Need of Systems Approach for Biological Explanation of Anti-learnable Signatures

We present simple formal models explaining unusual properties of several biological classification tasks as follows. For these datasets the whole range of supervised learning techniques generate predictive models which classify independent test samples systematically below the performance of random guessing (hence the name anti-learning). We show that explanation of such “counter-intuitive” supervised learning results is possible by modelling interactions between multiple subsystems (a system biology approach). The presentation is focussed on a particular cancer genomics example and a matching synthetic dataset, an outcome of a simple zero-sum game that models species competing for limiting resources. Some other biologically plausible models are discussed briefly.