How does REM sleep fit the synaptic homeostasis hypothesis when most downscaling evidence comes from NREM or total sleep?

Why REM is not the main downscaling stage

The synaptic homeostasis hypothesis proposes that sleep serves to renormalize synaptic strength, and it explicitly considers a possible role for REM sleep in synaptic downscaling [4]. That framing places REM inside the hypothesis as a candidate contributor, not as the primary evidence base. The hypothesis does not require that all downscaling happen in one stage; it allows downscaling to be distributed or to interact with other sleep processes.

In computational models of continual learning, the three sleep components are kept distinct: NREM is modeled as veridical memory replay, REM as generative memory replay, and synaptic downscaling as a separate size-dependent weight adjustment [1][2]. This separation mirrors the idea that REM contributes replay and reorganization, while downscaling is a distinct mechanism that can be modeled on its own. The models show benefits from including all three components together, which is consistent with REM playing a complementary role rather than being the downscaling stage itself [1][2].

What REM may contribute instead

A review of sleep EEG oscillations describes synaptic homeostasis as global downscaling that occurs alongside selective strengthening of memory-relevant connections, and it links REM sleep to theta oscillations [3]. That description suggests REM may be part of a broader sleep-wide process that both weakens and strengthens synapses, rather than being solely a downscaling stage. The review's abstract does not specify that REM itself performs the global downscaling; it presents downscaling and selective strengthening as concurrent features of sleep plasticity.

The modeling work gives a concrete example of what REM might do: generative replay of previous tasks. In the models, REM-like generative replay helped reduce catastrophic forgetting when combined with NREM replay and downscaling [1][2]. This is a proposed function in an artificial system, not a direct measurement of biological REM, but it illustrates how REM could support memory integration without being the downscaling mechanism.

How downscaling and REM might interact

The models also reveal a trade-off in downscaling strength: more aggressive downscaling better protected early tasks, while less downscaling enhanced learning of new tasks, and intermediate levels gave the highest overall accuracy [1][2]. If REM contributes replay that helps recover or integrate earlier tasks, then the balance between downscaling and replay could matter for how well REM's contribution is expressed. This is a model-based suggestion, not a demonstrated biological interaction.

Separately, a Drosophila study of Alzheimer's-related APP overexpression found defects in synaptic downscaling and in sleep deprivation-induced sleep rebound, both restored by inhibiting IP3R [5]. That study does not examine REM sleep, so it does not directly link REM to downscaling. It does show that downscaling and sleep homeostasis can share molecular machinery, which is consistent with the idea that downscaling is not confined to one sleep stage but is tied to homeostatic sleep processes more broadly.

What REM adds to the hypothesis

Taken together, the evidence supports a view in which REM is not the primary downscaling stage but a proposed partner in sleep-dependent plasticity. The hypothesis itself leaves room for REM in downscaling [4], reviews describe downscaling as concurrent with selective strengthening across sleep [3], and models treat REM as generative replay distinct from downscaling [1][2]. The molecular study adds that downscaling can be linked to homeostatic sleep mechanisms without specifying REM [5]. The practical answer is that REM fits the hypothesis as a complementary process—likely involving replay and reorganization—while the strongest downscaling evidence remains associated with NREM or total sleep.

Sources used in this answer

Related research questions