Complex Adaptive Systems (CAS) are dynamical networks of interacting agents which as a whole determine the behavior, adaptivity and cognitive ability of the system. CAS are ubiquitous and occur in a variety of natural and artificial systems (e.g., cells, societies, stock markets). To study CAS, Holland proposed to employ an agent-based system in which Learning Classifier Systems (LCS) were used to determine the agents behavior and adaptivity. We argue that LCS are limited for the study of CAS: the rule-discovery mechanism is pre-specified and may limit the evolvability of CAS. Secondly, LCS distinguish a demarcation between messages and rules, however operations are reflexive in CAS, e.g., in a cell, an agent (a molecule) may both act as a message (substrate) and as a catalyst (rule). To address these issues, we proposed the Molecular Classifier Systems (MCS.b), a string-based Artificial Chemistry based on Holland’s broadcast language. In the MCS.b, no explicit fitness function or rule discovery mechanism is specified, moreover no distinction is made between messages and rules. In the context of the ESIGNET project, we employ the MCS.b to study a subclass of CAS: Cell Signaling Networks (CSNs) which are complex biochemical networks responsible for coordinating cellular activities. As CSNs occur in cells, these networks must replicate themselves prior to cell division. In this paper we present a series of experiments focusing on the self-replication ability of these CAS. Results indicate counter intuitive outcomes as opposed to those inferred from the literature. This work highlights the current deficit of a theoretical framework for the study of Artificial Chemistries.