Inference of transcriptional regulatory networks is an important filed of research in systems biology, and many computational methods have been developed to infer regulatory networks from different types of genomic data. One of the most popular classes of computational network inference methods is expression based network inference. Given the mRNA levels of genes, these methods reconstruct a network between regulatory genes (called transcription factors) and potential target genes that best explains the input data. However, it has been shown that the networks that are inferred only using expression, have low agreement with experimentally validated physical regulatory interactions. In recent years, many methods have been developed to improve the accuracy of these computational methods by incorporating additional data types. In this dissertation, we describe our contributions towards advancing the state of the art in this field. Our first contribution, is developing a prior-based network inference method, MERLIN-P. MERLIN-P uses both expression of genes, and prior knowledge of interactions between regulatory genes and their potential targets, and infers a network that is supported by both expression and prior knowledge. Using a logistic function, MERLIN-P could incorporate and combine multiple sources of prior knowledge. The inferred networks in yeast, outperform state of the art expression based network inference methods, and perform better or at a par with prior based state of the art method. Our second contribution, is developing a method to estimate transcription factor activity from a noisy prior network, NCA+LASSO. Network Component Analysis (NCA), is a computational method that given expression of target genes and a (potentially incomplete and noisy) network structure that describes the connection of regulatory genes to these target genes, estimates unobserved activity of the regulators (transcription factor activities, TFA). It has been shown that using TFA can improve the quality of inferred networks. However, our prior knowledge in new contexts could be incomplete and noisy, and we do not know to what extent presence of noise in input network affects the quality of estimated TFA. We first show how presence of noise in the input prior network can decrease the quality of estimated TFA, and then show that by adding a regularization term, we can improve the quality of the estimated TFA. We show that using estimated TFA instead of just expression of TFs in network inference, improves the agreement of inferred networks to experimentally validated physical interactions, for all state of the art methods, including MERLIN-P. Our final contribution, is developing a multi-task inference method, Dynamic Regulatory Module Network (DRMN), that simultaneously infers regulatory networks for related cell lines, while taking into account the expected similarity of the cell lines. Many biological contexts are hierarchically related, and leveraging the similarity of these contexts could help us infer more accurate regulatory programs in each context. However, the small number of measurements in each context makes the inference of regulatory networks challenging. By inferring regulatory programs at module level (groups of co-expressed genes), DRMN is able to handle the small number of measurements, while the use of multi-task learning allows for incorporation of hierarchical relationship of contexts. DRMN first infers modules of co-expressed genes in each cell line, then infers a regulatory network for each module, and iteratively updates the inferred modules to reflect both co-expression and co-regulation, and updates the inferred networks to reflect the updated modules. We assess the accuracy of the inferred networks by predicting the expression on hold out genes, and show that the resulting modules and networks, provide insight into the process of differentiation between these related cell lines. For all the developed methods, we validate our results by comparing to known experimentally validated networks, and show that our results provide useful insight into the biological processes under consideration. Specifically, in chapter 2, we evaluated our inferred networks based on both network structure and predictive power, identified TFs that all tested methods fail to recover their target sets, and explored potential reasons that can explain this failure. Additionally, we used our method to infer stress specific networks, and evaluated predictions using stress specific knock-down experiments. In chapter 3, we evaluated our inferred networks based on both network structure and predictive power, and furthermore used our inferred networks to identify potential regulators that could be important for pluripotency state in mESC. We tested the effect of these regulators using shRNA experiments, and experimentally validated some of their predicted targets. Finally, in chapter 4, we evaluated our inferred models based on their predictive power and ability to predict gene expression in hold out data.