Fig. 1

Transitioning to a MBC treatment model incorporating history of response. a Current treatment model for MBC, where patients switch treatments upon progression following NCCN guidelines and physician choice. b Two example patient trajectories, the top showing a typical sequence with progressively decreasing progression free intervals (PFI) from the time of metastatic recurrence. The bottom illustrates a potential departure, with a longer PFI (green) following a shorter PFI (orange). In the top trajectory, the progression free interval PFIⁱ/PFI^i-1 is always < 1, whereas in the bottom, PFIⁱ/PFI^i-1 > 1 for a transition. c Schematic showing a learning/treatment model for MBC. The left-hand side shows a network of possible treatments, where each therapeutic class has multiple agents (e.g., Tx_1.2 = treatment Class 1, agent 2). Upon progression, patients being treated with Agent I in Class J (Tx_i.j) can potentially transition to any other treatment in the network. Transitions are selected based on probabilities generated by an adaptive randomization engine (right). This engine generates predictive probabilities based on the patient’s treatment-response history and molecular phenotype using (1) cell line-based phenotype- and history-dependent response predictions, and (2) a community database of MBC patient Tx and PFI sequences. All treatment-PFI sequences are fed back into the engine to refine predictions. Within the engine, the endpoint is progression free interval ratio PFIⁱ/PFI^i-1. At a higher level of resolution/abstraction, this schema constitutes a SMART MBC learning/treatment system with the goal of identifying the optimal strategy to maximize PFIⁱ/PFI^i-1 (short term) and ∑PFI (long term) using response history and molecular tumor features. In a trial setting the comparator would either be the typical treatment trajectory for MBC patients (A), or possibly the same treatment network with equal probabilities assigned to all treatment transitions, adjusting for subtype