Recent years have seen a growing interest in intelligent game-based learning environments featuring virtual agents. A key challenge posed by incorporating virtual agents in game-based learning environments is dynamically determining the dialogue moves they should make in order to best support students' problem solving. This paper presents a data-driven modeling approach that uses a Wizard-of-Oz framework to predict human wizards' dialogue acts based on a sequence of multimodal data streams of student interactions with a game-based learning environment. To effectively deal with multiple, parallel sequential data streams, this paper investigates two sequence-labeling techniques: long short-term memory networks (LSTMs) and conditional random fields. We train predictive models utilizing data corpora collected from two Wizard-of-Oz experiments in which a human wizard played the role of the virtual agent unbeknownst to the student. Empirical results suggest that LSTMs that utilize game trace logs and facial action units achieve the highest predictive accuracy. This work can inform the design of intelligent virtual agents that leverage rich multimodal student interaction data in game-based learning environments.