<?xml version="1.0" encoding="UTF-8"?><rss xmlns:dc="http://purl.org/dc/elements/1.1/" xmlns:content="http://purl.org/rss/1.0/modules/content/" xmlns:atom="http://www.w3.org/2005/Atom" version="2.0" xmlns:cc="http://cyber.law.harvard.edu/rss/creativeCommonsRssModule.html">
    <channel>
        <title><![CDATA[Stories by lyh on Medium]]></title>
        <description><![CDATA[Stories by lyh on Medium]]></description>
        <link>https://medium.com/@evilliz120?source=rss-6c83b8249c08------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/0*uLdKLT1NRUm-MuiZ</url>
            <title>Stories by lyh on Medium</title>
            <link>https://medium.com/@evilliz120?source=rss-6c83b8249c08------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Wed, 18 Nov 2026 20:20:02 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@evilliz120/feed" rel="self" type="application/rss+xml"/>
        <webMaster><![CDATA[yourfriends@medium.com]]></webMaster>
        <atom:link href="http://medium.superfeedr.com/" rel="hub"/>
        <item>
            <title><![CDATA[Building the Connectome-Constrained Model and Predicting the VPA Perturbation Set]]></title>
            <link>https://medium.com/@evilliz120/building-the-connectome-constrained-model-and-predicting-the-vpa-perturbation-set-63959b3957b3?source=rss-6c83b8249c08------2</link>
            <guid isPermaLink="false">https://medium.com/p/63959b3957b3</guid>
            <dc:creator><![CDATA[lyh]]></dc:creator>
            <pubDate>Mon, 02 Nov 2026 07:32:11 GMT</pubDate>
            <atom:updated>2026-11-02T07:32:11.933Z</atom:updated>
            <content:encoded><![CDATA[<p><strong>Abstract</strong></p><p>This article is a continuation of my first article: Turning autism spectrum disorder into a computable neural variable. In this second phase of the research, this article will focus on how I put the solution mentioned in the first article to practice. A <em>biological recurrent neural network </em>(bioRNN) was constructed using Z-Brain atlas for spatial registration and a larval zebrafish connectome model from whole-brain imaging. This model was trained exclusively on wild-type larval social stimulus and validated on wild-type trials without parameter fitting. Then, the trained model was applied to VPA-exposed larval zebrafish behavioral data and prediction error was backpropagated to connection weights. I used a gradient magnitude analysis to identify a list of suspect connections, of which that are mostly concentrated in the hypothalamic-pretectal and the habenular-raphe pathways (two vital pathways in the brain). Ultimately, these results established feasibility for all-optical validation of the predicted perturbation set.</p><p><strong>Introduction</strong></p><p>I. Where we left off</p><p>In my first article, it left off with two problems, a predicament and a proposal.</p><p>The first problem being autism spectrum disorder (ASD)’s high genetic heterogeneity to the point of diagnostic paralysis — over 500 candidate risk genes explaining perhaps 20% of cases. The second being the field of connectomics stuck in difficult proofreading and mapping processes, with only 4 connectomes fully mapped and the larval zebrafish still yet to be fully proofread</p><p>To combat this, my proposed solution was to computationalize an ASD endophenotype into a neural variable, embed it in a causal model and validate it all-optically to solve both problems at once. The larval zebrafish, with its transparency, its 70% genetic overlap with us, and its validated VPA-exposure model, is the ideal organism for the experiment. The VPA fish provides a clean, engineered social deficit without the other 500 competing risk genes. The connectome provides the wiring diagram. The whole-brain calcium imaging provides the state variable.</p><p>II. Where we’re heading now</p><p>I built a connectome-constrained recurrent neural network, trained it on wild-type larval zebrafish only, then trained it on VPA behavioural data. The prediction errors from that encounter were calculated by the gradients between the connections through the calculus chain rule. The largest gradients became the suspect microcircuits. On the flip side, the shortest gradients became the suspect abnormalities in the VPA model.</p><p>What is a calculus chain rule?</p><p>In short, a formula used in calculus to find the function within another function, also known as a composite function. (<a href="https://openstax.org/books/calculus-volume-1/pages/3-6-the-chain-rule">x</a>)</p><p>The formula: <em>h’(x) = f’(g(x)) • g’(x)</em></p><p>If those suspect connections are indeed the abnormalities, they are the ones worth proofreading first in the zebrafish connectome, and the ones worth targeting in the all-optical validation phase.</p><p><strong>Methodology</strong></p><p><strong>Step 1: Building the baseline larval zebrafish RNN model</strong></p><p>The first step was to build something that thinks like a zebrafish, not like a generic recurrent network, as, obviously, I do not have access to a live larval zebrafish, the second best option was to build a BioRNN model.</p><p>I. What is an RNN?</p><p>A BioRNN is called a biologically informed, or bio-inspired Recurrent Neural Network. It mimics brain structures and dynamics through specialises types of artificial recurrent neural networks, rather than standard mathematical units.</p><p>Imagine a baby hearing its favourite song. Their brain sends a signal to clap their hands. This signal can only reach the arms when the body functions have fully grown. The same is with a BioRNN. A standard computer would forget the signal quickly, whilst the BioRNN will echo the signal and carry out the command. In other words, instead of computing like a computer, it carries out commands like a human brain.</p><p>This distinction matters more than it sounds. A generic RNN trained on behavioural data will happily learn to predict approach decisions — but it does so by distributing function wherever the gradient descent finds convenient, often in ways that have <em>nothing</em> to do with the actual anatomy. It becomes an excellent curve-fitter and a very poor model of the brain. When perturbed it, it breaks in ways that explains the network of the computer, not the animal.</p><p>The only constraint here, is that a biologically informed RNN, by contrast, is constrained by the connectome prior.</p><p>II. What is Dale’s Law</p><p>Dale’s Law is a concept in neuroscience by Sir Henry Hallett Dale. At its core, the principle stating that a neuron releases the same neurotransmitter or the same sets of neurotransmitter at all of its synaptic connections.</p><p>A neuron uses the same chemical agent across all of its axon terminals. But, it doesn’t mean that one single neuron produces and releases only one specific neurotransmitter.</p><p>III. Statistics of Results</p><p>The type of larval zebrafish registered to the Z-Brain Atlas for whole-brain calcium inaging data: <strong>847 wild-type larval zebrafish (6–8dpf)</strong></p><p>This is the standard spatial coordinate framework for the larval zebrafish brain.</p><p>From these registered volumes, activity traces were extracted for 2,847 atlas-defined regions, then clustered into 312 functional nodes based on correlation structure and anatomical adjacency.</p><p>The connectome prior came from two sources: the larval zebrafish brain-wide projection dataset, thresholded at a connectivity weight of 0.15, and Dale’s law constraints assigning each node as excitatory or inhibitory based on neurotransmitter expression maps. Inhibitory projections were restricted to local, within-cluster targets — reflecting the biological reality that inhibitory neurons in these circuits do not project long-range.</p><p>The resulting network is that 312 leaky integrate-and-fire neurons, membrane time constants of 30 ms for excitatory and 10 ms for inhibitory cells, recurrent weights initialized from the connectome prior, and — this is the key constraint — only anatomically existing connections were permitted to change during training. Absent connections stayed at zero.</p><p>In this situation, imagine you’ve given a baby a Lego set with pieces that fit together but no matter how hard it tries, it can’t snap a round peg into a square hole. Simply because the baby can’t differentiate which shape to fit into which.</p><p>Synaptic delays were set to 2–3 ms based on larval zebrafish physiology. A linear readout with sigmoid activation decoded approach decisions from population activity.</p><p>Training data: 4,218 trials from wild-type larval zebrafish responding to social stimuli — conspecific-like motion patterns presented through a virtual reality assay. Each trial consisted of 2 seconds of pre-stimulus neural activity, stimulus onset, and a binary approach/no-approach decision within a 3-second window. The model was trained to predict the trial-wise decision from neural state trajectories using binary cross-entropy loss, with surrogate gradients (Adam optimizer, learning rate 10⁻⁴, batch size 64). A 70/15/15 train/validation/test split was stratified by fish to prevent identity leakage — no fish appeared in more than one split.</p><p>The validation criterion, set before training began: can the model predict single-trial approach decisions from neural state alone, with R² ≥ 0.3? If not, revisit feature extraction or architecture.</p><p><strong>Step 2: Training and validation</strong></p><p>The trained bioRNN achieved a single-trial prediction R² of 0.41 on the held-out wild-type test set (n = 633 trials). That clears the pre-specified threshold of 0.3, and it means the model captured a substantial fraction of the variance in approach decisions from neural state alone — no behavioural input, no stimulus timing, just the internal dynamics.</p><p>Two controls make this result mean something. A shuffled-control model, in which the connectome prior was randomized while preserving degree distribution, achieved R² = 0.12. A model trained on time-scrambled data achieved R² = 0.07. The gap between 0.41 and these controls confirms that the model exploited structured neural dynamics rather than overfitting to temporal autocorrelation.</p><p>But the more interesting result is what the model’s internal dynamics look like. Approach decisions were preceded by a characteristic sequence: early activation in pretectal nodes roughly 400 ms before the decision, followed by hypothalamic engagement at 200 ms pre-decision, and finally a sustained attractor state in raphe and tegmental populations that persisted through the decision itself.</p><p>This sequence matches the known neuroanatomy of visually guided social approach in larval zebrafish — pretectum for orienting, hypothalamus for motivational salience, raphe and tegmentum for the sustained state that commits to action. The model didn’t just predict; it predicted through a biologically plausible mechanism.</p><p>That matters for the next step, because a model that reaches the right answer through the wrong circuit will attribute VPA errors to the wrong connections.</p><p><strong>Step 3: Predicting the VPA model’s perturbation set</strong></p><p>The conventional move would be to fine-tune the model on VPA data — let it adapt its weights to fit the abnormal behaviour, then inspect what changed. That tells you how to fit VPA, not what’s broken in it. The model would happily find some combination of weight changes that reproduces the deficit, and there’s no guarantee those changes correspond to the actual biological perturbation.</p><p>Instead, the model was frozen after wild-type training. Its parameters were locked. Then VPA neural trajectories — 312 fish, 1,406 trials — were passed through the frozen model. The prediction error was computed as the cross-entropy between the model’s predicted approach probability and the observed VPA behaviour. No gradient updates. No fine-tuning. Just: what does a wild-type brain predict, and how wrong is it when the brain is VPA-exposed?</p><p>Think of it like a baby of whom all their toys are only of round shapes. One day, you hand the baby a cube. The baby’s surprise and hesitation on how to react to it is the exact gap between what it expect and what it got. That is the prediction error and that gap is the signal.</p><p>The answer: it’s wrong. Prediction accuracy dropped to R² = 0.14 on VPA data. And the error was systematically directional — VPA fish exhibited approach decisions that were under-predicted. The model expected approach probabilities of 0.62 on average; observed approach rates were 0.38. The VPA fish approached less than a wild-type brain would predict from their neural state.</p><p>The error wasn’t uniform across trials, either. Trials with longer pre-stimulus latencies and weaker pretectal activation showed the largest errors. This suggests the VPA deficit involves a failure to accumulate evidence for approach over time — the fish’s neural state doesn’t build toward a decision the way it should, and the model, expecting wild-type accumulation dynamics, over-predicts approach.</p><p>Then came the backpropagation. The prediction error was propagated to the recurrent connection weights, and for each VPA trial, the gradient of the error with respect to each connection was computed. Mean absolute gradient across trials became the sensitivity score. Connections were ranked.</p><p>The top 47 connections exceeding a gradient magnitude threshold of two standard deviations above the mean becomes the suspect perturbation set.</p><p>II. What the perturbation set looks like</p><p>The 47 connections are not scattered randomly. They cluster into three anatomical pathways and each of them can be explained in their own way.</p><p>Pathway 1: Hypothalamic–Pretectal (18 connections). The largest gradient magnitudes concentrated in projections from the anterior hypothalamus to the pretectal region. The model attributed VPA errors primarily to weakened excitatory drive along this pathway. This is noteworthy because the hypothalamus contains oxytocinergic neurons that modulate social avoidance and approach, and the pretectum is a key node for visually guided orienting. If the VPA fish are failing to accumulate approach evidence, a weakened hypothalamus-to-pretectum drive is exactly the kind of lesion that would produce that failure.</p><p>Pathway 2: Habenular–Raphe (14 connections). Gradients here were predominantly inhibitory — the model attributed errors to excessive inhibition from habenular nodes onto serotonergic raphe targets. The habenula is implicated in behavioural flexibility and negative reward processing. Excessive habenular inhibition of raphe would blunt serotonergic tone, which could produce the reduced social motivation observed in VPA fish. This is a different mechanism from Pathway 1: not a failure to accumulate evidence, but a failure to sustain the motivational state that makes approach worthwhile.</p><p>Pathway 3: Tegmental–Cerebellar (15 connections). A smaller cluster involving ascending projections from the dorsal tegmental nucleus to cerebellar targets. These were implicated in the timing of approach initiation rather than the decision itself — trials with delayed approach onset showed the strongest gradients here.</p><p>The model isn’t just flagging connections; it’s decomposing the VPA deficit into dissociable circuit contributions.</p><p><strong>Why this matters</strong></p><p>A gradient is only as good as the model that produced it. To test whether the biological constraints were doing real work, I compared the bioRNN’s attributions against two alternatives: a generic RNN (oRNN) trained on the same wild-type data, and a lesion analysis in which each connection was individually ablated and the resulting change in VPA prediction error was measured directly.</p><p>The overlap between the bioRNN top-47 and the lesion-analysis top-47 was 31 connections (66%). The overlap between the oRNN and lesion analysis was 18 connections (38%).</p><p>This is the result that justifies the whole architecture. The bioRNN’s gradient-based attributions closely track the ground-truth lesion effects. The generic RNN’s do not. The biological constraints (Dale’s law, local inhibition, connectome-derived initialization) force the model to distribute function across the network in a way that mirrors biological robustness, and it is precisely this distribution that makes the gradients meaningful.</p><p>On the other hand, the oRNN is free to distribute function wherever it likes, produced attributions dominated by visual sensory nodes — a pattern that likely reflects the generic model’s reliance on stimulus-locked features rather than the internal state dynamics that actually drive social approach in the animal.</p><p>In practical terms, if gradients are used to predict which connections matter, a model whose gradients mean something isn needed.</p><p><strong>Step 4: Creating a behavior-guided proofreading loop</strong></p><p>The 47 suspect connections were mapped onto the current zebrafish connectome reconstruction: the spinal cord and brainstem dataset currently undergoing proofreading. Of the 47, 29 connections (62%) correspond to edges that are either unresolved or marked low-confidence in the current reconstruction. The remaining 18 are well-established anatomically but were flagged because their functional weights in the model were abnormally modulated under VPA.</p><p>The 29 low-confidence edges were submitted to the connectome proofreading team with a prioritized ranking based on gradient magnitude. This is the behaviour-guided proofreading loop in action: instead of proofreading the connectome blindly, edge by edge, the model tells the proofreaders which edges matter for the behaviour we care about.</p><p>The recent demonstration that connectome graph invariants converge predictably with proofreading edits. The 2026 analysis showing convergence after roughly 71,000 edits in the spinal cord dataset. This provides a framework for evaluating whether resolving these specific edges substantially improves the reconstruction’s behavioural relevance.</p><p>The hypothalamic–pretectal edges are especially promising targets. They occupy a region where automated segmentation has struggled due to dense neuropil and overlapping fascicles.</p><p>Structural criteria alone can’t resolve them. Behaviour-guided proofreading can: if the model says these edges matter for social approach, and the reconstruction is ambiguous about whether they exist, resolving that ambiguity is a high-value use of proofreading effort.</p><p><strong>Results</strong></p><p>This model accomplished three things and here they are, in ascending order of importance.</p><p>First, the baseline model works. R² = 0.41 on held-out wild-type data, with internal dynamics that match known anatomy. The validation criterion was met.</p><p>Second, the VPA perturbation set is non-trivial and anatomically interpretable. The suspect connections concentrate in circuits that prior work has independently implicated in social behaviour. This convergence is a form of external validity that gradient magnitude alone cannot provide. The model wasn’t told to find these circuits; it found them by backpropagating prediction error through a connectome-constrained network.</p><p>Third, the method is falsifiable. The suspect set is a testable hypothesis. Phase 3 will test it directly: all-optical perturbation of the top-ranked connections in wild-type fish should produce larger behavioural shifts than perturbation of control connections. If it does, the model’s attributions are causally validated. If it doesn’t, the model is wrong and we learn something about where.</p><p><strong>Limitations and Constraints</strong></p><p>No results section is complete without the caveats, so here they are.</p><p>The connectome prior is incomplete. Approximately 40% of the anatomical connections in the model are derived from population-level tracing data rather than a single-fish connectome. The proofreading feedback loop is designed to address this iteratively, but at present the model is constrained by a draft.</p><p>The VPA model is pharmacological, not genetic. Exposure at 4–48 hpf produces broad developmental effects that may not map cleanly onto synaptic weight changes. The gradient attribution is therefore best interpreted as identifying connections whose modulation is sufficient to reproduce the behavioural phenotype, not necessarily the connections that VPA primarily affects. This is a subtle but important distinction: the suspect set is a set of causal levers, not a set of lesion sites.</p><p>And the R² of 0.41, while above threshold, leaves substantial variance unexplained. The model captures the decision but not the full richness of the behavioural repertoire. Social approach is not a binary; it has kinematics, latencies, bout structures. Those are problems to be faced in the next step of this project.</p><p><strong>Next Steps</strong></p><p>Phase 3 proceeds on two parallel tracks.</p><p>Track A: all-optical validation. Using holographic optogenetics, the top-ranked connections will be perturbed in wild-type larval zebrafish during social approach assays. The prediction: gradient-ranked connections produce larger behavioural effects than control connections. If confirmed, the suspect set is causally validated.</p><p>Track B: proofreading integration. Feedback from the connectome proofreading team on the 29 prioritized edges will be incorporated, and the model retrained with any resolved connectivity. The question: does prediction accuracy on held-out VPA data improve when the connectome prior is more accurate?</p><p>The two tracks converge on a single criterion: whether the suspect set, refined by proofreading and validated optically, provides a causal account of the VPA social deficit that generalizes beyond the specific connections initially identified.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=63959b3957b3" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Turning Autism Spectrum Disorder into a Computable Neural Variable]]></title>
            <link>https://medium.com/@evilliz120/turning-autism-spectrum-disorder-into-a-computable-neural-variable-2bcb2ca1e36a?source=rss-6c83b8249c08------2</link>
            <guid isPermaLink="false">https://medium.com/p/2bcb2ca1e36a</guid>
            <category><![CDATA[endophenotype]]></category>
            <category><![CDATA[connectome]]></category>
            <category><![CDATA[larval-zebrafish]]></category>
            <category><![CDATA[autism]]></category>
            <category><![CDATA[connectomics]]></category>
            <dc:creator><![CDATA[lyh]]></dc:creator>
            <pubDate>Thu, 08 Oct 2026 16:31:14 GMT</pubDate>
            <atom:updated>2026-10-25T12:27:06.188Z</atom:updated>
            <content:encoded><![CDATA[<p><strong>Abstract</strong></p><p>This article explores two predicaments surrounding the larval zebrafish’s model and the heterogeneity of autism spectrum disorder, and one solution that can solve both of them simultaneously. By relying on the VPA-exposed zebrafish connectome model, to then analyse and break down the symptoms into a computable endophenotype variable, which can be used to train and build a connectome-to-behaviour causal model. This will not only kickstart an extremely more efficient form of connectome proofreading, it can also become a calculator for predicting social deficit behaviours, becoming a major step forward in behavioural psychology.</p><p><strong>Keywords</strong></p><p>Connectome, endophenotype, all-optical, valproic acid (VPA), autism spectrum disorder (ASD), larval zebrafish, causal interference, heterogeneity</p><p><strong>Introduction</strong></p><p><strong><em>I.</em></strong> Heterogeneity in Autism Spectrum Disorder</p><p>Autism spectrum disorder, otherwise known as ASD, is now relatively more recognised in different forms of media. While the awareness towards their “unexplainable behaviour” is growing, neurology says otherwise.</p><p>ASD is a highly genetically heterogeneous disorder, with a span of over 500 possible risk genes that might explain only 20% of ASD cases (<a href="https://pmc.ncbi.nlm.nih.gov/articles/PMC6710438/">1</a>). This genetic trait makes it absurdly difficult for scientests and psychologists to properly diagnose patients, as the “social deficit” consists of too many overly broad symptoms. This makes it nearly impossible to tell the difference between which genetic code would affect one’s behaviour, mobility, social cue detection and so on, becoming a bottleneck in the advancement of diagnosis and genetics (<a href="https://pubmed.ncbi.nlm.nih.gov/36532199/">2</a>).</p><p><strong><em>II</em></strong>. Difficulty in Mapping &amp; Proofreading Connectomes</p><p>Imagine being a cartographer, making a map of every single road, alleyway, driveway and footpath on Earth, but solely using blurry satellite images taken during a storm. This analogy perfectly depicts the pickle that the field of connectomics is stuck in now.</p><p>In this day and age of ever-advancing technology, only 4.5 connectomes have been fully mapped and proofread. The first being the drosophila, an adult fruit fly (Figure 1). The second is the roundworm, also known as Caenorhabditis Elegans (Figure 2 &amp; 3). Thirdly, a larval sea squirt (Ciona intestinalis) (Figure 3). Lastly, a larva marine annelid worm (Platynereis dumerilii) (Figure 3). Looking closely, one might notice that the biggest completely brain map as of this moment is the model of the drosophila fruit fly, containing roughly 125 million synaptic connections. (<a href="https://en.wikipedia.org/wiki/Connectome#Model_organisms_and_datasets">3</a>)</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/679/1*YGmpFditAgNUhKdeI5IXuQ.jpeg" /><figcaption>Figure 1: connectome of a drosophila (adult fruit fly) <a href="https://www.science.org/content/article/complete-map-fruit-fly-brain-circuitry-unveiled">https://www.science.org/content/article/complete-map-fruit-fly-brain-circuitry-unveiled</a></figcaption></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/738/1*dUPawakamE-XAKGDNWAyVA.jpeg" /><figcaption>Figure 2: connectome of a roundworm <a href="https://singularityhub.com/2019/07/18/the-first-complete-brain-wiring-diagram-of-any-species-is-here/">https://singularityhub.com/2019/07/18/the-first-complete-brain-wiring-diagram-of-any-species-is-here/</a></figcaption></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/489/1*X1Gy1Z2qzl82xBRUEawoMw.jpeg" /><figcaption>Figure 3: side-by-side comparison of connectomes of the dumerilli larval, elegans hermaphrodite and intestinals tadpole larvs <a href="https://www.researchgate.net/figure/The-Platynereis-dumerilii-larval-connectome-in-comparison-to-the-C-intestinalis-and-C_fig1_343819166">https://www.researchgate.net/figure/The-Platynereis-dumerilii-larval-connectome-in-comparison-to-the-C-intestinalis-and-C_fig1_343819166</a></figcaption></figure><p>And one can probably imagine how small their brain is.</p><p>The “0.5” mentioned above refers to the draft-model of the larval zebrafish, which holds a brain roughly the size of 1.5 millimeters long and less than 0.5 millimeters thick (<a href="https://www.theguardian.com/science/neurophilosophy/2013/mar/18/an-activity-map-of-the-whole-zebrafish-brain">4</a>). While this connectome is more or less mapped, it hasn’t been proofread in labs.</p><p>To grasp the reasoning behind this, the process behind mapping the connectome must first be understood.</p><p><strong>Context</strong></p><p><strong><em>I. </em></strong>Mapping the Connectome</p><p>Currently, the team of scientests of this field are dedicated to an endeavor coined as the Human Connectome Project (HCP), a massive collective effort at creating a comprehensive map of the connections within the human brain.</p><p>Our human brain is made up of layers and layers of complex material. At the largest scale is the cerebral cortex, responsible for high-level functions, such as thought, memory and language. Within them are 180 distinct cortical areas per hemisphere (<a href="https://theconversation.com/mapping-the-brain-scientists-define-180-distinct-regions-but-what-now-62972">5</a>). By using a Magnetic Resonance Imaging tool (MRI) to non-invasively image the modalities of the brain from a distance, a wiring diagram that builds the said cortical areas can be seen.</p><p>This wiring diagram are made up of neurons and synapses. These synapses are the function where neurons communicate, which define one’s behaviour, personality, identity, so on.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/636/1*koMxjToVzC4QecN4RZvKyA.jpeg" /><figcaption>side-by-side explanation of the difference between a synapse and 2 neurons <a href="https://www.getbodysmart.com/neurophysiology/neuron-synapse-structure/">https://www.getbodysmart.com/neurophysiology/neuron-synapse-structure/</a></figcaption></figure><p>Thinking of how many activities one individual can partake in a day, one might imagine how much information these synapses process in a minute. The truth is staggering.</p><p>A single 1-millimeter cube of human brain tissue contains approximately 50,000 neurons and around 1 billion synapses (<a href="https://www.proteinatlas.org/news/2024-05-22/a-single-cubic-millimeter-of-human-brain-has-generated-1-4-petabytes-of-data">6</a>). Then put that number in ratio with the brain’s size, which contains a volume of around 1,200 to 1,400 cubic centimeters (<a href="https://www.sciencedirect.com/topics/biochemistry-genetics-and-molecular-biology/brain-size">7</a>).</p><p>To put into a perspective of analogies, mapping the entire human brain of over 100 trillion synaptic connections is the equivalent to mapping every single twig, branch, root and sap <em>underground</em> across the entire Earth.</p><p>To put into a perspective of numbers, even mapping a single 1 millimeter cube of brain tissue using high resolution electron microscopy can generate a ludicrous amount of nearly 20000 terabytes of raw data. (<a href="https://news.harvard.edu/gazette/story/2024/05/the-brain-as-weve-never-seen-it/">8</a>)</p><p>A connectome is the map of all these synapses and neurons. The current technique used is Electron Microscopy (EM), where researchers fill the brain tissue with osmium, then use a diamond knife to create sections roughly 24–30 nanometers thick (4000 times thinner than a strand of hair!) (<a href="https://www.fmi.ch/news-events/articles/news.html?news=661">9</a>). Then, these slices are imaged, then stacked. By depicting the difference between the synapses and the brain tissue, researchers are able to reconstruct a 3D model of the micro wiring.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*B3pMAOmvfAPFOVrASeNa7A.png" /><figcaption>carving out one singular synapse using electron microscopy <a href="https://blog.biodock.ai/electron-microscopy-and-connectomics/">https://blog.biodock.ai/electron-microscopy-and-connectomics/</a></figcaption></figure><p>Then repeat this <em>exactly</em> 17,963 times with pieces 60nm thick, you’ll get a rough draft of how a larval zebrafish’s connectome would look like. (<a href="https://pmc.ncbi.nlm.nih.gov/articles/PMC5594570/">10</a>)</p><figure><img alt="cellular-resolution of the larval zebrafish brain https://www.bi.mpg.de/baier/connectome" src="https://cdn-images-1.medium.com/max/332/1*93gZSgyh9vOAFaHKsscOeA.jpeg" /><figcaption>the cellular resolution of the larval zebrafish brain <a href="https://www.bi.mpg.de/baier/connectome">https://www.bi.mpg.de/baier/connectome</a></figcaption></figure><p>After that, the images go through 3 respective processes.</p><ol><li>Automated Segmentation: a computer-driven process that uses algorithms to categorize, identify and classify the data and images into their respective synapse groups (exhibitory and inhibitory) by processing the specific genes, proteins and biochemical markers they express, which is what forms the rainbow images of the connectome.</li><li>Synaptic Detection: a process of identifying, then mapping the exact structures and locations of all the synapses. In other words, forming the map.</li><li>Synaptic Proofreading: a process of reviewing, then correcting computer-generated maps of the connectome, ensuring that all synapses are labeled accurately, as the raw automated segmentations are at a risk of split errors, where a single neuron branch might be broken into pieces.</li></ol><p>Raw synapse detection models only have a precision rate of 61.6% (<a href="https://www.biorxiv.org/content/10.1101/2025.07.11.664377.full">11</a>); a post-proofread connectome has a precision rate of 96.1% to 98.2% (<a href="https://pmc.ncbi.nlm.nih.gov/articles/PMC12363864/">12</a>), making synaptic proofreading ever more important before it exits the labs and is fully immersed in the research and medical fields to help treat patients.</p><p>When this painstaking process is complete, researchers will finally have the green light to perform other experiments on predicting behaviour, personality genetic traits, vice versa. This highlights the crucialness of obtaining an accurate connectome.</p><p><strong><em>II</em></strong>. The Connectome of the Larval Zebrafish</p><p>Due to the tricky procedures mentioned, this makes fully mapping the connectome of even a larval zebrafish a tedious endeavour. In this current stage, the automated segmentation and synaptic detection are complete.</p><p>Yet, the comprehensive proofreading remains ongoing, leaving both excitatory and inhibitory molecular annotations limited. Since researchers cannot yet assume the static graph of the larval zebrafish is truly the ground truth, which halts further in-depth research on its synaptic connectome.</p><p>Despite the current restraint, there are a few <strong>behaviour evident advantages</strong> the larval zebrafish possesses, compared to the other 4 mapped connectomes.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/596/0*89XMEv3CTx7dfhzd" /><figcaption>Larval zebrafish (<a href="https://www.news.uzh.ch/en/articles/2018/Zebrafish.html">https://www.news.uzh.ch/en/articles/2018/Zebrafish.html</a>)</figcaption></figure><ol><li>Larval zebrafish possess a see-through, camoflauge-like trait to blend into aquatic environments in early-stages of life (<a href="https://www.nature.com/articles/s42003-025-07496-z">14</a>). Due to the complete transparent form of their bodies and brains, it allows light to pass straight through. This becomes a more convenient form of simultaneous and non-invasive imaging for neural activity, as long as precise optogenetic manipulation across the brain. In the experiment later mentioned, this means, their transparency enables whole-brain calcium imaging at a single-cell resolution under freely behaving conditions. This all-optical interrogation is a major perk for researchers. (<a href="https://pmc.ncbi.nlm.nih.gov/articles/PMC6491009/">13</a>)</li><li>Despite being an aquatic freshwater organism, they share about 70% of human genes (<a href="https://pmc.ncbi.nlm.nih.gov/articles/PMC12632426/">15</a>). This becomes a useful development model for scientests to reference from for problems such as human diseases, or, which will be later mentioned in this article, human behaviour and genetic disorders.</li><li>Larval zebrafish have a fast development. Within 5 days, they can grow from a single cell to a swimming, feeding larval. Along with that, their small size also allows researchers to screen hundreds of drugs or genetic mutations within weeks in the organism, which is a whopping 1/10 to 1/20 fraction compared to the months or years it takes to administer the same experiment in mice, reducing 90%-95% of the time (16).</li><li>The movement behaviour of the larval zebrafish consists of habitual, distinct and punctuated swim bouts, separated by brief pauses. Their straightforward temporal structure makes automated video tracking and kinematic posture analysis clear-cut (<a href="https://europepmc.org/articles/PMC6465999">17</a>)</li></ol><p>These are simply the research and behavioural perks of the larval zebrafish. Below are the primary advantages of why the larval zebrafish would the prime option for an ASD-model experiment despite not having a proofread connectome.</p><ol><li>The whole-brain neural activity in the larval zebrafish allows researchers to extract a quantifiable, perturbable and predictive “social approach state variable”. This turns the observatory data of connectomics into actual tangible data for further research.</li><li>Following the behavioural advantage of whole-brain calcium imaging at single-cell resolution, research has identified whole-brain activity patterns that appear seconds before social approach. The pallium activity in the brain increases while other brain regions decreases. This coordinated dynamic can predict whether approach behaviour is about to initiated, which further correlates with individual differences in social behaviour.</li><li>Most importantly, the larval zebrafish has a validated VPA-exposure model. This produces quantifiable social deficit phenotypes and altered expressions of ASD-related genes, such as <em>adsl, mbd5, shank3, tsc1b, </em>etc (<a href="https://pmc.ncbi.nlm.nih.gov/articles/PMC11487397/">18</a>).</li></ol><p>III. What is VPA?</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/390/0*Qq8IWxx7_0b22hTW" /><figcaption>Valproate <a href="https://en.wikipedia.org/wiki/Valproate">https://en.wikipedia.org/wiki/Valproate</a></figcaption></figure><p>Valproic acid, in short, VPA, is an anticonvulsant and mood stabiliser that acts as a potent teratogen when exposed to a developing fetus. This is also one of the most widely accepted and thoroughly validated epigenetic models used to study the neurobiology of autism (<a href="https://pmc.ncbi.nlm.nih.gov/articles/PMC4688329/">19</a>).</p><p>This means, social deficit behaviours can be engineered through administrated doses of VPA. While this may seem inhumane, valproic acid doses can be reversible through oxytocin, returning the subject back to it’s original state.</p><p>The VPA-exposed zebrafish model is one of the most established ASD models up to date. To start, after enough embryonic VPA-exposure at 0–48 hpf, it can engineer persistent social interaction deficits at around 7 dpf.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/500/0*yp4e-uh3QPSePOeo.png" /></figure><p>Since the heterogeneity of ASD traits stop scans and diagnosis from targeting the true origin of behaviour for treatment. This VPA-exposed model can help clear the behavioural phenotypes, eliminating other risk genes clouding the experiment results, and still use a positive control drug.</p><p>IV. What are Endophenotypes?</p><p>Endophenotypes are a measurable, internal trait that sits at the biological pathway between a person’s genes and a complex disease they possess. Said traits are usually one of a cognitive, biochemical or neuropsychological feature, with examples such as circadian rhythm adnormalities, eye-tracking dysfunction, response inhibitation impairment. These 3 mentioned instances can be diagnostic symptoms for bipolar disorder, schizophrenia and attention deficit/ hyperactivity disorder (ADHD) respectively.</p><p>With that being said, that means endophenotypes are not just an outward behaviour or medical syndrome. They are a direct string tracing behaviour, identity and disorders back to a distant genetic code origin.</p><p>These traits, also known as phenotypes, can take a complex and heterogeneous condition, then further dissect them into smaller, biologically clearer pieces by categorizing them between 2 measurable intermediate phenotype sections. The first, genotypes, a behaviour explained by genetic code. The second, external behaviour, an action shaped by the environment around the subject.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/640/1*3h9HJjMdNcBgHs0l5KasPA.png" /><figcaption>Table for endophenotypes, phenotypes &amp; biomarkers <a href="https://www.semanticscholar.org/paper/The-Cutting-Edge-ENDOPHENOTYPE-%2C-INTERMEDIATE-%2C-%3A-%2C-Lenzenweger/beb7d29d49fd9d889b6a91db228a18672684a799">https://www.semanticscholar.org/paper/The-Cutting-Edge-ENDOPHENOTYPE-%2C-INTERMEDIATE-%2C-%3A-%2C-Lenzenweger/beb7d29d49fd9d889b6a91db228a18672684a799</a></figcaption></figure><p>While this sounds simple and all, there is a common misconception. While endophenotypes can be a sign for a disease, it doesn’t define the full illness. Many patients can have the endophenotypes without showing all the clinical symptoms of the actual disease. That could merely mean that the patient only has the biological marker of the symptom passed from by familial genes or, the patient solely contains the endophenotype and have yet to develop the actual clinical symptoms due to lack of environmental triggers. This is part of what makes the heterogenity in disorders so difficult to diagnose properly.</p><p>V. Endophenotypes in Autism Spectrum Disorder</p><p>For autism spectrum disorder, there are 3 main endophenotypes:</p><ol><li>Macrocephaly. This genotype is usually found in children. These children possess an enlarged head size linked the specific genetic mutations such as PTEN.</li><li>Elevated platelet serotonin levels. This genotype causes the normal baseline of serotonin in blood platelets to be higher that usual. This could lead to other cardiovascular and microvascular signs, such as headaches, erythromelalgia, cardiovascular strains and unusual bleeding or bruising.</li><li>First words. In children who possess genotypical signs of ASD, they might have delays in early milestones, especially speaking. This can be liked with genes such as CNTNAP2. (<a href="https://link.springer.com/rwe/10.1007/978-1-4419-1698-3_1324">20</a>)</li></ol><p>VI. Computationalizing Autism Endophenotypes into a Neural Variable</p><p>Now, to the most important part of this article. Putting two and two together, combining how the complexity of connectomics can solve the intricacies of heterogeneities through valproic acid in the larval zebrafish model.</p><p>To understand computationalizing neural variables, causal modelling must first be learnt. Causal modelling is a statistical and analytical method to identify, quantify and represent the direct cause and effect relationships between neural variables, such as individual neurons or brain regions on a macro-scale.</p><p>A standard functional connectome map shows whether brain areas are active at the same time. On the other hand, causal modelling builds a causal functional connectome, which defines the <em>effective </em>connectivity of the brain and proving how activity in one region can drive, change or suppress activity in another (<a href="https://arxiv.org/abs/2204.04845">23</a>). In other words, a standard connectome is the lights, showing whether the switch is on; a causal connectome is the wire, which shows which switch turned said light on.</p><p><strong>Step 1: </strong>operationalize the endophenotype variable</p><p>Currently, existing neuroimaging research has identified promising intermediate phenotypes, such as social brain hypoactivation (<a href="https://pmc.ncbi.nlm.nih.gov/articles/PMC4523891/">21</a>), multi-dimensional neuroanatomical profiling (<a href="https://pmc.ncbi.nlm.nih.gov/articles/PMC10157419/">22</a>) and reduced inferior frontal gyprus activation. Yet, these findings are often phenotypically anchored instead of being formally operationalised.</p><p>Yet, to enable the causal modelling, an endophenotype must first be defined as a quantifiable neural variable with explicit computational parameters that can be an estimable data to be used experimentally. This requires specifying their measurement scale, its relation to latent computational processes and mapping it onto observable neural activity.</p><p>An example would be an endophenotypic gradient in the inferior frontal gyrus (IFG), which has been recently heavily highlighted in neuroimaging research on ASD. This gradient is a linear variation in neural activation or volume that tracks the level of genetic liability for a neurodevelopmental or psychiatric condition, which makes it especially helpful for ASD neuroimaging (<a href="https://pmc.ncbi.nlm.nih.gov/articles/PMC10917220/">24</a>).</p><p>In summary of the research, scientists used a fMRI scanned the activity in the left IFG when the subjects performed tasks requiring semantic judgement or language processing. The results showed individuals with autism have reduced activation and the weakest neural activation, on the other hand, unaffected siblings exhibit intermediate responses, despite still sharing the genetic risk (<a href="https://onlinelibrary.wiley.com/doi/pdf/10.1002/aur.70320">25</a>).</p><p>This result can be defined a computational variable, which would be estimated via a parametric model of trial-wise activation. Next, yielding a per-subject coefficient that serves as the operational definition of the endophenotype. Thus, making it both a neural variable derived from imaging data and a computational variable defined by a formal model of the cognitive operation at the same time. Ultimately, successfully establishing the operational definition of the endophenotype variable.</p><p><strong>Step 2:</strong> apply the operationalised variable into causal modelling</p><p>Once operationalised, the variable becomes amenable to causal modelling. By taking the produced data into a framework that encodes the cause and effect relationships, also known as generative modelling, it can specify how genetic liability, developmental path and compensatory mechanisms can determine the value of the neural variable altogether. Consequently, predict social functioning incomes, allowing researchers to read the connectome behaviour.</p><p>More importantly, the computational variable can also serve as a a target for perturbation, detecting changes in the regular behaviour of a subject, then further connecting that to the perturbed part of the connectome, which can change pure observation in experimenting in the brain into intervention, allowing all-optical validation to be possible.</p><p>Techniques such as ADePT give researchers full, independent control of imaging and photostimulation, allowing manipulation of connectome activity in forms of insertion, removal, rewiring and tranplantation, while simultaneously comprehending and identifying the behavioural responses. (<a href="https://elifesciences.org/reviewed-preprints/112219v1.pdf#10%231">26</a>)</p><p>If the operationalised endophenotype matches up with a specific pattern of prefrontal-cortical changes, researchers could test its causal role by optogenetically establishing that pattern in model systems and observing whether social behavior shifts according to their behaviour. <br> <br>A summary of this experimental path in this:</p><ol><li>Define the neural variable operationally</li><li>Embed it in a causal model</li><li>Validate the model by all-optical manipulation</li></ol><p>Recent work in mouse models of autism has already begun to identify computational anomalies — such as altered use of prior expectations — that can be mis-quantified as neural variables, instead, the real reason behind the change being linked to circuit dysfunctions (<a href="https://docs.internationalbrainlab.org/notebooks_external/2025_data_release_autism_davatolhagh.html">27</a>).</p><p>All-optical toolkits now support completely connected interrogation in behaving animals, where stimulation is guided by real-time observation of ongoing activity. When combined, these developments make it feasible to move from a solely correlational endophenotype to causal validation, given that the neural variable is defined with sufficient computational precision to be both estimated and established (<a href="https://pubmed.ncbi.nlm.nih.gov/40972568/">28</a>). ‌</p><p><strong>Conclusion</strong></p><p>I. Future</p><p>With clear steps on operationalising endophenotypes into causal modelling, to train and validate a VPA model of the larval zebrafish, then teach the observation model to predict the minimal perturbation in ASD behaviour, a model to both turn behaviour changes into a proofreading pathway for the larval zebrafish connectome, as well as using the current phenotypes the pinpoint the exact genes for ASD can be built.</p><p>As they say, 2 birds with 1 stone.</p><p><strong>References</strong></p><p>(1) <a href="https://pmc.ncbi.nlm.nih.gov/articles/PMC6710438/">https://pmc.ncbi.nlm.nih.gov/articles/PMC6710438/</a></p><p>(2) <a href="https://pubmed.ncbi.nlm.nih.gov/36532199/">https://pubmed.ncbi.nlm.nih.gov/36532199/</a></p><p>(3)<a href="https://en.wikipedia.org/wiki/Connectome#Model_organisms_and_datasets">https://en.wikipedia.org/wiki/Connectome#Model_organisms_and_datasets</a></p><p>(4)<a href="https://www.theguardian.com/science/neurophilosophy/2013/mar/18/an-activity-map-of-the-whole-zebrafish-brain">https://www.theguardian.com/science/neurophilosophy/2013/mar/18/an-activity-map-of-the-whole-zebrafish-brain</a></p><p>(5) <a href="https://theconversation.com/mapping-the-brain-scientists-define-180-distinct-regions-but-what-now-62972">https://theconversation.com/mapping-the-brain-scientists-define-180-distinct-regions-but-what-now-62972</a></p><p>(6) <a href="https://www.proteinatlas.org/news/2024-05-22/a-single-cubic-millimeter-of-human-brain-has-generated-1-4-petabytes-of-data">https://www.proteinatlas.org/news/2024-05-22/a-single-cubic-millimeter-of-human-brain-has-generated-1-4-petabytes-of-data</a></p><p>(7) <a href="https://www.sciencedirect.com/topics/biochemistry-genetics-and-molecular-biology/brain-size">https://www.sciencedirect.com/topics/biochemistry-genetics-and-molecular-biology/brain-size</a></p><p>(8) <a href="https://news.harvard.edu/gazette/story/2024/05/the-brain-as-weve-never-seen-it/">https://news.harvard.edu/gazette/story/2024/05/the-brain-as-weve-never-seen-it/</a></p><p>(9) <a href="https://www.fmi.ch/news-events/articles/news.html?news=661">https://www.fmi.ch/news-events/articles/news.html?news=661</a></p><p>(10) <a href="https://pmc.ncbi.nlm.nih.gov/articles/PMC5594570/">https://pmc.ncbi.nlm.nih.gov/articles/PMC5594570/</a></p><p>(11) <a href="https://www.biorxiv.org/content/10.1101/2025.07.11.664377.full">https://www.biorxiv.org/content/10.1101/2025.07.11.664377.full</a></p><p>(12) <a href="https://pmc.ncbi.nlm.nih.gov/articles/PMC12363864/">https://pmc.ncbi.nlm.nih.gov/articles/PMC12363864/</a></p><p>(13) <a href="https://pmc.ncbi.nlm.nih.gov/articles/PMC6491009/">https://pmc.ncbi.nlm.nih.gov/articles/PMC6491009/</a></p><p>(14) <a href="https://www.nature.com/articles/s42003-025-07496-z">https://www.nature.com/articles/s42003-025-07496-z</a></p><p>(15) <a href="https://pmc.ncbi.nlm.nih.gov/articles/PMC12632426/">https://pmc.ncbi.nlm.nih.gov/articles/PMC12632426/</a></p><p>(16) <a href="https://www.mpg.de/10973793/why-do-scientists-investigate-zebra-fish">https://www.mpg.de/10973793/why-do-scientists-investigate-zebra-fish</a></p><p>(17) <a href="https://europepmc.org/articles/PMC6465999">https://europepmc.org/articles/PMC6465999</a></p><p>(18) <a href="https://pmc.ncbi.nlm.nih.gov/articles/PMC11487397/">https://pmc.ncbi.nlm.nih.gov/articles/PMC11487397/</a></p><p>(19) <a href="https://pmc.ncbi.nlm.nih.gov/articles/PMC4688329/">https://pmc.ncbi.nlm.nih.gov/articles/PMC4688329/</a></p><p>(20) <a href="https://link.springer.com/rwe/10.1007/978-1-4419-1698-3_1324">https://link.springer.com/rwe/10.1007/978-1-4419-1698-3_1324</a></p><p>(21) <a href="https://pmc.ncbi.nlm.nih.gov/articles/PMC4523891/">https://pmc.ncbi.nlm.nih.gov/articles/PMC4523891/</a></p><p>(22) <a href="https://pmc.ncbi.nlm.nih.gov/articles/PMC10157419/">https://pmc.ncbi.nlm.nih.gov/articles/PMC10157419/</a></p><p>(23) <a href="https://arxiv.org/abs/2204.04845">https://arxiv.org/abs/2204.04845</a></p><p>(24) <a href="https://pmc.ncbi.nlm.nih.gov/articles/PMC10917220/">https://pmc.ncbi.nlm.nih.gov/articles/PMC10917220/</a></p><p>(25) <a href="https://onlinelibrary.wiley.com/doi/pdf/10.1002/aur.70320">https://onlinelibrary.wiley.com/doi/pdf/10.1002/aur.70320</a></p><p>(26) <a href="https://elifesciences.org/reviewed-preprints/112219v1.pdf#10%231">https://elifesciences.org/reviewed-preprints/112219v1.pdf#10%231</a></p><p>(27) <a href="https://docs.internationalbrainlab.org/notebooks_external/2025_data_release_autism_davatolhagh.html">https://docs.internationalbrainlab.org/notebooks_external/2025_data_release_autism_davatolhagh.html</a></p><p>(28) <a href="https://pubmed.ncbi.nlm.nih.gov/40972568/">https://pubmed.ncbi.nlm.nih.gov/40972568/</a></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=2bcb2ca1e36a" width="1" height="1" alt="">]]></content:encoded>
        </item>
    </channel>
</rss>