A primary goal in systems neuroscience is to understand how neuronal connectivity impacts the activity and dynamics in mechanistic models of neural processing. New connectome datasets of the fruit fly provide access to the connectivity of every neuron in its neural circuits. However, the functional role of recurrent connections in the fruit fly remains incompletely understood. For instance, although its visual system is predominantly feedforward, strong feedback connections have been identified in its motion detecting pathways [1]. What is the role of recurrent connections in the visual perception of fruit flies? We investigate the impact of recurrent connectivity in ON-motion selectivity using connectome-constrained deep mechanistic networks (DMNs) of the fruit fly’s visual system [2]. Every neuron and synaptic connection in the fly have a one-to-one correspondence in the DMN, enabling detailed predictions of neural function. To capture the nonlinear neural computations in motion detection [3], we incorporated more biologically-realistic conductance-based synapse dynamics, instead of current-based ones. Indeed, we found that the conductance-based models correctly predict supralinear input integration, in addition to accurately predicting known single-neuron tuning properties. We then performed ablations of cell-type connectivity in the DMNs to generate experimentally testable hypotheses about the computational role of these connections. To validate our model, we first confirmed that ablations of major T4 input cell types capture existing experimental findings [3]. We then ablated the excitatory lateral T4 to T4 connections, expecting a decrease in the amplitude of T4 responses to motion. Indeed, we found a weakened response of T4 cells to motion towards all directions, maintaining T4 direction selectivity. Finally, we investigated the unknown computational role of the recurrent inhibition between Mi4 and Mi9 cells, the major inhibitory T4 inputs, by ablating these connections. Notably, these ablations decrease T4 direction selectivity, weakening the preferred direction response and strengthening non-preferred direction responses. This suggests that the Mi4-Mi9 recurrency sharpens T4 motion detection. In summary, we equip DMNs with more biologically-realistic synapse dynamics and predict known ablation effects on ON-motion computation. We then make new detailed predictions for the role of recurrent connections in the fruit fly’s motion detection pathway.