Proteostasis modules causally enable memory consolidation

When a neuron converts a fleeting experience into a lasting memory, which of its cellular subsystems switch on — and in what order? Scored across seven transcriptomic datasets with the predictions fixed in advance, consolidation resolves into a clean, layered cellular program — and one arm of it is shown to be causally required.

Pre-registered · 26 predictions 7 datasets + 2 replication Causal · TLR9 knockout 32 literature experiments · 0 contradictions

Memory consolidation is usually studied as two separate literatures — the signaling/immediate-early-gene story (CREB, ERK) and the proteostasis story (chaperones, the unfolded protein response, autophagy) — that rarely meet. Read through one common vocabulary of 44 cellular sensing modules, they turn out to be halves of a single, temporally-ordered architecture.

Headline. Consolidation engages a three-layer architecture: a perception layer (cAMP/CREB, ERK/FOS, NFAT, NPAS4) that senses the encoding trigger — transient, post-translational; a sentinel sub-layer (UPR-ATF6, ATF4, PERK) that senses accumulated proteostatic load — persistent; and a constructive sub-layer (HSF1, NRF2, autophagy, IRE1) that actively builds new proteostatic machinery. The three separate cleanly in time, mechanism, and cell-biological role — and TLR9 knockout selectively removes one arm, with a matching memory deficit.

predictions registered before analysis

d = +1.40

HSF1 (build machinery) at 24 h — larger than CREB

60–75%

of ER-chaperone induction lost in TLR9 knockout

28 / 32

published perturbations predicted correctly · 0 contradictions

The three layers, and how they separate

Every module is read two ways: readiness (are the components transcribed?) and activity (are the target genes actually induced?). That distinction is what resolves the architecture — several key modules activate post-translationally, invisible to component-only scoring.

Perception — cAMP/CREB, ERK/FOS, NFAT, and NPAS4 sense the encoding event. Activation is transient and post-translational: CREB target activity is d = +0.80 in dentate gyrus at 24 h and back to d = −0.19 by 96 h.
Sentinel — the UPR-ATF6 / ATF4 / PERK axis senses accumulated protein-folding load. It persists: UPR-ATF6 is still climbing at d = +1.04 at 96 h, when perception has fully relaxed — and the elevation is enriched in engram-like neurons, not a uniform tissue-wide stress response.
Constructive — HSF1, NRF2, autophagy, and IRE1 build the folding and quality-control machinery. This is the largest signal in the whole perceptome during active consolidation (HSF1 d = +1.40, exceeding CREB by 75%), and it re-engages at recall.

The layers dissociate along a temporal axis (transient vs persistent), a regulatory axis (post-translational vs transcriptional), and a functional axis (sense the trigger → sense the load → build the machinery).

Same architecture, region-specific implementation

Four amygdala predictions — that hippocampal architecture would simply generalize — were pre-registered and all failed, in a structured way that is the point rather than a disappointment.

Hippocampus (dentate gyrus) induces both ER and cytoplasmic chaperones plus autophagy — the signature of building new synapses.
Amygdala induces ER chaperones but suppresses cytoplasmic chaperones and autophagy, against a 1.8-fold higher baseline chaperone reserve — the signature of swapping receptors in existing synapses. It already has the capacity; it reallocates rather than builds.

Two opposite transcriptional solutions to the same constraint, against different starting reserves.

The causal test

Correlations across timepoints establish the architecture; one genetic manipulation makes it causal. In Tlr9⁻ᐟ⁻ mice (Jovasevic et al., 2024), ER-chaperone induction is selectively blunted by 60–75% across hippocampal subfields while cytoplasmic chaperones are preserved — a clean, arm-specific lesion of the infrastructure — and those mice show impaired contextual fear memory. Remove one branch of the constructive layer, lose the memory.

It agrees with three decades of perturbations

The model was checked against 32 published proteostasis-manipulation experiments — HSP90 inhibitors, HSP70, chemical chaperones, XBP1, ISRIB, PERK, autophagy, mTOR, NRF2, HSF1 knockouts — spanning mouse, rat, and fly. It predicted the correct memory direction in 28 (full match), 4 partial, and zero contradictions. It even resolves a standing paradox: HSP90 inhibitors are anti-cancer drugs that enhance memory — because inhibiting HSP90 releases HSF1 and turns the constructive layer up.

A method that travels

The readiness vs activity distinction — scoring a module by its target-gene output, not just component abundance — is a general refinement. It captured post-translational activation that component scoring misses, and it validates cleanly out of domain: zero sign reversals in Alzheimer’s neurons (0/8) and a 9.4% reversal rate across four aging tissues. NPAS4 is proposed as the 44th canonical perceptome module — the first tissue-restricted one.

The framework yields a falsifiable therapeutic prediction: chemical chaperones (4-PBA, TUDCA) should rescue memory deficits in TLR9-deficient and aging-impaired contexts.

Built on the perceptome framework · sole-author manuscript, submission-ready · pre-registration and code archive to accompany the preprint. Contact tspiro@vaika.org.