In this paper, we propose a framework of biome-chanics analysis by estimating unknown dynamics from partial sensing information, exploring the minimalism of motion cap-ture sensing. A preliminary experiment reveals that Centroidal Dynamics in a cyclic motion such as running can be extracted as a simple time-series profile. Based on this result, we apply an LSTM network to the estimation of unknown upper-body Centroidal Dynamics from measured lower-body information. Estimated result allows us to compute the inverse dynamics of the human musculoskeletal model. We report an experimental validation for an outdoor running motion using an IMU motion capture suit.