The highest progression of a spectrum of non-epidemic disorders is observed in countries where chemical substances are consumed the most, including perfumes, detergents and pesticides . With forty times more cases today than in the 60s, and ratio estimates that shifted from 1.7% (in 2010) to 2.4% (weighted prevalence ) of the US population, these neurodevelopmental disabilities referred to as Autism Spectrum Disorders (ASD) seemingly followed environmental changes. Indeed, among 250 susceptibility genes already analyzed by geneticists, none is involved in more than 2% of ASD cases . However, more than autistic twins who share both genetic and prenatal environmental influences , this is the male prevalence (greater than 4:1) that led scientists to suspect a heritable genetic disease implying DNA mutations . However, there is increasing evidence that variations in genes expression caused by the environment can be inherited in humans, with a number of studies reporting transgenerational transmission of these variations without genetic mutations (epigenetics) .
Neuromodulators possibly unbalanced by epigenetic regulations
Monoamines under influence
Mostly focused on synapses and neurons, genetic investigations have concerned the controlled networking of neurons by neuromodulators, including serotonin (5-HT) for its recognized contribution to the brain development . Several studies of autistic troubles have therefore addressed the production, transport and metabolism of 5-HT, depending themselves upon the expression of specific genes. In the blood of autistic patients, high 5-HT levels can be attributed to its increased synthesis by tryptophan hydroxylase, enhanced uptake into platelets through transporter, or decreased metabolism by the monoamine oxidase type A (MAOA) enzyme . MAOA preferably metabolizes 5-HT among other monoamines, namely Norepinephrine (NE) and Dopamine (DA). Given its likely coupling with NE, serotonin should be considered in its neuromodulation context. Accordingly, 5-HT and NE variations would tend to accompany each other, except in case of repeated use of drugs of abuse . MAOB, the other type of monoamine oxidase, may also be involved in ASD, since its brain concentrations, barely detectable at birth, reach a high at around 2 years of age , when autism symptoms start being revealed.
In order to avoid synaptic overload, neuromodulators are broken down by enzymes once they have contributed to transmit the nervous signal. The balanced concentration of monoamines in the synaptic cleft thus partly depends on the ongoing enzymatic activity. Compared with control groups, significant reductions of the MAOA enzyme have been recorded in autistic children, either in plasma  or in both cerebellum and frontal cortex , inducing lower metabolism (degradation, deactivation) of serotonin. The key-role of MAOA is also evidenced in genetically-modified mice with MAO knock-out (KO) gene, displaying autistic features associated with brain abnormalities . Similarly, mutations and deletion in the MAOA gene, as found in the Brunner Syndrome (BS) , have recently been observed to induce autism symptoms . Essential here as a potential cause of neuromodulation imbalance, the following piece of knowledge must be emphasized: unlike other monoamines, 5-HT cannot be metabolized by another enzyme, namely the catechol-O-methyltransferase (COMT) .
Variations in the expression of the MAOA enzyme have been identified which do not involve DNA mutations. Located in the promoter region of genes, allelic variants are known as functional polymorphisms, namely variable number tandem repeat (uVNTR): the 3-repeat allele is associated with decreased transcriptional activity, whereas the 4-repeat allele shows the opposite effect . Several genes in charge of neuromodulators show allelic variants, including the 5-HT transporter  and the COMT enzyme: the A (or Met) and G (or Val) alleles respectively exhibit low and high genetic expressions for breaking down the synaptic DA and NE. About 25% of people carry the AA (Met/Met) polymorphism which brings the lowest metabolism of both DA and NE . Besides a possible synergetic effect with other genetic variants , the 3-repeat low-activity allele of the MAOA promoter has been linked with cortical enlargement (up to a factor 1.5)—a recognized hallmark of autism —, as well as with ASD severity, namely a tendency towards lower intellectual abilities and more severe behavioral problems . Another specific MAO polymorphism (rs6323) with low-activity allele exhibits significant association with ASD, posing higher risk in males . However, despite their involvement in the ASD heterogeneity, these genetic polymorphisms exist independently of the condition, and should not be regarded as its main cause.
The MAOA promoter not only displays built-in variants; it is also sensitive to environmental factors which participate in its epigenetic programming. A number of situations have been shown to modify the MAOA expression, including Major Depressive Disorder (MDD) , stressful events, diet changes, tobacco smoking, and social environment . At the molecular level, mechanisms have been identified which explain how a xenobiotic—foreign chemical substance—can contribute to the acute promotion of MAOA. For instance, Valproic acid (VPA) can directly bind to a transcription factor of the gene promoter, among other possibilities of action . Quite relevant to the present study, the pre-natal exposure to xenobiotics with affinity for endogenous transcription factors is likely to induce persistent regulations of genes involved in the enzymatic activity . With respect to interactions between maternal and fetal neuromodulators, exposure to maternal inflammation or MDD imply atypical concentrations of placental 5-HT. This situation may also affect the fetal brain development [28, 29]. Accordingly, both external and internal factors can be involved in the embryonic development of ASD .
Based on the above facts and hypotheses, it is now time to address a main point of the present study. Mutations and promoter variants of the MAOA gene can undoubtedly be inherited. As an alternative, the promoter region of this gene may undergo prenatal acute modifications caused by either xenobiotic intrusion or maternal condition. Among potential xenobiotics, VPA has been found to stimulate the production of MAOA and, consequently, the degradation of 5-HT . Unexpectedly, if taken during pregnancy, this small molecule is known to possibly generate autistic features in the offspring , which have been linked with high 5-HT synaptic levels, instead. This apparent contradiction can be resolved by discriminating two separate contexts in which VPA may reach the brain. The first one is during sensitive periods of the prenatal development, when genetic programming can still counteract a temporary acute imbalance of monoamines (MAOA+ in Fig. 1) . As a small molecule with two methyl groups (CH3), VPA can cross the fetus blood-brain barrier and bind to DNA transcription factors (e.g.: Sp1 ) together with endogenous proteins. Among several epigenetic processes that are able to fight back against the resulting overexpression of MAOA, protein R1 is known to act through Sp1 sites as well . Now, regarding the postnatal context, monoamine metabolisms may then have been permanently programmed through reactions to an initial contact with VPA. Although still inducing MAOA, VPA would not trigger the same epigenetic regulation as in the prenatal context, namely within critical periods of the fetal brain development. The scenario proposed in Fig. 1 shows how a coincident under-metabolism of 5-HT and over-metabolism of DA can result from the meeting of epigenetic marks with an allelic variant. Following a temporary xenobiotic intrusion (e.g.: VPA), a MAOA down-regulation allows monoamines to return to their baseline value. But when the temporary/accidental increase of MAOA is over, monoamines are less metabolized again, hence a significant increase of their average level. Both DA and NE under-metabolisms can then be compensated by another enzyme (COMT) before MAOB gradually takes part in DA degradation, after birth. This scenario could explain the relative excess of synaptic 5-HT observed long after the prenatal intake of a drug which paradoxically tends to clear the synapse from 5-HT, among other monoamines.
Now, there is a connection waiting to be made between these early molecular events and the late onset of autistic conditions. If a theory could bridge the gap between a prenatal molecular accident and postnatal behavioral deficiencies, it should also account for the delayed occurrence of symptoms, several years after birth in children who manifest developmental regression . Among relevant factors to be specified hereafter, low MAOB levels during gestation could mask the uncoupling of NE and 5-HT metabolisms which is assumed here to underlie ASD (Fig. 1).
Computer simulations of impaired encoding
Reactions of the nervous system against the poor degradation of 5-HT at issue may include the reduction of serotonin receptors evidenced in limbic and neocortical regions of the autistic brain . Along the same line, anatomical studies reveal reciprocal connections between the 5-HT generator (raphe nuclei) and its cortical targets, allowing a feedback control of cortical 5-HT release . Exercised ‘online’ in the awake brain, this control loop is likely to compensate for excess 5-HT, together with the reuptake mechanism which is known to contribute to 5-HT synaptic clearance. However, an enzymatic deficit cannot be counteracted when the release of 5-HT is interrupted whatever the cortical feedback, as across sleep cycles: there, 5-HT neurons decrease their firing and stop activity during the last phase of each cycle . Consequently, if this interruption fulfilled a key function during sleep, the faulty persistence of monoamines in the synaptic cleft could not be counteracted by ‘online’ mechanisms. Such irrelevant synaptic accumulation of serotonin during sleep cycles is central in the theory proposed here.
Sleep and parallel processing
From a broader perspective, the medical descriptions of ASD have evolved over last decades. After having enhanced social deficits, they shifted towards communication problems and repetitive/restricted behaviors . Today, sleep problems are also increasingly reported. Whereas healthy adults sleep across 4 or 5 cycles of distinct phases, autistic children are observed to hardly fall asleep, sleep a few, with delayed phases and significantly less ‘deep-sleep’. Although identified in animal models as well , these specific troubles are surprisingly not mentioned among the official criteria of autism : sleep disorders are rather thought to originate in the daytime over-activity and stress , only worsening the condition disabilities .
The lifelong alternation of wake and sleep involves memory processes that are respectively referred to as encoding, consolidation and recall. Sleep is widely thought to contribute to the consolidation stage by strengthening fragile memories , then available on the long-term for effective recall. Such consolidation during sleep may accompany the transfer of information between the hippocampus and cortical areas , and sleep phases can respectively be associated with distinct types of memory . Contrary to waking in which brain areas are busy interacting, sleep would allow the independent consolidation of memories in different neural structures . During Slow-Wave-Sleep (SWS), the hippocampus is evidenced to form memories of events in time , and the ‘two-stage memory’ system postulates a role of short-term recipient for information to be distributed in cortical areas . In the final Rapid Eye Movement (REM) phase of a sleep cycle, the brain paradoxically shows greater activity than during wake, whereas the body remains still. Across these distinct phases, a saw-teeth neuromodulation pattern makes the production of both 5-HT and NE monoamines decline down to zero from the beginning of each cycle until REM sleep , while DA levels are maintained.
Whatever the model, a division of labor is currently supported whereby the part of sleep would pertain to the strengthening of memories encoded during prior wakefulness [42,43,44,45,46,47]. Although few structural modifications of the neural tissue have been observed during sleep so far (i.e.: the expansion of synaptic boutons in drosophila ), it is argued here that offline periods could also take part in memory encoding. Indeed, with evidence of shortly delayed activities in distant neural networks (e.g.: hippocampus and neocortex [43, 50]) during sleep phases, it appears feasible for brain areas to encode inner stimuli while the latter are generated by the hippocampus. Sleep would thus contribute significantly to the proper development of the brain, a function vital enough to justify that one third of the human life is spent asleep, albeit in a competitive environment.
Genesis of a learning constraint
When initiated in 1983, the Guided Propagation (GP) system implemented concept-cells named Event-Detectors (EDs), considered as potential targets for inner-flows of activity. The other main feature of this deterministic approach was the recruitment of new cells for growing the memory paths ending in EDs, in response to requests of a dynamic learning algorithm . In this formalism, the direction/path taken by a given inner-flow depends on the occurrence of afferent stimuli, one after the other, hence the expression: “guided propagation”. Within module k, the inner-flow is initiated by a root-cell, and flows along the parallel paths chains of GP-cells (formerly named Elementary Processing Units) coding for sequences of events (e.g.: [root-cellk0] => [cellk1:‘M’] => [cellk2:‘A’] => [cellk3:‘O’] => [cellk4:‘B’] => concept-cellk5: ‘MAOB’). In order to control the necessary balance between inner (contextual) flows and stimuli, only two control parameters are associated with every cell, namely the ratio R between the respective weights of Context (C) and Stimulus (S) inputs, and Excitability E which defines the cell response threshold . Surprisingly, this minimalist design could later on be used to model coupling between 5-HT and NE (through R), as well as a hypothetical global function of DA neuromodulation (through E) . Definitions of the GP system components are given in Table 1, while a video movie shows a glimpse of running computer simulation .
GP networks could hardly be considered as neurobiological models at the time they were first issued. Altogether, concept-cells in a self-growing network did not meet the notion of “distributed memory” in fashion at that time, neither the dogma stating a fixed number of neurons at birth. Since then, concept-cells have regained interest with the discovery of neurons that respond selectively to either high-level notion  or action . As well, the genesis of neurons across a lifetime became plausible with the relation found between song learning and the production of new neurons in a vocal nucleus of the adult canary brain . More recently, a strong adult neurogenesis has even been evidenced in the human hippocampus . Provided environmental enrichment, experimental results in adult guinea pigs further suggested that new neurons are generated in the cortex of mammals . In a functional perspective, newborn neurons ensure the formation of non-overlapping representations for distinct patterns, whereas older cells mediate pattern completion , a biological hypothesis consistent with the Differentiation/Generalization alternative implemented in GP-networks (Fig. 10 of the “Methods” section).
The first implementation of GP-based model concerned reading , and enabled an additional functionality. Memory paths built for perceiving words were found to possibly produce them, through offsets of propagation thresholds. The expression “guided propagation” found a new meaning then: whereas a given series of stimuli could guide an inner-flow forwardly towards a specific ED, modulation signals targeting this ED could aspirate the upstream inner-flow and “generate” the same sequence of stimuli. It thus appeared that an ED could be used as Event Generator (EG). This dual function evokes mirror neurons, firing either for the perception of someone acting or for the production of the same action . Action production with GP could later on be applied to the modelling of four symptoms of the Parkinson disease . This study originated from the assumption that a local lack of dopamine could functionally compare to a GP-cell high Excitability. Not only this parameter, but also the dynamic weighting of GP flows could be linked with the reciprocal control, or coupling, between noradrenergic and serotonergic neurons . Based on the same simplified representation of neuromodulation, a computer implementation of decision-making has more recently been designed, in line with the ascending spiral neurobiological model . For this purpose, parallel GP-channels can be connected through cross-circuits aimed at propagating modulating signals from “emotional” to “conditioner” and eventually “sensorimotor” channels where decisions are thus quickly facilitated/repressed [65, 66]. But this anticipation skill brought an inescapable constraint of the original learning algorithm to the surface.
The GP ability to encode stimuli sequences actually relies on their synchrony with the host-module inner-flow. As long as these two flows run in pace, a “never-ending learning” algorithm can apply to various tasks ; but the picture changes when unconditioned stimuli have to be anticipated for making decisions, implying that the internal representations of future possible events are activated in advance. By boosting their inner-flows, the forward-looking emotional/conditioner channels can run ahead of the current input. In the context of lifelong learning, the challenge is thus to deal with the implementation of two skills, anticipation and encoding, which share the same memory space but require incompatible regulations (different values of the same parameters). If a learning episode were triggered during anticipation, even known patterns would hardly match their internal references. This lack of coincidence being the criteria for the system to upgrade its knowledge, the network would tend to grow too much. Furthermore, the system proactivity would forbid the accurate sprouting of new memory-paths for implementing Differentiation, while favoring the over-Generalization of existing paths by several stimuli. A solution to this problem was inspired by the waking/sleep alternation. Proactivity can actually be shut down across ‘offline’ periods during which memory is disconnected from external stimuli, making anticipation temporarily useless. A new channel (named Hippo in the following) is required for the online “recording” of online events to be replayed offline and possibly encoded in relevant parallel channels, including their cross-connections (Fig. 2). Of note, these Hippo replays appear especially useful for the long-term encoding of behaviors that cannot be rehearsed online. On the biological side, while the overnight reinforcement of neural structures still remains a controversial issue , the stronger assumption of a possible involvement of sleep in neural encoding had seemingly not yet been made, so far.
Offline cycles of memory encoding: a new model
As previously stated, sleep disorders are commonly believed to originate in the daytime over-activity and stress generated by the autistic condition. The opposite causal relationship is supported here. Provided a new interpretation of the sleep architecture (Fig. 3), impaired sleep would disturb the completion of connections within and between parallel neural structures. The resulting miscoded internal representations, as well as their cross-links, would impede several functions relating Perception, Emotion and Language. In fact, these functions are more or less impacted across the autism spectrum. In regular perception, the eyes saccades rely on peripheral vision for the central/foveal vision to run properly . Similarly in everyday decision-making, sensorimotor circuits can instantaneously be driven by emotional channels through an ascending spiral pathway , while language processing requires at least semantic and syntactic representations to interact. For these various skills, the GP parallel architecture suggests that approximate-and-fast channels guide quick decisions made by accurate-and-slow channels. In autism, neural counterparts of GP processing channels and their cross-links may not properly develop. Indeed, the more severe the injury, the worse an autistic individual performs at tasks requiring parallel and interacting processes, such as decision-making based on emotional conditioning, visual tracking, and the production of speech enriched by prosody.
The decrease of R (w5ht/wne) displayed in both Figs. 3 and 4 represents 5-HT/NE during a sleep cycle. The regular timing of offline phases within each cycle results from the maximum value of R at the beginning of each cycle: lowest in sensorimotor modules, medium in forward-looking modules, and highest in cross-circuits. This timing is disturbed when R does not decrease fast enough, which is a way to model a MAOA deficiency entailing an insufficient degradation of 5-HT.