Purpose. The aim of this study is to verify the effectiveness of the learning strategy for collective classifiers based on population dynamics in the coexistence regime of competition driven by training samples, when applied to a classifier with response described by a model of a synthetic gene circuit, and to compare the learning strategies based on the winner-take-all and coexistence competition regimes.

Methods. We use the models of competitive population dynamics, which were suggested previously for describing collective classifier training, in the form of ordinary differential equation systems. The classifier response is described by a previously suggested model of a synthetic gene circuit. The efficiency of training is characterized by the trained classifier’s correct answer probability, which is estimated by its responses to testing samples. We compare the results for the two learning strategies, using two classification problems as examples, with different ways to determine the classification threshold (namely, by an analytical expression and by optimization), and varying classifier parameters.

Results. The coexistence competition learning strategy produces a better or similar result as compared to the winnertake-all strategy, dependent upon the specifics of the classification problem and of the learning process. The former strategy, unlike the latter, does not require neither the use of an external cell sorter nor terminating the learning process in proper time, and admits a quasioptimal analytical estimate of the classification threshold. This is achieved at the cost of an increased complexity of the system, imposing a limitation on the quantity of cell types constituting the classifier, and slowing down the learning process.

Conclusion. A collective classifier trained by competitive dynamics in the coexistence regime is essentially an artificial ecological system with competition driven by training samples. This learning strategy is complicated in terms of biological implementation, but it provides a step towards creating self-learnable multicellular classifiers.