern the information content of the data and ways to extract that information. Previously, statistical techniques have revealed that compounds with similar patterns of activity against the 60 cell lines are often similar in structure and mechanism of action. Feed-forward, back-propagation neural networks have been trained on this type of data to predict broadly defined mechanisms of action of chemotherapeutic agents.
Purpose and method: In this report, we examine the information that can be extracted from the screening data by means of another type of neural network paradigm, the Kohonen self-organizing map. This is a topology-preserving function, obtained by unsupervised learning, that nonlinearly projects the high-dimensional activity patterns into two dimensions. Our dataset is almost identical to that used in the earlier neural network study.
Results: The self-organizing maps we constructed have several important characteristics. 1) They partition the two-dimensional array into distinct regions, each of which is principally occupied by agents having the same broadly defined mechanism of action. 2) These regions can be resolved into distinct subregions that conform to plausible submechanisms and chemically defined subgroups of submechanism. 3) These results (and exceptions to them) are consistent with those obtained with the use of such deterministic measures of similarity among activity patterns as the Euclidean distance or Pearson correlation coefficient.
Conclusions: Our results indicate that the activity patterns obtained from the screen contain detailed information about mechanism of action and its basis in chemical structure. The self-organizing map can be used to suggest the mechanism of action of compounds identified by the screen as potentially useful chemotherapeutic agents and to probe the biology of the cell lines in the cancer screen. Kohonen self-organizing maps, unlike the previously applied neural networks, preserve and reveal the relationships among compounds acting by similar mechanisms and therefore have the potential to identify compounds that act by novel cytotoxic mechanisms.