Instructional quality dimensions of cognitive demands and instructional support have been shown to have an impact on students' learning gains. Existing operationalizations of these dimensions have mostly used comprehensive ratings that combine various subdimensions of task quality and interaction quality. The current study disentangles interaction quality in a video data corpus study (of 49 middle school classrooms sharing the same tasks) to identify those quality features that predict students' learning gains in conceptual understanding. The regression analysis reveals that quality features of students' individual engagement do not predict individual student learning, whereas teachers' support of learning content-relevant vocabulary predicts the small groups' learning. For at-risk students, the collective time spent on conceptual practices (i.e. explaining meanings of concepts) on students' learning is significantly predictive. The observation that different operationalizations (for similar aspects of interaction quality) lead to different impacts on the learning gains contributes to ongoing research efforts to refine and increase insight into aspects of interaction quality.