Autism Spectrum Disorder as a Signed Failure of Developmental Encoding

1. Introduction

Autism spectrum disorder has been described as a critical-period disorder for more than a decade (LeBlanc and Fagiolini, 2011). The proposal is attractive: ASD is neurodevelopmental, its behavioral signature emerges during the interval in which cortical circuits are refined by experience, and the molecular machinery that opens and closes critical periods — excitatory–inhibitory balance, parvalbumin interneuron maturation, perineuronal net deposition — is disturbed in ASD models and in postmortem tissue (Gogolla et al., 2009).

The proposal has nonetheless retained a specific representational commitment that we believe is now inconsistent with the data. In the critical-period framing, the developmental abnormality is expressed as a modification of a curve: onset may be precocious or delayed, duration increased or decreased, degree of plasticity increased or decreased, or the period may fail to open or to close (LeBlanc and Fagiolini, 2011). Every one of those possibilities is a change in the amplitude or timing of a single quantity. Plasticity is more, or less, or earlier, or later.

The most direct structural evidence from ASD brain tissue does not have this form. Tang et al. (2014) reported increased dendritic spine density in layer V pyramidal neurons of postmortem ASD temporal lobe, arising specifically from reduced developmental pruning, and correlated with mTOR hyperactivation and reduced basal autophagy. In Tsc2 haploinsufficient mice, in which mTOR is constitutively overactive, the same postnatal pruning defect appeared alongside ASD-like social behaviors, and rapamycin corrected both — but not in neuronal autophagy-deficient animals. Critically, they showed that neuronal autophagy enables spine elimination with no effect on spine formation.

A defect that removes elimination while sparing formation is not a smaller amount of plasticity. It is a change in what plasticity does, and a single scalar term for plasticity has no direction to lose.

This is not a claim that developmental plasticity has never been represented as directional. Bidirectional, separately-thresholded synaptic modification has been formalized since the sliding-threshold theory of Bienenstock, Cooper, and Munro (1982), and long-term potentiation and depression, metaplasticity, and homeostatic synaptic scaling (Turrigiano and Nelson, 2004) are all explicitly signed phenomena in the broader plasticity literature. The claim made here is narrower: the critical-period framing specifically applied to ASD has not imported this formalism. As stated by its own proponents, the representational vocabulary of that account is limited to onset, offset, duration, and degree — a scalar family — with no term corresponding to a direction of encoding (LeBlanc and Fagiolini, 2011). We take BCM-type and homeostatic accounts as prior art for the general claim that plasticity can be signed, and argue that the ASD structural data require importing that formalism into a specific literature that has so far modeled the disorder without it. This distinction matters beyond terminology. BCM-type sliding-threshold rules and homeostatic scaling are signed, but the sign is carried by a single, activity-dependent threshold governing potentiation and depression together; moving that threshold changes when synapses potentiate versus depress, not whether the two processes can be independently disabled. Applied to ASD, a pure sliding-threshold or scaling account would predict correlated, threshold-linked changes in formation and elimination as activity history shifts — not the pattern in the Tang and Batista data, where one arm is removed by mTOR dysregulation while the other proceeds normally. Explaining that dissociation requires an asymmetric gating term separate from the threshold itself, which is what v(t) and its mTORC1-linked components supply and what BCM-type accounts alone do not.

This paper develops that argument. We first show, from an independent animal literature in which the encoding window can be opened and closed pharmacologically, that developmental encoding is signed and that its arms are separably controlled. We then argue that ASD is better modeled as a disturbance of the direction of encoding than of its magnitude, derive a partition among phenotypically convergent presentations, and specify what would falsify the proposal.

2. Evidence That Developmental Encoding Is Signed

2.1. The Imprinting Preparation

Filial imprinting in the domestic chick provides an unusual combination of properties: a short and well-delimited encoding window, an identified forebrain substrate in the intermediate medial mesopallium (IMM), a permissive state that can be opened and closed by experimental manipulation, and structural readouts measured within the window (Horn, 2004; Solomonia and McCabe, 2015).

The window is not a chronological endpoint. Imprinting training itself raises brain triiodothyronine (T3), converted from circulating thyroxine by type 2 deiodinase; exogenous or focally delivered T3 restores imprintability in chicks that have passed the normal window, and deiodinase inhibition impairs imprinting (Yamaguchi et al., 2012). GABA-A and GABA-B receptors in IMM play developmentally opposing roles in opening and closing the window (Aoki et al., 2018). Pharmacological AKT activation restores imprinting after normal closure (Batista et al., 2018). The window is a manipulable state, and this is what makes the preparation informative for the present argument: the permissive state can be set independently of the training experience.

2.2. The Structural Record Has Two Arms

The structural consequence of imprinting is not an increase in synaptic quantity. Auditory imprinting is associated with dendritic spine loss in rostral forebrain (Wallhäusser and Scheich, 1987), and NMDA-receptor blockade suppresses learning-induced synaptic elimination (Bock and Braun, 1999). Batista et al. (2018) resolved the composition: in both the mediorostral nidopallium/mesopallium and IMM, imprinting reduced thin spines and increased mushroom spines, and rapamycin blocked the increase in mushroom spines without blocking the decrease in thin spines.

Two features of this result matter here. First, the record of a single learning episode is composed of opposing changes occurring together. Second, the two arms are separably controlled by a single pharmacological manipulation, which means they are not two aspects of one underlying quantity.

A natural alternative reading treats spine maturation as a serial pipeline rather than two coupled arms: thin spines are the immature precursor of mushroom spines, and "thin spines decrease, mushroom spines increase" describes one conversion process rather than two independently controlled ones. On that reading, rapamycin's failure to block thin-spine loss is unsurprising — thin spines still enter the pipeline and are still removed, but fail to consolidate into the mushroom state and are eliminated instead. A single gate positioned at the consolidation step, with mTORC1 controlling that step alone, would suffice; no direction vector is required.

This reading makes a prediction the reported data do not bear out. If thin-spine loss under rapamycin reflected failed consolidation rather than an independently controlled elimination process, blocking consolidation should increase, not merely preserve, the population of thin spines lost without replacement, and the abortive intermediate should appear somewhere — as an increase in stubby spines, or a failure to clear filopodia. Batista et al. (2018) report neither. Thin-spine reduction under rapamycin remained statistically significant in both MNM and IMM, at a magnitude the report treats as equivalent to the vehicle-trained condition rather than amplified by it, and no change in stubby or filopodial counts was detected in any condition. The authors' own interpretation treats formation and elimination as independently regulated, explicitly analogizing the pattern to a form of hippocampal homeostatic plasticity in which potentiation of one synapse population is coordinated with, but mechanistically separable from, elimination elsewhere on the same dendrite (Bourne and Harris, 2011) — a coordination-without-shared-mechanism relationship, not a single conversion pipeline. A one-gate serial account also leaves unexplained why the elimination component is mTORC1-independent while the consolidation component is mTORC1-dependent: if both stages belonged to a single process passing through the same gate, both should show the same pharmacological sensitivity. The dissociation reported is the signature two separably controlled arms predict, not the signature a serial pipeline predicts.

3. Formal Statement

We adopt the minimal formalism required by the argument, developed in full elsewhere (Rahman, 2026a; Rahman et al., 2025). Immediate behavioral execution and durable neural encoding are represented separately. Encoding drive at the neural level is distinguished from the permissive state that allows that drive to produce durable change, so that writability is

Wₙ(t) = Φₙ(t) × Gₙ(t),

where Gₙ is activity-dependent encoding drive and Φₙ is encoding permissiveness — the quantity that the imprinting window manipulations set. Substrate change is then

dS/dt = Wₙ(t) · v(t) − Λ[S − S₀],

where S is a multidimensional substrate state, Λ a consolidation and decay term, and v(t) a direction vector specifying which substrate components are increased, decreased, or redistributed by the current writing episode.

In its minimal form, v(t) is two-dimensional, with one component tracking elimination and one tracking formation, v(t) = [v_elim(t), v_form(t)]ᵀ, so that a disturbance can independently scale, reverse, or decouple the two. This minimal partition is not a formal convenience; it is what the evidence in Section 2 forces. v_form is set by mTORC1-dependent consolidation — blocked by rapamycin in the chick, hyperactivated in Tsc2-haploinsufficient mice — and v_elim is set by a separate, mTORC1-independent process, unaffected by rapamycin in the chick and impaired via autophagy in ASD. A higher-dimensional partition, by dendritic compartment, spine subtype, or cortical layer, is not excluded and may ultimately be required as more separable control points are identified, but the two-component division is the minimal structure that distinguishes a directional-failure account from a scalar one, and it is the one the present data can support without overreach.

The critical-period literature has, in effect, been modeling Φₙ alone. Onset, offset, duration, and degree of plasticity are all properties of Φₙ. The vector v has no counterpart in that framing. The imprinting data show that v exists and is separately manipulable; the ASD data, we argue next, show that it is what is disturbed.

An objection should be met here. One may reply that a scalar account can accommodate directional failure simply by positing separate scalars for formation and elimination. That is correct, and the concession is the argument: once separate parameters per direction are required, the object being modeled is a vector, and the developmental literature has not drawn the consequences of that. Chief among them is that a manipulation which raises overall plasticity and a manipulation which restores the balance between arms are different interventions with different predicted outcomes, and the two have not been distinguished in either the ASD treatment literature or the critical-period reopening literature.

4. Autism as a Directional Failure

4.1. The Structural Evidence

Tang et al. (2014) provide the pivotal observations. In postmortem ASD temporal lobe, layer V pyramidal neurons showed increased spine density attributable to reduced developmental pruning, a density increase independently reported using Golgi impregnation rather than Tang et al.'s methods, and across additional cortical regions, by Hutsler and Zhang (2010). Basal autophagy was reduced across childhood and adolescence: LC3-II, indexing autophagosome abundance, was lower in ASD than control, and p62, an autophagy substrate, was higher. In Tsc2⁺ᐟ⁻ mice, postnatal spine pruning defects, blocked autophagy, and ASD-like social behavior co-occurred; rapamycin corrected ASD-like behaviors and pruning defects in Tsc2⁺ᐟ⁻ mice but not in Atg7 conditional knockouts or double mutants. Neuronal autophagy enabled spine elimination with no effect on spine formation.

Downstream consequences at the network level have since been demonstrated: in Tsc2 haploinsufficient mice, mTOR-dependent increased spine density is accompanied by ASD-like stereotypies and cortico-striatal hyperconnectivity on resting-state fMRI, rescued by mTOR inhibition (Pagani et al., 2021). The directional structural defect therefore has a macroscale functional signature and is not confined to the level of the spine.

4.2. The Mapping, and Its Limit

Placed beside the imprinting result, the correspondence is close but not identical, and the difference should be stated rather than elided. In the chick, mTOR inhibition removed the formation arm and spared elimination. In ASD and its models, mTOR hyperactivation removes the elimination arm and spares formation. The claim licensed by both together is not that mTOR does the same thing in the two systems, but that mTOR signaling sets the balance between the two arms of encoding, and that pushing it in either direction produces a one-sided record.

This is the sense in which ASD is proposed to be a disturbance of v rather than of Φₙ. Encoding is not failing to occur; social and non-social experience continues to be written. What is absent is the elimination component that, in typical development, accompanies and shapes what is retained.

4.3. Why This Matters for the Critical-Period Account

Under the scalar account, therapeutic reasoning proceeds by asking whether plasticity is too low, in which case one reopens the window, or too high, in which case one closes it. Under a signed account, neither is the right question. The relevant intervention restores the balance between arms, and the appropriate readout is direction-specific: an intervention that raised total plasticity while leaving the imbalance intact would be predicted to worsen rather than improve the substrate, because the added writing would again be one-sided.

The reopening literature is instructive on this point. Interventions that restore critical-period plasticity in adulthood — local reduction of inhibition, fluoxetine, lynx1 deletion, enzymatic degradation of perineuronal nets — are characterized in terms of restoring plasticity, and their effects on the balance of formation and elimination are, to our knowledge, not routinely reported (LeBlanc and Fagiolini, 2011; Hensch, 2005). The signed account predicts that this omission conceals heterogeneity in their effects.

5. Three Routes to a Convergent Phenotype

If encoding depends jointly on a permissive state, an encoding drive supplied by experience, and a direction vector, then phenotypically similar presentations may arise from disturbances at different points. Three routes are distinguishable in principle and, we argue, in the existing human literature.

Route	Architecture	Permissive state Φₙ	Input/drive Gₙ	Direction v	Predicted profile
Directional failure	variable	near-typical	present	disturbed	one-sided structural record; persistent
Progressive input loss	intact	open	declines over months	typical	initially typical performance, then divergence
Input absence	intact	intact	absent	typical	execution intact; less persistent under restored input

5.1. Directional Failure: Idiopathic ASD

This is the route developed in Section 4, characterized by the pruning deficit and its mTOR-autophagy correlates.

5.2. Progressive Input Loss: The Prospective Infant Literature

Jones and Klin (2013) followed 59 infants at high familial risk and 51 at low risk, and found that attention to others’ eyes was not diminished in the first months of life in infants later diagnosed with ASD; rather, eye fixation declined from 2 to 6 months of age, a pattern absent in infants who did not develop ASD. The authors explicitly note that this falsifies the hypothesis that eye-looking is immediately reduced.

In the present terms, this is a trajectory in which the encoding gate is presumably open while the drive supplied by social experience progressively falls. Whether this is a separate route or an early consequence of the first is not settled by existing data; the framework treats it as separable and predicts it should be identifiable prospectively (Section 7).

5.3. Input Absence: Deprivation-Related Quasi-Autism

In the English and Romanian Adoptees study, 26 of 101 adoptees exposed to extended severe institutional deprivation displayed autistic characteristics, termed quasi-autism (Rutter et al., 1999; Rodriguez-Perez et al., 2023). The original description reported features that distinguished it from idiopathic autism: an unusual degree of social approach, a relative lack of communication problems, intense circumscribed interests, a weaker association with male sex, and less persistence between ages 4 and 6.

Intact and indeed elevated social approach, in the presence of absent durable relational structure, is what the framework predicts when architecture and gate are intact and input is absent: execution is unimpaired and the substrate is simply unwritten. The reported reduction in persistence is consistent with a substrate that remains partly writable once input is restored.

This route is not hypothetical, and it has been tested directly. Rodriguez-Perez et al. (2025) compared the quasi-autism group against a community sample of early-diagnosed autistic individuals, characterizing the quasi-autism symptom profile at ages 11, 15, and 23 through item-level analysis of the Social Communication Questionnaire against non-deprived UK adoptee controls, and comparing latent structure against the QUEST cohort. We propose that this comparison be read as a test of route separation rather than as a description of two clinical variants.

5.4. An Incidental Alignment: Sex Distribution

The framework predicts that whatever confers male preponderance in idiopathic ASD should attach to the encoding-layer route and not to the input-absence route, because the latter involves no disturbance of the encoding machinery. The reported distributions are consistent: the quasi-autism group was 61.5% female (Rodriguez-Perez et al., 2025), against the substantial male preponderance characteristic of community autism, and the original quasi-autism description already noted a weaker association with male sex. We present this as an alignment rather than as evidence, since sex ratios in small samples are unstable and the comparison was not designed to test it.

6. A Case the Framework Excludes: Maternal Thyroid Status

Because T3 is the best-characterized implementer of the encoding gate in the chick, the association between maternal thyroid status and autism risk invites direct incorporation. Rómán et al. (2013), in the Generation R cohort, assessed thyroid function in 5,100 women at a mean gestational age of 13.4 weeks and reported that severe maternal hypothyroxinemia was associated with an adjusted odds ratio of 3.89 (95% CI 1.83–8.20) for probable autism in offspring.

The framework nonetheless assigns this finding elsewhere, and we state the exclusion explicitly because it is a case where the framework constrains rather than accommodates. The mechanism proposed for the human association is first- and early second-trimester disruption of neuronal migration and cortical layering, with animal models showing blurred and irregular cortical layers, absent barrels, and neurons in abnormal locations including white matter. That is thyroid hormone constructing architecture. The chick encoding gate is perihatch, acts within minutes through non-genomic signaling, and determines whether activation produces durable change in an already-assembled circuit.

The same hormone therefore occupies different positions at different developmental times, and the assignment is made by function and timing rather than by chemical identity. Treating the maternal hypothyroxinemia association as evidence for a thyroid-dependent encoding gate in human ASD would conflate an architectural effect with a permissive one. We note this because the conflation is available and superficially attractive.

7. Predictions

The proposal is testable in five respects, each addressable with existing methods and, in two cases, with existing samples.

First, structural measures in ASD should show dissociation rather than proportional reduction. Indices of spine formation and spine elimination should diverge, with elimination reduced and formation near-typical, rather than both reduced in proportion. The relevant methodology already exists and needs only to be applied to this question: longitudinal two-photon imaging in vivo can track individual spines across days and classify each as newly formed, eliminated, or persistent (Holtmaat et al., 2005; Zuo et al., 2005), rather than inferring turnover from a single cross-sectional density measure as the postmortem ASD literature has done to date. Applying this approach to a mouse model carrying an ASD-associated mTOR-autophagy lesion, with formation and elimination rates reported separately rather than collapsed into a net density change, would constitute a direct test. A finding of proportionate reduction in both arms would favor the scalar account.

Second, the three routes should differ biologically where they converge behaviorally. Deprivation-related quasi-autism should not show the mTOR-autophagy pruning signature that characterizes idiopathic ASD, despite phenotypic overlap. This is the single most decisive available test and it requires no new behavioral data — only tissue or in vivo markers applied across groups already characterized.

Third, sex preponderance should track route rather than phenotype. In samples large enough to estimate it stably, the male bias should be present in the directional-failure route and attenuated or absent in the input-absence route.

Fourth, interventions should be evaluated on direction-specific readouts. An mTOR-normalizing intervention should shift the ratio of elimination to formation rather than raise a global plasticity index, and its behavioral benefit should track the ratio rather than the total.

Fifth, in prospective infant-sibling cohorts, infants whose social attention declines from an initially typical level should be separable from infants whose social attention is low from the outset. The framework assigns these to different routes — progressive input loss versus architectural or early directional disturbance — and predicts different structural correlates and different responses to input-restoring intervention.

8. Limitations

The central structural evidence rests on a small postmortem series. The seizure confound is narrower than it might first appear: Tang et al. (2014) report that spine density itself was not correlated with seizure history, postmortem interval, cause of death, brain pH, or tissue storage (their Table S2); the correlation with seizure activity is specific to the autophagy markers, LC3-II and p62, not to the structural finding those markers are invoked to explain. That leaves the structural claim — increased spine density from reduced pruning — less exposed to the confound than the proposed autophagy mechanism connecting it to mTOR signaling, which remains a substantial concern in a population with elevated epilepsy comorbidity and has not to our knowledge been resolved by replication in seizure-free ASD tissue. This concern is only partly offset by Hutsler and Zhang (2010), who independently reported increased spine density in ASD cortex — including layer V of temporal cortex, the same lamina and region implicated by Tang et al. (2014) — using Golgi impregnation in a separate postmortem series. That is a replication of the structural finding, increased density from reduced pruning, not of the mTOR-autophagy mechanism or the autophagy-dependent rescue. We were unable to determine from their published methods whether Hutsler and Zhang's series reports seizure status at all, so it cannot be treated as either confirming or ruling out the confound for the structural finding, and it does not bear on the confound affecting the autophagy markers specifically.

The quasi-autism comparison rests on 26 individuals, and the sex-distribution alignment in Section 5.4 is correspondingly fragile.

The inference from imprinting to human ASD is architectural rather than homological. We do not claim that the human cortex and the chick IMM implement the same mechanism; we claim that the imprinting preparation demonstrates that developmental encoding can be signed and separably controlled, and that this possibility must therefore be represented in accounts of human developmental disorder. The demonstration establishes that the vector exists somewhere, not that it is the same vector.

The direction vector v is a latent construct, inferred from the separable manipulability of formation and elimination rather than measured directly. No human study has yet measured formation and elimination indices separately in a social-encoding context, which is why the first prediction in Section 7 is the foundational one.

Finally, the proposal is descriptive at the level that matters most for etiology. It states that the direction of encoding is disturbed in ASD; it does not explain why. The genetic heterogeneity of ASD is not addressed, and it is entirely possible that convergence on a directional failure is the common downstream consequence of many upstream causes rather than a mechanism in its own right.

Interventional caution is also warranted. Attempts to correct developmental gate variables in ASD have not to date produced durable clinical benefit: the largest randomized trial of an mTORC1 inhibitor in a TSC-associated neurodevelopmental population, a 12-month placebo-controlled trial of everolimus in 32 children, found no benefit on full-scale IQ or on secondary autism and behavioral measures (Overwater et al., 2019). That trial targeted mTORC1 activity broadly rather than the ratio of elimination to formation specifically, so it is not a direct test of the present proposal, but it is the closest existing test of gate-level intervention in this population, and its null result should temper expectations for the direction-targeting strategy proposed here.

9. Conclusions

Autism has been described as a disorder of critical periods, and the description has organized a productive literature. But the representation of plasticity that the description carries — a quantity that may be raised, lowered, advanced, or delayed — cannot express the structural finding that is best established in ASD neuropathology. Reduced pruning with spared formation is not a smaller amount of plasticity. It is a one-sided record.

Developmental encoding, where it has been manipulated within a defined window in an animal preparation, turns out to be composed of opposing changes whose arms can be dissociated by a single pharmacological intervention. The formal object implied by that finding is a direction, not a magnitude. We suggest that autism is a disturbance of that direction, that phenotypically convergent presentations may be partitioned by which term is disturbed, and that the most informative next measurement is not of how much plasticity remains but of what it is doing.