Estimating the real dose distribution during proton therapy through experimental measurements is key to optimizing treatment planning and ensuring treatment quality. Positron emission correlates with dose distribution, but this correlation is non-linear and difficult to estimate in practical situations. We propose a deep learning-based method that utilizes raw coincidence data from readouts of palm-sized PET modules and predicts the positron activity and dose distributions bypassing traditional reconstruction steps. An extensive simulation dataset was generated using the GATE/Geant4 10.4 toolkit and a human CT phantom with a compact in-beam PET imaging setup for realistic single-fraction dose delivery. Conditional generative adversarial networks were trained to generate dose maps conditioned to the coincidence distributions on the detector. The generated dose maps are used to validate the range against simulated output. We evaluated the model performance through mean relative error (MRE), absolute dose fraction difference, and shift in Bragg peak position, and present it as a function of the detector threshold, the number of coincidences required, and other conditions that optimize the predictive power of the model. The results show that the model can predict the Bragg peak position and dose for mono-energetic irradiation between 50 MeV and 122 MeV with uncertainties less than 1% and 2%, respectively, with 10 5 coincidences acquired for five minutes post-irradiation. An important aspect of this simulation study is the use of compact detectors applicable in a treatment scenario, which operate on very low counts and the demonstration of a method for direct reconstruction. Through our analysis, we can continue to expand the scope of inverse modeling to simulate realistic data by learning imaging and experimental physics in radiation therapy from observations.