This dissertation presents novel computational methods to infer accurate functional models of musculoskeletal systems from minimal experimental data focussing on the tendon networks of the human fingers as an example. State of the art biomechanical modeling consists of assuming a fixed structure for the system being modeled and measurement or regression of specific parameter values. However, the assumed structure may not be the best representation of the system and hence lead to a functionally less-accurate model. The objective here is to simultaneously infer both the structure and the parameter values directly from experimental input-output data. We present novel methods to infer computational models of two kinds– analytical models capturing input-output behavior without specifically modeling the mechanics of the system, and anatomy-based models that explicitly capture the mechanics of interactions of the constitutive elements. Using experimental data from a tendon-driven robotic system and synthetic data from simulated musculoskeletal systems, a novel method based on symbolic regression using genetic programming that simultaneously infers the form and parameter values of mathematical expressions is presented and shown to outperform polynomial regression, the state of the art method used in musculoskeletal modeling. This method is then implemented on experimental data collected from a cadaveric index finger to obtain accurate analytical functions for the tendon excursions of the seven tendons of the finger. Whether the goal is to obtain accurate subject-specific models or to obtain generalizable models, the functions obtained using this novel technique are more accurate than both polynomial regressions and Landsmeer-based models, both of which have been used in the literature. Experimental control of a cadaveric index finger to produce simple finger movements gives some insight on how a muscle-like spring based control can be more advantageous than using simple force or position control. Two different kinds of equilibria is demonstrated in the human cadaveric index finger, for the first time to our knowledge, and their relationship to the null space of the moment arm matrix is studied. An experimental validation, of some common models of the index finger in their ability to predict fingertip output is presented and it is shown that the fingertip force is sensitive to moment arm values. A novel non-linear finite element method based solver is developed that can be used to model the interactions of elastic tendon networks on arbitrarily shaped bones. This solver which is validated using experimental data is then used to model the extensor mechanism draped on the finger bones and a local sensitivity analysis is performed to see how the fingertip force output is affected by changes to the properties of the network. It is concluded that fingertip force output is most sensitive to topology and the resting lengths of the bands of the extensor mechanism. Finally, a novel inference algorithm that is based on the co-evolution of models and tests is used to infer the parameters and topology of a 3 dimensional model of the extensor mechanism directly from cadaveric data with minimal number of experimental data points through intelligent testing. It is shown that the inferred models are more accurate in predicting fingertip force magnitude and direction compared to a model popularly used in the literature.