Transcranial Magnetic Stimulation (TMS) is a noninvasive brain stimulation technique that is increasingly used in clinical neuroscience research. Since the invention of TMS in 1985, new applications and stimulation patterns for this technique have been developed and advanced very fast. However, compared with other research techniques used in neuroscience, the use of computational tools for TMS is limited. The studies introduced in this thesis applied computational methods adapted from the field of machine learning to TMS experimental procedures and data analysis, and inspired the idea of computational TMS. These studies showed that computational TMS improved the accuracy and efficiency of TMS, and could be used to investigate neuroscience problems. In Chapter 1, the background, basic principles, and variations of TMS techniques are introduced, and the idea of computational TMS is motivated. Then three example computational TMS studies are demonstrated: Chapter 2 introduces a new protocol facilitating the TMS motor threshold (MT) estimation. This protocol uses Bayesian framework to incorporate prior knowledge of MT and is two to five times faster than the fastest existing method. Chapter 3 introduces a fast protocol for TMS mapping. This protocol uses high-resolution coordinate data obtained from a neuronavigation system and non-parametric regression techniques to generate TMS maps. Chapter 4 is a quantitative analysis of the relationship between Motor Evoked Potential (MEP) and TMS coil placement, which highlights the importance of TMS spatial information. This analysis also supports the neurophysiology hypothesis that TMS excites the column structures on the anterior bank of central sulcus. Chapter 5 includes conclusion and future work of computational TMS.