Supplementary Materials from Automated multi-objective calibration of biological agent-based simulations

Computational agent-based simulation is increasingly used to complement laboratory techniques in advancing our understanding of biological systems. Calibration, the identification of parameter values that align simulation with biological behaviours, becomes challenging as increasingly complex biological domains are simulated. Complex domains cannot be characterised by single metrics alone, rendering simulation calibration a fundamentally multi-metric optimisation problem that typical calibration techniques cannot handle. Yet calibration is an essential activity in simulation-based science; the baseline calibration forms a control for subsequent experimentation, and hence is fundamental in the interpretation of results. Here we develop and showcase a method, built around multi-objective optimisation, for calibrating agent-based simulations against complex target behaviours requiring several metrics (termed <i>objectives</i>) to characterise. Multi-objective calibration delivers those sets of parameter values representing optimal tradeoffs in simulation performance against each metric, in the form of a Pareto front. We use MOC to calibrate a well-understood immunological simulation against both established <i>a priori</i> and previously unestablished target behaviours. Further, we show that simulation-borne conclusions are broadly, but not entirely, robust to adopting baseline parameter values from different extremes of the Pareto front, highlighting the importance of MOC's identification of numerous calibration solutions. We devise a method for detecting overfitting in a multi-objective context, not previously possible, used to save computational effort by terminating MOC when no improved solutions will be found. MOC can significantly impact biological simulation, adding rigour to and speeding up an otherwise time-consuming calibration process, and highlighting inappropriate biological capture by simulations that cannot be well calibrated. As such, it produces more accurate simulations that generate more informative biological predictions.