Introduction

To prevent or slow down disease, therapies must target the biological processes that cause the disease. Genomic approaches like family studies or genome-wide association studies (GWASs) can help in this quest for causal processes by exposing genes that are mutated before disease onset. For finding causal processes, studying the genome, rather than the transcriptome or epigenome, is advantageous because the chronology from genomic variant to disease is unambiguous, providing a strong argument for causality. In theory, all we need to do after a genomic study is follow the thread from each gene to the biological process in which it is involved. We know from the genomic study that this process increased or reduced risk when it was altered by mutations in the gene, so modulating more forcefully the process with a drug may unlock a larger therapeutic benefit. In practice, the path from genomic variant to druggable process is far from straightforward.

Research on Alzheimer’s disease (AD) exemplifies well both the benefit and challenge of translating genomic studies into druggable biological processes. Family studies of early-onset AD in the 1990’s¹ identified causal mutations in amyloid precursor protein (APP)^2,3, presenilin 1 (PSEN1)⁴, and presenilin 2 (PSEN2)⁵. Subsequent work demonstrated that amyloid beta (Aβ), a small peptide which forms aggregates in AD patients’ brains⁶, was generated by cleavage of APP by γ-secretase^7,8, of which the catalytic subunit is PSEN1 or PSEN2^9,10. Consequently, the field naturally converged on the amyloid hypothesis of AD, which posits that the disease is caused by toxic aggregates of Aβ¹¹. Today, antibodies against Aβ such as lecanemab show promise in slowing down disease progression¹². This story shows that, although a great challenge, genomic studies (family studies of early-onset AD) can be successfully translated into a causal process (Aβ aggregation) that is now targeted by disease-modifying drugs (lecanemab).

However, the beneficial effects from targeting Aβ aggregation currently remain modest despite substantial reductions in brain amyloid burden¹². To completely stop disease progression, anti-amyloid therapy will likely need to be combined with drugs modulating other processes that contribute to the disease¹³. GWASs have identified tens of genomic loci where sequence variation is associated with late-onset AD, offering an opportunity to discover new causal processes of AD that potentially go beyond the amyloid hypothesis. For example, analysis of cell types enriched for open chromatin at AD-associated loci pointed to a possible critical role of monocytes, macrophages, and microglia in AD progression^14,15. Although GWASs are designed to generate new hypotheses, AD-associated loci have rarely been exploited to find new causal processes in an unbiased, systematic manner. Given the challenges inherent to linking genomic variants to causal biological events, new genomic associations are often first more narrowly interpreted in the context of the amyloid hypothesis. While this interpretation may be correct, it tends to create a self-fulfilling prophecy which leaves little room for the discovery of new causal processes^16,17.

As AD is primarily a disease of the old age, research often focuses on patients or animal models after disease onset. However, this approach also hinders the discovery of new causal processes from genomic studies because it largely annuls the advantage of unambiguous chronology—since genomic variants were unambiguously present before disease onset, they must have modulated a process which contributed causally to disease initiation. In patients or animals after disease onset, many biological processes are disrupted, but only a small proportion may be genuinely causal for disease progression and therefore make suitable targets for disease-modifying therapies. For example, one may find that dopamine is lacking in the basal ganglia of patients and animal models of Parkinson’s disease. Treatment with levodopa, a dopamine precursor, temporarily relieves motor symptoms, but does not slow down disease progression, so dopamine deficiency was not in fact a causal process¹⁸. Therefore, perhaps counterintuitively, studying the consequences of AD-associated mutations early in life seems more likely to identify processes that are genuinely causal to disease, as any disrupted process that is identified is less likely to be a secondary disease consequence.

In practice, how can we quickly follow the thread from a disease-associated gene to a (druggable) biological process in which it is involved? In this work, we describe a behavioural pharmacology approach to link genes associated with a disease to likely causal biological processes using zebrafish larvae. Our strategy predicts from the behavioural profile of knockout zebrafish the likely causal processes and potential compounds that rescue the phenotype. In previous work¹⁹, we introduced the use of zebrafish F0 knockouts to study complex traits such as behaviour. In this study, we demonstrate how the F0 knockout method renders this behavioural pharmacology approach fast and scalable to the parallel study of tens of disease-associated genes, rather than one at a time^20,21. As genomic studies of AD can likely be further exploited to find causal processes, we used Alzheimer’s risk genes as a case study for our strategy. The strategy is not specific to any one disease or set of genes. In theory, any measurable change in behaviour could be used to predict the underlying causal pathways and small molecules that normalise this change.

Results

Most Alzheimer’s risk genes are present in zebrafish and expressed early in development

GWASs point to small portions of the genome where variation in sequence is associated with variation in disease risk, but do not readily specify the genes whose altered function are responsible for this association. Therefore, as a starting point, we used a meta-analysis of GWASs on AD that found 37 significant loci and annotated each with the most likely causal gene using mainly statistical colocalisation^22,23. To add confidence to these calls, we cross-referenced these causal gene predictions with a transcription-wide association study and risk gene pre-mRNAs that undergo differential splicing in AD brains²⁴. Finally, we added to our list the three genes which can cause early-onset AD when mutated: PSEN1⁴, PSEN2⁵, and APP^2,3, yielding a list of 40 genes associated with AD risk (Supplementary file 1).

Of these 40 Alzheimer’s risk genes, 30 (75%) had at least one annotated orthologue in the zebrafish genome (source: Ensembl). Of those, 17 had one orthologue (e.g. the only zebrafish orthologue of human SORL1 was sorl1); 11 had two orthologues (e.g. the zebrafish orthologues of human APP were appa and appb); and 2 had more than two orthologues (e.g. the zebrafish orthologues of human MS4A6E included ms4a17a, ms4a17c.2, tmem176l, and more) (Fig. 1a). A human gene often has two or more zebrafish orthologues because of a teleost-specific whole-genome duplication event around 340 million years ago²⁵. There were no annotated orthologues for 10 Alzheimer’s risk genes, including TREM2.

Next, we used a published single-cell RNA sequencing (scRNA-seq) dataset of the developing zebrafish brain²⁶ to ask whether the orthologues of the Alzheimer’s risk genes were expressed in zebrafish embryos and larvae. Most of the genes (33/42) were detectable as early as 12 hours post-fertilisation (hpf) and remained expressed throughout development (Fig. 1b). At 5 days post-fertilisation (dpf), 38 of the 42 orthologues (90%) were expressed.

From these observations, we selected seven high-confidence Alzheimer’s risk genes for further study in zebrafish; the orthologues of the three early-onset Alzheimer’s genes: psen1, psen2, appa/appb; and four genes associated with late-onset AD: apoea/apoeb, cd2ap, clu, sorl1. We chose APOE as it is the most well-known genetic risk factor for late-onset AD²⁷. CD2AP, CLU, and SORL1 were chosen because non-coding variants within or near those genes are repeatedly found by GWASs^28,29. At the CD2AP locus, CD2AP was highly likely to be the causal gene by colocalisation²². At the CLU locus, both PTK2B and CLU were likely causal²², but differential splicing of CLU correlated with risk of AD²⁴. The top variant at the SORL1 locus was within an intron of SORL1²², but rare protein-coding variants in SORL1 were likely causal for some early-onset AD patients through haploinsufficiency or deleterious effects on protein function^30–32.

For these selected Alzheimer’s risk genes, we more carefully examined their expression patterns in larval zebrafish. From the scRNA-seq dataset²⁶, most of these genes were broadly expressed in the 5-dpf larva in different neuronal populations and other cell types (Fig. 1–supplement 1 and 2). apoeb was highly expressed specifically in epidermis progenitors, Müller glia in the retina, and microglia (Fig. 1c), as observed previously^33–36. Across clusters, the highest expression of clu was in radial glia (Fig. 1–supplement 2c, cluster 30). To confirm and extend these observations, we used in situ hybridization chain reaction (HCR) to label mRNA in 6-dpf larvae. As we observed in the scRNA-seq data, most genes tested (appa, appb, psen1, psen2, apoea, cd2ap, sorl1) were broadly expressed throughout the 6-dpf brain (Fig. 1d and Fig. 1–supplement 3 and 4). apoeb expression was restricted to cells in the forebrain and optic tectum (Fig. 1d and Fig. 1–supplement 4b). Based on the scRNA-seq dataset (Fig. 1c) and literature^33,37, these cells are likely ccl34b.1⁺ amoeboid microglia derived from primitive myeloid precursors from the rostral blood island. Additionally, apoeb was detected in cells bordering the hindbrain ventricle, which are likely radial glia/astrocytes^38,39. Contrary to the widespread expression found in the scRNA-seq dataset, clu expression was largely restricted to the diencephalic and myelencephalic choroid plexuses (Fig. 1d and Fig. 1–supplement 4d), confirming a previous report⁴⁰.

In summary, around 75% of Alzheimer’s risk genes had at least one clear orthologue in zebrafish and most of these were expressed in the brain of 5–6-dpf zebrafish larvae, suggesting they play a role in early brain development or function.

Fig. 1:

Fig. 1–suppl. 1:

Fig. 1–suppl. 2:

Fig. 1–suppl. 3:

Fig. 1–suppl. 4:

The FramebyFrame R package for analysis of sleep/wake behaviour from high-throughput video-tracking data

The next stage in our behavioural pharmacology strategy is to measure sleep/wake behaviour of F0 knockout larvae for the genes under study. To uncover even subtle behavioural phenotypes caused by the loss of Alzheimer’s risk genes, we developed a high-throughput sleep/wake tracking assay for zebrafish larvae capable of analysing behaviour at the sub-second resolution over multiple days and nights. To achieve this, we combined previous sleep/wake analysis methods^41–43 with some aspects of the frame-by-frame analysis developed by Ghosh and Rihel (2020)⁴⁴ into a single software tool, the FramebyFrame R package (github.com/francoiskroll/FramebyFrame). We also designed a 3D-printed mesh-bottom plate that supports long-term (up to 8.5 days tested) tracking of larvae with minimal intervention by regulating the water level with a small pump and delivering paramecia for feeding through the mesh from the water bath below (Fig. 2a,b).

The FramebyFrame package extracts and analyses 17 behavioural parameters from frame-by-frame Δ pixel data, which represent the number of pixels that changed intensity at each frame-to-frame transition. These parameters capture both behaviours that unfold over multiple minutes or hours, such as sleep (Fig. 2c), and actions at smaller time scales (< 1 sec), such as individual swimming bouts (Fig. 2d). The parameters are grouped into three categories: activity parameters (Fig. 2–supplement 1a), which describe the overall activity of each larva during a complete day or night; active bout parameters (Fig. 2–supplement 1b), which describe the structure of individual swimming bouts; and sleep parameters (Fig. 2c) such as sleep latency (Fig. 2–supplement 2d). In zebrafish larvae, sleep is defined as any period of inactivity (Δ pixel = 0) lasting longer than one minute, a definition based on increases in arousal threshold and homeostatic rebound following deprivation^43,45. This frame-by-frame analysis has several advantages over previous methods that analysed activity data at the one-minute resolution. First, individual swimming bouts could not be resolved by these methods as single bouts last ∼0.2 sec on average. Second, the one-minute methods missed around one third of all sleep bouts because of how the Δ pixel data is binned in one-minute epochs (Fig. 2–supplement 2a). At the frame-by-frame resolution, the start and end frame of each sleep bout can be precisely determined, which also improves the measurement accuracy of sleep bout duration and sleep latency, the delay from lights-off to first sleep bout. For example, during a sample 10-hr night (Fig. 2–supplement 2b), the frame-by-frame analysis detected 35 ± 14 more sleep bouts, longer average sleep bout lengths (+0.2 ± 0.2 min), and shorter sleep latencies (1.6 ± 2.3 min earlier), resulting in a 42% increase in total sleep time (+2.3 ± 0.4 hr). During the day, the FramebyFrame analysis also detected more sleep and more sleep bouts (Fig. 2–supplement 2c). Most behaviour plots included in this study were created using the FramebyFrame package in just a few commands from the raw frame-by-frame data.

Fig. 2:

Fig. 2–suppl. 1:

Fig. 2–suppl. 2:

psen2 knockouts sleep more during the day

Next, we used CRISPR-Cas9 to generate F0 knockouts for the zebrafish orthologues of the three genes that can cause early-onset AD when mutated in humans—psen1, psen2, appa/appb—and tested these mutants for behavioural and other phenotypes. We focused first on psen1 and psen2.

Zebrafish Psen1 and Psen2 are highly similar to their human counterparts (∼70% identical amino acid sequence, Fig. 3a,b), with the same critical amino acid at the four annotated active sites (two per presenilin). Mismatches between the human and zebrafish proteins were largely clustered into disordered domains, suggesting that the zebrafish presenilins have the same catalytic activity as in humans. The Cas9/guide RNA ribonucleoproteins (RNPs) used to generate F0 knockouts mutated virtually every copy of the genome at each of the three targeted sites. At each targeted locus of psen1, 99.0 ± 2.7% of reads were mutated and 78.6 ± 29.7% of all reads had a frameshift mutation (Fig. 3c). At each targeted locus of psen2, 99.9 ± 0.1% reads were mutated and 82.0 ± 33.6% of all reads had a frameshift mutation (Fig. 3d). As each gene is mutated at three loci, the biallelic knockout probability, that is, the probability of having at least one frameshift mutation on both alleles, cumulatively reached > 98% (francoiskroll.shinyapps.io/frameshiftmodel/), indicating that most psen1 and psen2 F0 knockout larvae were complete loss-of-function mutants.

In humans, PSEN1 or PSEN2 acts as the catalytic subunit of the γ-secretase protein complex, which is responsible for the cleavage of more than 90 substrates, such as Notch. It also performs the last cleavage of APP to release Aβ, which aggregates in the brains of patients with AD^7,9,46,47. Do presenilins also cleave zebrafish Appa/Appb into Aβ? Using a sensitive ELISA-based assay, Aβ40 and Aβ42 were detectable in control (uninjected and scrambled-injected) larvae but not in two clutches of psen1 F0 knockouts (Fig. 3e), confirming that zebrafish Psen1 is required for the cleavage of Appa/Appb into Aβ. In contrast, Aβ40 and Aβ42 were detectable in psen2 F0 knockouts, suggesting that most Aβ is generated by Psen1 in zebrafish and that Psen2 alone cannot compensate. This result extends previous findings that Aβ production in zebrafish is blocked by the γ-secretase inhibitor DAPT⁴⁸ and matches closely observations in mice, in which loss of PSEN2 had no measurable effect on Aβ40/42 production⁴⁹.

In mice, knockout of Psen1, but not of Psen2, causes severe skeletal malformations and brain haemorrhages, likely because of impaired Notch signalling⁵⁰. Homozygous animals die within minutes after birth⁵¹. Morphologically, our psen1 F0 knockout larvae developed normally (tested up to 16 dpf) and were indistinguishable from their wild-type siblings, as observed in a psen1 stable knockout line⁵². psen2 F0 knockout larvae had a mild pigmentation defect (Fig. 3f and Fig. 3–supplement 2a), which has been previously reported in a psen2 stable knockout line⁵³. This difference suggests a specific function of Psen2 in melanophores which cannot be fulfilled by Psen1. We also generated psen1/psen2 double F0 knockouts but most were lethal early in development (∼5 dpf) with severe defects in eye development (Fig. 3–supplement 1a), in which the retinal pigment epithelium appeared patchy and some larvae developed oedema around the eye. The tail was severely curved outwards, in exactly the same way as larvae treated with a high dose of the γ-secretase inhibitor DAPT⁵⁴. Incidentally, Psen1 knockout mice also display defects in the axial skeleton⁵¹. Overall, these results suggest that zebrafish presenilins are more readily capable of compensating each other during development than in mice, as only double psen1/psen2 knockout larvae showed a severe morphological phenotype.

We next video-tracked two clutches of psen1 and psen2 F0 knockout larvae over multiple day-night cycles and applied our frame-by-frame analysis to detect behavioural phenotypes over long and short time scales. The loss of psen1 only had mild effects on behaviour. At night, psen1 F0 knockout larvae slept slightly less (−8% vs. scrambled, Fig. 3–supplement 1b) and instead spent more time active than control injected siblings (+22%), mainly because they performed more swimming bouts (+20%). The psen1 knockouts also showed a slightly reduced startle response at lights-off (−4%), in line with some data obtained from stable psen1 knockout larvae⁵². Behaviour during the day was not affected (Fig. 3–supplement 1c). In contrast, psen2 F0 knockouts of both cutches were substantially less active than controls during the day (total activity: −26% vs. scrambled, Fig. 3g,h and Fig. 3– supplement 1d), performing both fewer (−17%) and more subdued swimming bouts (active bout mean: −12%). psen2 F0 knockouts also slept more during the day than controls (Fig. 3g,h, Fig. 3–supplement 1d), both spending more time asleep (+178%) and initiating more frequent sleep bouts (+150%). Loss of Psen2 did not strongly affect night-time behaviour. In summary, psen2 F0 knockout larvae were substantially less active and sleeping more than controls during the day.

Since psen2 knockout larvae were less pigmented (Fig. 3–supplement 2a), we tested whether the reduction in activity could be an artefact due to fainter detection by the camera. We extracted the maximum Δ pixel value each larva reached during the startle response at lights off (Fig. 3–supplement 2b). These vigorous swimming bouts can displace the whole larva in a tenth of a second, providing a measure of how many dark pixels are detectable for each larva. psen2 F0 knockout larvae indeed displaced fewer pixels during the startle response than scrambled-injected controls (Fig. 3–supplement 2c, clutch 1: psen2 knockouts displaced 78 ± 10 pixels vs 82 ± 8 pixels for scrambled-injected controls, clutch 2: 71 ± 8 pixels vs 80 ± 8 pixels). We then downscaled the Δ pixel values of the scrambled-injected larvae proportionally to the ratio of the startle response means (Fig. 3–supplement 2c,d), akin to artificially making the scrambled-injected larvae appear less dark. Even after this adjustment, psen2 F0 knockouts were measurably less active and sleeping more during the day, although the parameter effect sizes were somewhat reduced (Fig. 3–supplement 2e). In summary, the psen2 behavioural phenotype reflects a genuine difference in behaviour.

Finally, we examined the impact of loss of both Appa and Appb. As described previously^48,55, zebrafish Appa and Appb are highly similar to human APP (64–68% identical amino acid sequence, Fig. 3–supplement 3a,b). We generated appa/appb double F0 knockout larvae by mutating each gene at two loci, rather than three, to limit unviability¹⁹. The CRISPR-Cas9 RNPs were highly mutagenic (across all four loci: 95.8 ± 7.0% mutated reads, 76.5 ± 21.7 of all reads had a frameshift mutation, Fig. 3–supplement 3c,d). The appa/appb double F0 knockout larvae looked morphologically normal. Behaviourally, appa/appb double knockouts were less active during the day than control siblings (–14% vs. scrambled, Fig. 3–supplement 3e) and showed shorter swimming bouts across the day/night cycle (active bout duration during the day: –4%; at night: –3%). Sleep was not consistently affected by the loss of appa/appb (Fig. 3–supplement 3f).

In summary, knockout larvae for the three genes associated with early-onset AD had distinct morphological and behavioural phenotypes, with the strongest behavioural changes observed for psen2 knockouts. Since both presenilins are broadly expressed in the larval brain (Fig. 1d and Fig. 1–supplement 3c,d), these results indicate that zebrafish Psen1- and Psen2-γ-secretases likely cleave different substrates, such as Appa/Appb which is primarily cleaved into Aβ by Psen1-γ-secretase. Moreover, the behavioural phenotypes of appa/appb and psen1 knockout larvae had little overlap while they presumably both resulted in the loss of Aβ. The appa/appb day phenotype could be primarily caused by loss of some Appa/Appb cleavage product not relying on γ-secretase. Alternatively, the net effect of loss of all other Psen1-γ-secretase products could have masked the effect of loss of Aβ.

Fig. 3:

Fig. 3–suppl. 1:

Fig. 3–suppl. 2:

Fig. 3–suppl. 3:

sorl1 knockouts are less active during the day and sleep less at night

To test whether knockout of genes associated with late-onset AD also impacted behaviour in larvae, we then generated F0 knockouts for the zebrafish orthologues of APOE (apoea/apoeb), CD2AP (cd2ap), CLU (clu), and SORL1 (sorl1).

APOE shares 22–25% amino acid identity with its two zebrafish orthologues, apoea and apoeb (Fig. 4–supplement 2a,b). To make apoea/apoeb double F0 knockouts, we targeted each gene at two loci in separate exons with highly mutagenic CRISPR-Cas9 RNPs (Fig. 4–supplement 2c,d). Consistently across three clutches, double apoea/apoeb F0 knockout larvae performed slightly more subdued swimming bouts during the day than scrambled-injected siblings (active bout mean: −4%; active bout maximum: −5% vs. scrambled; Fig. 4–supplement 2e). This slight hypoactivity is unlikely to reflect a motor defect because at night, apoea/apoeb knockouts performed more swimming bouts (+31%) and had less sleep (−7%, Fig. 4–supplement 2e) than controls.

Knockout of other early-onset AD risk genes also impacted behaviour, especially at night. We mutated zebrafish cd2ap (44% identical amino acid sequence vs. human CD2AP; Fig. 4–supplement 3a), at three loci on distinct exons (Fig. 4–supplement 3b). While the two clutches of cd2ap F0 knockouts gave generally inconsistent results, knockout larvae of both clutches were more active (time active: +38%; total activity: +44% vs. scrambled) and slept less at night than control siblings (−13%), mainly because they performed more swimming bouts (+34%, Fig. 4–supplement 3c). Similarly, knockout of clu (39% identical amino acid sequence vs. human CLU, Fig. 4–supplement 4a) did not strongly affect behaviour during the day (Fig. 4–supplement 4c), despite a particularly high rate of frameshift mutations at all three targeted loci (Fig. 4–supplement 4b). At night, clu F0 knockouts were slightly more active (time active: +12% vs. scrambled) and slept less than control siblings (−7%, Fig. 4–supplement 4c). Finally, we generated F0 knockouts for sorl1 (63% identical amino acid sequence vs. human SORL1, Fig. 4a) by mutating three exons (Fig. 4b). sorl1 F0 knockout larvae were less active during the day but slept less at night (Fig. 4c and Fig. 4–supplement 1). Specifically, during the day (Fig. 4d), sorl1 F0 knockouts spent less time active than scrambled-injected larvae (−15% vs. scrambled), performing fewer swimming bouts (−15%) of approximately the same duration and intensity as controls. Sleep during the day was unaffected. At night (Fig. 4d), sorl1 F0 knockouts spent more time active than control siblings (+20%), largely because their swimming bouts tended to last longer (+6%). They also slept less (−5%), initiating fewer sleep bouts (−9%) of roughly the same duration.

In summary, we video-tracked the sleep/wake behaviours of F0 knockout larvae in four genes associated with late-onset AD: apoea/apoeb, cd2ap, clu, and sorl1. Remarkably, loss of all four genes produced a fairly consistent phenotype at night (Fig. 4–supplement 5), with all knockout larvae spending 5–13% less time asleep and instead spending 12–38% more time active because they were moving 13–34% more often. In contrast, each mutant had distinct day-time behavioural alterations. Comparing with early-onset genes, psen1 knockouts had similar night-time phenotypes, but loss of psen2 or appa/appb had no effect on night-time sleep. Therefore, at least some late-onset (and one early-onset) Alzheimer’s risk genes have common effects on sleep from an early age, despite being expressed in different tissues and having distinct biochemical properties.

Fig. 4:

Fig. 4–suppl. 1:

Fig. 4–suppl. 2:

Fig. 4–suppl. 3:

Fig. 4–suppl. 4:

Fig. 4–suppl. 5:

From behavioural fingerprint to causal process: serotonin signalling disruption by the loss of Sorl1

From genomic studies of AD, we know that mutations in genes such as SORL1 modify risk by disrupting some biological processes. Presumably, the same processes are disrupted in zebrafish sorl1 knockouts, and some caused the behavioural alterations we observed. Can we now follow the thread backwards and predict some of the biological processes in which Sorl1 is involved based on the behavioural profile of sorl1 knockouts?

To predict disrupted biological processes from the sorl1 knockout behavioural profile, we developed a behavioural pharmacology approach based on a database of 5,756 small molecule behavioural fingerprints (3,674 unique compounds) obtained in wild-type larvae⁴². First, we used information from the Therapeutic Target Database⁵⁶ to annotate each compound with its indications (e.g. triprolidine is used to treat hay fever), targets (e.g. triprolidine targets the histamine H1 receptor), and the pathways it affects through its targets (e.g. triprolidine affects the “inflammatory mediator regulation of TRP channels”). Second, we converted the frame-by-frame behavioural fingerprint of sorl1 knockouts to the one-minute format used by the database. We then compared the mean sorl1 knockout fingerprint with each small molecule behavioural fingerprint, creating a ranked list from the small molecule fingerprint most similar to the sorl1 fingerprint (SU6656: cosine similarity = 0.83) to the small molecule fingerprint most opposite to the sorl1 fingerprint (flumazenil: cos = −0.80). Third, we tested using a custom permutation test whether specific indication, target, or pathway annotations were significantly enriched at the top and/or the bottom of the ranked list. More present among the small molecules most (anti-)correlating with the sorl1 fingerprint were drugs used to treat depression (Fig. 5–supplement 1a, simulated p-value = 0.0129), targeting the serotonin transporter SLC6A4 (Fig. 5a,b, simulated p-value = 0.0112), and affecting the “serotonergic synapse” pathway (Fig. 5– supplement 1b, simulated p-value = 0.0003). Thus, sorl1 knockout larvae behaved similarly to larvae treated with small molecules targeting serotonin signalling, suggesting that the loss of Sorl1 disrupted serotonin signalling.

If serotonin signalling is disrupted in sorl1 knockouts, they should react differently to serotonergic drugs than wild-type animals. To test this hypothesis, we treated sorl1 F0 knockouts and controls with citalopram, a selective serotonin reuptake inhibitor used to treat depression⁵⁷. Citalopram binds the serotonin transporter SLC6A4, preventing the reuptake of serotonin from the synaptic cleft. At a low dose of citalopram (1 µM, added directly to the fish water), sorl1 F0 knockouts and scrambled-injected controls reacted similarly, sleeping about 1.1 hr more during both day and night (Fig. 5c). In contrast, at a higher dose of citalopram (10 µM) sorl1 F0 knockouts had a stronger reaction than their control siblings. For example, sleep during the day increased 2.5× in sorl1 knockouts, while it only increased 2.2× in controls. The heightened sensitivity of sorl1 knockouts to citalopram was also apparent when taking all parameters into account by measuring each larva’s Euclidean distance from the average H2O-treated sibling of same genotype (Fig. 5d). Indeed, 10 µM citalopram displaced the sorl1 knockout larvae further from their behavioural baseline than it pushed controls from theirs. Thus, our behavioural pharmacology approach correctly predicted from behaviour alone that serotonin signalling was disrupted in sorl1 knockouts.

There are at least two ways sorl1 knockouts could be more sensitive to citalopram. First, compared to wild types, sorl1 knockouts could undergo a larger spike of serotonin in the synaptic cleft when reuptake is blocked. This may be because sorl1 knockouts synthesise more serotonin, either because they have more serotonergic neurons or a higher expression of the enzymes required to make serotonin, or because they produce less of the serotonin transporter, making them more sensitive to citalopram as a given dose would inhibit a larger proportion of the transporter than in wild-type animals. Second, sorl1 knockouts may be overly sensitive to serotonin itself because post-synaptic neurons have higher levels of serotonin receptors.

To distinguish between these two hypotheses, we used HCR to label serotonergic neurons by tagging mRNA coding for tryptophan hydroxylases (in zebrafish: tph1a, tph1b, tph2; Fig. 5e), an enzyme required for the synthesis of serotonin, and serotonin transporters by tagging slc6a4a and slc6a4b mRNA⁵⁸ (Fig. 5–supplement 1c). We registered the brain stacks to a common atlas (Zebrafish Brain Browser) and segmented them into 168 anatomical regions. In each region with expression, we counted the number of positive voxels and measured the total signal intensity in each channel (channel 1: tph1a + tph1b + tph2; channel 2: slc6a4a + slc6a4b). sorl1 F0 knockouts had slightly more (+2%) slc6a4a/b positive voxels than controls across regions, but this difference was not definite for any one region and not replicated when looking at total signal intensity (Fig. 5–supplement 1d). sorl1 F0 knockouts had on average 18% lower tph1a/1b/2 signal intensities across regions, particularly in the hypothalamus (hypothalamus – caudal zone 1: –24%, Fig. 5e,f). Therefore, although we see small differences in the levels of serotonin transporters or synthesis enzymes, these were opposite to our hypotheses so do not readily explain the heightened sensitivity of sorl1 to citalopram. Additionally, at least one study did not find changes in serotonin levels in the striatum of SORL1 knockout mice⁵⁹, and we did not find genes related to serotonin signalling to be differentially expressed in a bulk RNA-seq dataset of sorl1 knockout zebrafish brains⁶⁰. Ruling out the first hypothesis, sorl1 knockouts may react excessively to a given spike in serotonin.

Fig. 5:

Fig. 5–suppl. 1:

From behavioural fingerprint to candidate therapeutic: betamethasone rescues the psen2 behavioural phenotype

In addition to pointing to the disrupted pathways, behavioural pharmacology can also predict small molecules that may normalise mutant behavioural phenotypes by pointing to compounds that generate the opposite behavioural fingerprint in wild-type larvae.

Using this approach, we sought to identify compounds capable of normalising the behavioural alterations of psen2 knockouts, such as the increased day-time sleep. We ranked the 5,756 small molecule behavioural fingerprints in comparison with the mean psen2 F0 knockout fingerprint and focused on compounds that generated the most opposite (anti-correlating) fingerprint when applied on wild-type larvae. This identified three compounds with a high negative cosine compared to psen2 knockouts (Fig. 6a and Fig. 6– supplement 1a): tinidazole (minimum cos = −0.85), fenoprofen (minimum cos = −0.74), and betamethasone (minimum cos = −0.74). These compounds were also selected because they were each replicated at least once in the database, lending confidence to the prediction. We then applied the three compounds on psen2 F0 knockouts and measured their effects on sleep/wake parameters. For each treatment, we grouped the behavioural parameters into four categories: “rescue” if the parameter was significantly altered in DMSO-treated psen2 knockout larvae but got normalised by the drug; “missed rescue” if the psen2 knockout parameter remained altered after drug treatment; “side effect” if a parameter was unaffected in psen2 knockouts but got altered by the drug; and “no effect” if the parameter was unaffected in psen2 knockouts and remained unchanged by the drug.

All three compounds normalised at least some aspects of the psen2 knockout behavioural phenotype. Tinidazole did not reduce the abnormally high day-time sleep characteristic of psen2 knockouts (Fig. 6c left). It rescued the swimming bout alterations (active bout parameters) during the day but not at night. Overall, the compound barely caused any side effect but only rescued a small aspect of the psen2 knockout phenotype (9/18 altered parameters rescued). Fenoprofen performed worse (Fig. 6c middle, 8/18 altered parameters rescued), rescuing some of the swimming bout alterations during the day but aggravating them at night. Furthermore, fenoprofen made psen2 knockout larvae drowsier than they were during the day, causing sleep to increase further and activity to decrease. Strikingly, betamethasone completely resolved the excess day-time sleep without causing hyperactivity (Fig. 6b and Fig. 6–supplement 1b). Betamethasone rescued most, but not all, of the swimming bout alterations during both the day and night (Fig. 6c right, 14/18 altered parameters rescued). Betamethasone did cause a few side effects, likely by making psen2 knockout larvae overly aroused at the start of the night, extending sleep latency and decreasing the slope in activity and the startle response. Therefore, of the three drugs we selected using behavioural pharmacology, one almost completely normalised the psen2 knockout phenotype, albeit with a few side effects.

Fig. 6:

Fig. 6–suppl. 1:

Discussion

The F0 knockout and behavioural pharmacology approach was able to successfully predict a signalling pathway that was disrupted by knockout of the Alzheimer’s risk gene sorl1 and identify a drug capable of normalising the psen2 knockout behavioural phenotype. To allow researchers to generate pharmacological predictions from their own sleep/wake datasets, we built an online app (francoiskroll.shinyapps.io/zoltar/) that can plot behavioural fingerprints, rank all 5,756 small molecule fingerprints in the database in comparison to the query fingerprint, and perform permutation tests for enrichment of indications, drug targets, and KEGG pathways (Fig. 7–supplement 1a,b). While we used genes associated with AD as a case study, predictive behavioural pharmacology provides a generalisable framework that can be applied to any set of disease risk genes.

Disrupted sleep and serotonin signalling—causal processes of Alzheimer’s disease?

Of the seven genes tested, psen1 and all four late-onset Alzheimer’s risk genes tested decreased sleep duration at night when mutated in zebrafish larvae. Could disrupted sleep itself be a causal process in AD? AD patients often take longer to fall asleep, have higher sleep fragmentation, and spend sharply less time in NREM sleep^61,62. Sleep disruption can be present before the onset of cognitive deficits. For example, very high sleep fragmentation in elderly people is associated with subsequent diagnosis of AD⁶³. Experiments on humans and animal models also point to disrupted sleep being a causal process in some cases. In healthy adults, one night of sleep deprivation was sufficient to cause a 25–30% increase in Aβ signal on PET scans⁶⁴, a ∼30% increase in CSF concentrations of Aβ⁶⁵, and a ∼50% increase in CSF concentrations of tau⁶⁶. In wild-type rats and mice overexpressing mutated human APP and PSEN1, restricting sleep for 21 days increased Aβ deposits in the cortex^67,68. Conversely, pharmaceutically or chemogenetically consolidating sleep in AD mouse models for 1–2 months delayed Aβ plaque formation^67,69. Disrupted sleep is therefore likely to be a causal process in AD. Can it become a therapeutic target? An ongoing clinical trial (NCT04629547) is underway to test whether suvorexant, an orexin receptor antagonist that increases sleep, can reduce Aβ accumulation by consolidating sleep in older adults without dementia⁷⁰. Our observation that disruption of genes associated with AD diagnosis after 65 years reduces sleep in 7-day zebrafish larvae suggest that disrupted sleep may be a common mechanism through which these genes exert an effect on risk. Impaired sleep early in life may be especially deleterious as it is likely essential for brain development: infants spend most of their time asleep during the first year of life⁷¹. Incidentally, infants who are APOE ε4 carriers show differences in grey matter volume and myelin content in multiple brain regions⁷². Differences are also present in children 9–17 years old who carry the disease-causing PSEN1 E280A mutation⁷³. Future work should assess whether sleep is impaired in infants at higher genetic risk of AD and whether it is related to these structural differences.

Fig. 7–suppl. 1:

Our behavioural pharmacology approach also predicted that loss of Sorl1 impaired serotonin signalling, which was confirmed by increased sensitivity of sorl1 knockouts to citalopram. SORL1 acts as an adaptor protein between retromer and several cargo proteins, such as APP⁷⁴, GDNF receptor α1⁵⁹, and glutamate receptor 1 (GLUA1)⁷⁵. Retromer is a large protein complex which ‘rescues’ proteins from endosomes, targeting them to recycling instead of lysosomal digestion⁷⁶. By disrupting retromer function, mutations in SORL1 disrupt endosomal recycling, which appears as swelling of early endosomes⁷⁷. As a result, APP remains in endosomes for longer, where it is more likely to be cleaved by β-secretase, the first step towards production of Aβ⁷⁸. Similarly, retromer dysfunction also lowers the level of GLUA1 receptors on the neuron membrane, likely because endocytosed receptors are not recycled back to the membrane, which can result in both increased or decreased firing rates in vitro⁷⁵. We speculate that some serotonin (5-hydroxytryptamine, 5-HT) receptors are also recycled via retromer and SORL1. Consequently, sorl1 knockouts would react differently to a large spike in serotonin (citalopram treatment) as they have different levels of 5-HT receptors on the synaptic membrane. In support for this idea, both 5-HT type 4 receptor⁷⁹ and SORL1⁸⁰ interact with sorting nexin 27, a subunit of retromer. Furthermore, the 5-HT receptor type 2 was also an enriched target for sorl1 (simulated p-value = 0.046). Other observations are consistent with serotonin signalling being a causal process of AD. For example, treatment of an AD mouse model with the serotonin reuptake inhibitor (SSRI) citalopram caused a rapid decrease in Aβ concentrations in the ISF and eventually reduced amyloid plaque load. Additionally, humans treated with SSRIs within the past 5 years had significantly lower cortical amyloid plaque load on PET scans, suggesting a causal relationship⁸¹. While much work remains to be done, SSRIs can generally be used continuously with no adverse effects, so modulating serotonin signalling may be an attractive approach to reduce Aβ levels for years before the onset of cognitive symptoms⁸¹.

Limitations and future directions for predictive behavioural pharmacology

Currently, the hypotheses that our behavioural pharmacology approach can test are limited by the small molecule screen dataset⁴², which was biased towards specific themes and targets. For example, of 1,551 compounds with an annotated target, 60 (3.8%) target the dopamine D2 receptor but only 1 (0.06%) targets mitochondrial uncoupling proteins (UCP1, 2, 3). Consequently, the approach presumably has high sensitivity for disruptions in dopamine signalling, while some hypotheses, such as impairment in mitochondrial coupling, are never formally tested. Broadly, this means that a negative (non-significant) result may not be meaningful if the biological process in question is underrepresented. Additionally, about 30% of the annotated compounds in the database have more than one target. While this is not necessarily an issue when the goal is to find any small molecule which rescues the phenotype, this lack of specificity likely blurs the signal from modulation of a specific target to effect on behaviour. For example, all five compounds which target the histamine H4 receptor also target the histamine H3 receptor, so if the H3 receptor is detected as significantly enriched, the H4 receptor is certain to be significant too, regardless of whether the H4 receptor is indeed specifically disrupted. To develop the biological process prediction, future work should focus on selecting specific agonists and antagonists for a set list of targets, such as the compound VUF-6002 to selectively target the histamine H4 receptor without also activating H3⁸².

A way to improve the specificity of the approach is to enhance the resolution of the behaviour tracking so that targets which give similar but unique effects on behaviour can be distinguished. We introduced the FramebyFrame R package to enable analysis at sub-second resolution, but the original small molecule screen dataset was collected at the one-minute resolution⁴², so comparing a new behavioural profile to the small molecule database first involves losing resolution to make the fingerprints comparable. During a new small molecule screen, one could use SLEAP⁸³ or DeepLabCut⁸⁴ to extract additional behavioural parameters from video recordings such as speed or position in the well. A faster and higher definition camera could be introduced to allow fine tracking of tail and eye movements. Exposing the larvae to various stimuli is also a way to increase the number of parameters while avoiding redundancy⁸⁵. With more behavioural parameters, the prediction of therapeutics using behavioural pharmacology would also be more precise and the detection of possible behavioural side effects more thorough. Analytically, an improvement could be to represent larval behaviour as a sequence of movements, rather than parameter averages. For example, the Δ pixel time course from our experiments can be encoded as a sequence of bout types termed modules, themselves organised in motifs which are 2–20-module long⁴⁴. Alternatively, behaviour could be represented as sequences of known types of swimming bouts⁸⁶ or automatically discovered ‘syllables’⁸⁷. In mouse, this sequence representation was better than a fingerprint one at differentiating similar compounds or even doses⁸⁷, so should also improve the predictive power of the behavioural pharmacology approach.

This leads to a thought experiment: assuming absolute resolution in behaviour tracking, does every knockout give a unique effect on behaviour? In other words, are there ∼26,000 distinct knockout fingerprints or do multiple genes converge on a smaller set of possible fingerprints? How much resolution is required to distinguish every possible fingerprint?

Conclusion

In summary, we present a behavioural pharmacology approach which uses the sleep/wake behaviour of zebrafish larvae as a tool to translate genomic findings into druggable biological processes and candidate therapeutics. We used Alzheimer’s risk genes, especially SORL1 and PSEN2, as case studies and correctly predicted mutant-drug interactions from behaviour alone. Our strategy is both scalable thanks to the F0 knockout method and the FramebyFrame analysis package and generalisable beyond AD through the ZOLTAR app. Other neurological conditions for which many risk genes are known, such as autism, epilepsy, or schizophrenia, are prime candidates for the application of the strategy.

Methods

Animals

Adult zebrafish were reared by University College London’s Fish Facility on a 14 hr:10 hr light:dark cycle. To obtain eggs, pairs of one female and one male were isolated in breeding boxes overnight, separated by a divider. Around 9 AM (lights on) the next day, the dividers were removed and eggs were collected 7–10 min later. The embryos were then raised in 10-cm Petri dishes filled with fish water (0.3 g/L Instant Ocean) in a 28.5°C incubator on a 14 hr:10 hr light:dark cycle. Debris and dead or dysmorphic embryos were removed every other day with a Pasteur pipette under a bright-field microscope and the fish water replaced. At the end of the experiments, larvae were euthanised with an overdose of 2-phenoxyethanol (ACROS Organics). Experimental procedures were in accordance with the Animals (Scientific Procedures) Act 1986 under Home Office project licences PA8D4D0E5 and PP6325955 awarded to Jason Rihel. Adult zebrafish were kept according to FELASA guidelines⁸⁸.

Wild types refer to AB × Tup LF fish. Throughout, F0 knockouts refer to wild-type embryos that were injected with gRNA/Cas9 RNPs at the single-cell stage.

Generation of F0 knockout larvae

crRNA selection

The crRNA was the only component of the Cas9/gRNA RNP specific to the target locus.

In previous work¹⁹, we selected predesigned crRNAs from the Integrated DNA Technologies (IDT) database (eu.idtdna.com) based on predicted on-target and off-target scores calculated by a proprietary IDT algorithm. Subsequent benchmarking in zebrafish of various crRNA design algorithms revealed that the IDT on-target scores did not predict mutagenesis rates in vivo⁸⁹. All crRNAs used in F0 knockout experiments were therefore designed using CHOPCHOP (chopchop.cbu.uib.no)⁹⁰ implementing both the on-target score calculation by CRISPRScan⁹¹, which benchmarking determined was the best predictor of mutagenesis⁸⁹, and inDelphi⁹², which predicts to some extent the generated mutations in zebrafish embryos⁹³.

We prioritised targeting asymmetrical exons and only selected crRNAs without any off-targets with 0, 1, 2, or 3 mismatches, as off-targets with up to 3 mismatches may undergo mutation^19,94. We then selected the crRNAs based on their on-target score and the prediction of frameshift mutations by inDelphi (crRNAs used had 82.8 ± 5.7% predicted frameshift mutation rate). We also used SNPfisher, an online database of common single-nucleotide polymorphisms (SNP) in zebrafish wild-type strains, to check that putative target sites did not contain any SNP⁹⁵.

A protocol describing how to select crRNAs based on these developments is available at dx.doi.org/10.17504/protocols.io.81wgb6r5qlpk/v1. Sequences of the crRNAs and information about the targeted loci are provided in Supplementary file 1.

Cas9/gRNA preparation

A protocol describing how to generate F0 knockout larvae is available at dx.doi.org/10.17504/protocols.io.5qpvo52wdl4o/v3.

The CRISPR-Cas9 RNP was made of three components bought from IDT: the crRNA (Alt-R® CRISPR-Cas9 crRNA) and tracrRNA (Alt-R® CRISPR-Cas9 tracrRNA), together forming the gRNA, and the Cas9 (Alt-R® S.p. Cas9 Nuclease V3). The crRNA and tracrRNA were received as pellets, which were individually resuspended in Duplex buffer (IDT, received with the tracrRNA) to form 200 μM stocks. Stocks of crRNA and tracrRNA were stored at −70°C before use. Cas9 was stored at –20°C before use. Each crRNA was annealed separately with the tracrRNA by mixing 1 μL crRNA 200 μM; 1 μL tracrRNA 200 μM; 1.28 μL Duplex buffer. The mix was heated to 95°C for 5 min, then cooled on ice, to obtain a 61 μM gRNA solution. The gRNA solutions were then mixed in equal volumes with Cas9 (1 μL gRNA 61 μM; 1 μL Cas9 61 μM directly from the IDT vial), incubated at 37°C for 5 min then cooled on ice, generating three 30.5 μM RNP solutions. The three RNP solutions were pooled; the final concentration of each RNP in the pool was thus 10.2 μM and the total RNP concentration 30.5 μM.

The RNPs were usually kept overnight in a 4°C fridge on ice before injections the following day. Some experiments used RNPs stored at –70°C for a few weeks.

For experiments targeting two genes simultaneously (appa/appb, psen1/psen2, apoea/apoeb), we targeted each gene at two loci to potentially reduce unviability¹⁹. While mutating two loci instead of three is predicted to reduce rates of complete biallelic knockout animals, given the high rate of frameshift mutations achieved with the RNPs that were used, ∼90% of injected animals were still expected to be complete biallelic knockouts. Across loci that were mutated during double F0 knockout experiments, 97.5 ± 7.0% of the reads were mutated, and 78.9 ± 24.0% of all reads had a frameshift mutation; see francoiskroll.shinyapps.io/frameshiftmodel/ for the theoretical knockout rate prediction. The preparation followed the same steps as above. The four RNP solutions were pooled, so the final concentration of each RNP in the pool was 7.6 µM and the total RNP concentration was 30.5 µM.

Injections

Approximately 1 nL of the three-RNP pool was injected into the yolk at the single-cell stage before cell inflation. This amounts to ∼30.5 fmol of RNP (30.5 fmol [5029 pg] of Cas9 and 30.5 fmol [1070 pg] of total gRNA). Each unique RNP was present in equal amounts in the pool. Therefore, in the case of three RNPs, ∼10.2 fmol of each RNP were co-injected.

When targeting two genes simultaneously, approximately 1.3 nL of the four-RNP mix were injected so the amount of RNP per gene would remain equal to when a single gene is targeted at three loci.

The injected eggs were kept at ∼20°C for ∼50 min after injection before transfer to a 28.5°C incubator, as delaying the first cell division by lowering the temperature was tentatively shown to increase mutagenesis and reduce the diversity of alleles⁹⁶.

Scrambled RNPs

In all F0 knockout experiments, three or four non-targeting, ‘scrambled’, crRNAs were prepared into RNPs and injected following the same steps as above. These were created by shuffling the spacer sequence of existing crRNAs until the only predicted targets had 4 or more mismatches in protein-coding sequences, or 3 or more in non-coding sequences. Sequences of the scrambled crRNAs are available in Supplementary file 1.

Preparation of samples for Illumina MiSeq

We sequenced every targeted locus in samples of F0 knockout larvae using Illumina MiSeq. For psen1, psen2, apoea, and apoeb, the samples were from video-tracked larvae. For appa, appb, cd2ap, clu, and sorl1, we generated F0 knockout larvae expressly to be sequenced. For each locus, we prepared 5–6 F0 knockout samples and 2–3 scrambled-injected samples. Individual larvae were sequenced at the three or four mutated loci.

The larvae were anaesthetised and their genomic DNA extracted by HotSHOT⁹⁷ as follows. Individual larvae were transferred to a 96-well PCR plate or strips of 12 tubes. Excess liquid was removed from each well before adding 50 μl of 1× base solution (25 mM KOH, 0.2 mM EDTA in water). Plates were sealed and incubated at 95°C for 30 min then cooled to room temperature before the addition of 50 μL of 1× neutralisation solution (40 mM Tris-HCl in water). Genomic DNA was then stored at –20°C.

PCR primers were designed for each target locus using Primer-BLAST (NCBI) to amplify a window of 150–200 bp with at least 30 bp between each primer binding site and the predicted double-strand break site, as this is where most deletions are found¹⁹. PCR primers were ordered with a Nextera overhang at the 5′-end of each primer to allow indexing (see Supplementary file 1).

Each PCR well contained: 7.98 μL PCR mix (2 mM MgCl2, 14 mM pH 8.4 Tris-HCl, 68 mM KCl, 0.14% gelatine in water, autoclaved for 20 min, cooled to room temperature, chilled on ice, then added 1.8% 100 mg/mL BSA and 0.14% 100 mM d[A, C, G, T]TP), 3 μL 5× Phusion HF buffer (New England Biolabs), 2.7 μL dH2O, 0.3 μL forward primer (100 μM), 0.3 μL reverse primer (100 μM), 0.12 μL Phusion High-Fidelity DNA Polymerase (New England Biolabs), 1.0 μL genomic DNA; for a total of 15.4 μL. The PCR plate was sealed and placed into a thermocycler. The PCR program was: 95°C – 5 min, then 40 cycles of: 95°C – 30 sec, 60°C – 30 sec, 72°C – 30 sec, then 72°C – 10 min, then cooled to 10°C until collection. The PCR product’s concentration was quantified with Qubit (dsDNA High Sensitivity or Broad Range Assay) and its length was verified on a 2.5% agarose gel with GelRed (Biotium). Excess primers and dNTPs were removed by ExoSAP-IT (ThermoFisher) following the manufacturer’s instructions. The samples were then sent for Illumina MiSeq, which used MiSeq Reagent Nano Kit v2 (300 Cycles).

Illumina MiSeq data analysis

Illumina MiSeq data was received as two fastq files for each well, one forward and one reverse. The paired-end reads were aligned to the reference amplicon with bwa v0.7.17 and the resulting bam alignment file was sorted and indexed with samtools v1.11⁹⁸. Alignments were then filtered to keep only reads with less than 20% of its length soft-clipped and spanning at least 20 bp on each side of the predicted Cas9 double-strand break site, 4 bp upstream of the ‘N’ of the NGG protospacer adjacent motif. Whenever necessary, bam alignment files were visualised with IGV v2.16.0. The resulting filtered bam file was converted back to a forward and a reverse fastq file using bedtools v2.30.0⁹⁹. The filtered fastq files were used as input to the R package ampliCan v1.20.0¹⁰⁰, together with a csv configuration file containing metadata information about the samples. AmpliCan was run with settings min_freq = 0.005 (any mutation at a frequency below this threshold was considered as a sequencing error), cut_buffer = 12 (any insertion/deletion starting within ±12 bp of the PAM may be Cas9-generated), and event_filter = FALSE (no filtering of reads), as the reads were already filtered. AmpliCan can normalise mutation counts by ignoring any insertion/deletion found in control samples. On rare occasions, this could artefactually decrease mutation counts in F0 knockout samples because of low-level contamination in control samples (e.g. Fig. 4b, locus 1, sample scr1) so we turned this feature off. Instead, we manually checked the scrambled-injected samples in IGV to confirm that there were no insertions or deletions already present in the wild-type background. AmpliCan detected and counted mutations in the reads and wrote results files that were used for subsequent analysis. Any sample with less than 20× paired-end (40× single-read) coverage were excluded from plots and subsequent analysis. Figures like Fig. 3c,d plot the proportion of mutated reads and the proportion of reads with a frameshift mutation at each locus, as computed by ampliCan. If a read contained multiple indels, ampliCan summed them to conclude whether the read had a frameshift mutation or not.

3D-printed 96-well mesh-bottom plate

The 96-square well mesh-bottom plate (Fig. 2a) was designed using Fusion 360 and PrusaSlicer. Multiple copies were 3D-printed in clear polylactic acid (PLA) with a Creality Ender-3 3D printer. The 3D model is available at github.com/francoiskroll/FramebyFrame.

Behavioural video-tracking

In F0 knockout experiments, wild-type embryos from separate clutches were injected at the single-cell stage with CRISPR-Cas9 RNPs. Each clutch was from a unique pair of parents allowed to mate for 7–10 min. For each clutch, about half of the embryos were injected with non-targeting ‘scrambled’ RNPs to generate control siblings. Clutch-to-clutch variability in locomotor activity is substantial in zebrafish larvae, even between wild-type clutches of the same strain¹⁰¹. Therefore, keeping clutches from different parents/mating events separate is predicted to increase sensitivity of the assay by reducing variability. More generally, it is crucial never to perform comparisons where treatment larvae (e.g. mutant, transgenic, drug-treated) are from one clutch and the control larvae are from another (i.e. controls are not siblings), as any difference discovered with such an experimental design could be caused by clutch-to-clutch variability.

At 5 dpf, individual larvae were transferred to the wells of 3D-printed mesh-bottom plates (see 3D-printed 96-well mesh-bottom plate), each sitting in the water bath of a ZebraBox (ViewPoint Behavior Technology). To avoid any potential localisation bias during the tracking, F0 knockout and scrambled-injected larvae were plated in alternating columns of the 96-well plate. From each well, the video-tracking software (ZebraLab, ViewPoint Behavior Technology) recorded the number of pixels that changed intensity between successive frames. To be counted, a pixel must have changed grey value above a sensitivity threshold, which was set at 20. The metric, termed Δ pixel, describes each animal’s behaviour over time as a sequence of zeros and positive values, denoting if the larva was still or moving. In ZebraLab, the freeze setting was set to 3 and the burst setting was set to 200. However, note that these settings determine how the Δ pixel data is summarised into number of seconds active per minute by the software (middur parameter, see One-minute vs. frame-by-frame); therefore, their values are irrelevant when analysing the data using the FramebyFrame package. Tracking was performed at 25 frames per second on a 14 hr:10 hr light:dark cycle for ∼65 hours, generating sequences of roughly 5,850,000 Δ pixel values per animal. During the experiment, a pump on a timer (Kollea Automatic Watering System, Amazon UK) injected fish water from a reservoir to the bath every morning around 9 AM (lights on) to counteract water evaporation. Water could overflow through an exit pipe, whose height was adjusted to set the water level in the bath. The day light level was calibrated at 555 lux with a RS PRO RS-92 light meter (RS Components). Night was in complete darkness with infrared lighting for video recording. Temperature and light transitions were recorded with a HOBO Pendant (Onset Data Loggers) immersed in the water bath. Temperature throughout the experiment was 23–26°C.

At the end of the tracking, we inspected the larvae under a bright-field microscope and excluded from analysis any larva that did not appear healthy or was not responsive to a light touch with a P10 tip. We then randomly selected 2–3 scrambled-injected and 5–6 F0 knockout larvae for sequencing of the targeted loci (see Preparation of samples for Illumina MiSeq).

Behavioural video-tracking with drug treatment

This refers to the experiment treating sorl1 F0 knockouts with citalopram (Fig. 5c,d) and the experiment treating psen2 F0 knockouts with tinidazole, fenoprofen, and betamethasone (Fig. 6 and Fig. 6–supplement 1b). Broadly, video-tracking was performed as above (see Behavioural video-tracking) but larvae were housed in standard 96-square well plates (Whatman) in fish water with drug. Water was topped up every morning shortly after 9 AM (lights on).

Citalopram hydrobromide (Fisher Scientific, #15732987) was stored at room temperature before use. We chose the treatment concentrations (1 µM and 10 µM) based on previous work in zebrafish larvae¹⁰² that found that the minimum dose which had a discernible effect on locomotor activity was 373 µg/L (0.92 µM). A 500× stock (5 mM) for the 10 µM treatment was prepared by diluting 0.0203 g in 10 mL dH2O (molecular weight = 405.3 g/mol). The 5 mM solution was then diluted 1:10 (100 µL citalopram 5 mM + 900 µL dH2O) to obtain a 500× stock (500 µM) for the 1 µM treatment. Using a P1000 pipet with a tip whose end was cut-off, individual 5-dpf larvae were transferred in 650 µL fish water to the wells of clear 96-square well plates (Whatman). 1.3 µL of each 500× stock was added on top of each well, effectively diluting each stock 1:500 (5 mM stock diluted to 1 µM and 500 µM stock diluted to 10 µM). Control wells were topped-up with 1.3 µL dH2O.

Tinidazole (Fisher Scientific, #16594384) was stored at room temperature before use. We chose the treatment concentration (30 µM) based on a preliminary experiment where we video-tracked wild-type larvae treated with 10 µM, 30 µM, 100 µM tinidazole (data not shown). A 50 mM solution was prepared by diluting 0.0247 g in 2 mL DMSO (molecular weight = 247.27 g/mol), then diluted this solution to 15 mM in DMSO as 500× stock.

Fenoprofen calcium salt hydrate (Sigma-Aldrich, #F1517) was stored at room temperature before use. We chose 15 µM as the treatment concentration based on the results from the Rihel et al., 2010 small molecule behavioural screen database⁴². A 200 mM solution was prepared by diluting 0.1114 g in 1065 µL DMSO (molecular weight = 522.6 g/mol), then diluting this solution to 7.5 mM in DMSO as 500× stock.

Betamethasone (Cayman Chemical, #20363) was stored at –20°C before use. We chose 15 µM as the treatment concentration based on the results from the Rihel et al., 2010 small molecule behavioural screen database. A 33 mM solution was prepared by diluting 0.0204 g in 1,575 µL DMSO (molecular weight = 393.5 g/mol), then diluting this solution to 7.5 mM in DMSO as 500× stock.

For tinidazole, fenoprofen, and betamethasone, 100 µL of each 500× stock was then mixed in 20 mL fish water in a Falcon, and the solution transferred to one Petri dish per drug. A Petri dish for control larvae was prepared by mixing 100 µL of DMSO in 20 mL fish water. Using a P1000 pipet with a tip whose end was cut-off, individual 5-dpf larvae were transferred in 500 µL fish water to the Petri dishes. Each Petri dish was then topped-up with fish water to 50 mL final volume, which effectively diluted each drug to the final treatment concentration (100 µL 500× stock in 50 mL fish water). The final DMSO concentration was 0.002% (100 µL 100% DMSO in 50 mL fish water). Using a P1000 pipet with a tip whose end was cut-off, individual larvae were then transferred in 650 µL to the 96-square well plates. This process was to avoid adding drug in DMSO directly to the well as the solution sinks in water and may not mix properly.

Video-tracking was then performed with the same settings as above (see Behavioural video-tracking). Both mornings shortly after 9 AM (lights on), the wells were manually topped-up with fish water (no drug), which assumes that the drug does not evaporate and was stable in fish water at 23–26°C for a few days.

A protocol for behavioural video-tracking with drug treatment is available at dx.doi.org/10.17504/protocols.io.4r3l27p6pg1y/v1.

Behavioural data analysis

Behavioural data analysis was performed using the FramebyFrame R package v0.11.0. A tutorial and documentation are available at github.com/francoiskroll/FramebyFrame. The important steps of the analysis are summarised below.

The raw file generated by the ZebraLab software was exported into thousands of xls files each containing 1 million rows of data. Each row of data represented the Δ pixel of one larva at one frame transition. Using the vpSorter(…) function, these data were re-organised in a large csv file where each column was a well, each row was a frame transition, and each cell a Δ pixel value. To visualise activity over time, we smoothed the Δ pixel time course for each larva with a 60-min (∼45,000 rows) rolling average and binned the data by summing Δ pixels in 10-min epochs. We generated the activity trace of individual larvae using ggActivityTraceGrid(…) and excluded from subsequent analysis any larva that had an obviously aberrant behaviour, defined as having missed entirely a day-night transition (no jump/drop in the activity trace) or having shown activity during a day lower than the previous night. Each well was assigned to a group using a metadata file generated by genotypeGenerator(…). For each 10-min epoch, we calculated the mean and standard error of the mean (SEM) across larvae in each group to build the activity traces (e.g. Fig. 3g left).

Sleep was measured on the large csv file containing all the frame-by-frame Δ pixel data. A 60-second (∼1,500 rows) rolling sum was applied on the Δ pixel time course of each larva, so that any data point that became zero indicated that the previous 60 seconds of Δ pixels were all zeros, that is, a sleep bout started exactly 60 seconds ago. The data after the rolling sum was converted into a series of Booleans: true if the data point was 0 (i.e. that frame was part of a sleep bout); false if the data point was positive (i.e. that frame was not part of a sleep bout, as the larva moved at some point in the preceding 60 seconds). At this stage, the first 0 of each sleep bout marked the frame at the 60^th second of each sleep bout. Therefore, a correction was then applied to extend each sleep bout 60 seconds in the past by switching the ∼1,500 frames of false before each sleep bout start to true. The time spent asleep in 10-min epochs (∼45,000 rows) was calculated by counting the number of true frames in each epoch and multiplying the counts by 1/frame rate (typically 1/25 frames per second), which was converted into minutes for the sleep trace. For example, if larva #3 had 10,000 out of 45,000 frames marked true (i.e. these frames are part of a sleep bout) from minute 20 to 30 of tracking, that represented 10,000 × ¹/25 ×¹/60 = 6.67 minutes asleep for that epoch. As for the activity trace, the mean and SEM across larvae in each group were calculated for each 10-min epoch to build the sleep traces (e.g. Fig. 3g right).

15 behavioural parameters were then calculated for each larva and day/night. Two additional parameters, startle response at sunset and sleep latency, were only defined for nights, for a total of 32 unique parameters (15 day parameters and 17 night parameters). Full days/nights during the experiment were: night0, day1, night1, day2, night2. Larvae were 5 dpf during night0, 6 dpf during day1/night1, 7 dpf during day2/night2, and the experiment was stopped in the morning of 8 dpf. Night0 (5 dpf) was excluded from the analysis as a habituation period; therefore, two day data points and two night data points were calculated for each larva and behavioural parameter. Definitions of the behavioural parameters can be found in the documentation of the FramebyFrame package (github.com/francoiskroll/FramebyFrame).

To build the behavioural fingerprints, a z-score was calculated for each day/night, each larva, and each unique parameter. For example, to calculate the z-score for parameter active bout length of larva #5 for day1, we first calculated the mean and standard deviation of all active bout length data points for control larvae during day1. These were then used to calculate larva #5’s z-score as:

Where x was larva #5’s day1 data point, µ%&’ was the mean of control day1 data points, σ%&’was the standard deviation of control day1 data points. The calculation was repeated for day2. The day1 and day2 z-scores were then averaged to produce a common day z-score for active bout length of larva #5. The process was repeated for each unique parameter, producing 32 z-scores for each larva, which we term each larva’s behavioural fingerprint. To draw the fingerprint plots (e.g. Fig. 3–supplement 1c), the mean and SEM of all z-scores were calculated for each unique parameter and group of larvae. By definition of the z-score, the mean z-score for the control larvae was always 0 and the SEM constant for each experiment (dependent on the number of control larvae), so the controls’ fingerprint was omitted from the plot.

In Fig. 6c, for each parameter, the mean and SEM of all z-scores from each clutch were calculated separately. The two z-score means and SEMs were then averaged to draw each bar. See Statistics on behavioural parameters for the definition of each category.

Each larva’s fingerprint can be conceptualised as the coordinates of one data point in a 32-dimension space where each dimension represents one behavioural parameter. In Fig. 5d, the centre of the H2O-treated scrambled-injected larvae or the centre of the H2O-treated sorl1 F0 knockout larvae was set as the origin of this 32-dimension space (point at coordinates 0, 0, 0, …) by the z-scoring procedure. We then measured the Euclidean distance between each larva’s fingerprint and the origin.

Statistics on behavioural parameters

Each behavioural parameter was statistically compared between F0 knockout and scrambled-injected larvae using linear mixed effects (LME) modelling implemented in the lmer function of the R package lme4 v1.1.31¹⁰³. The data were parameter values per larva per time window, for example larva #3 slept 4 hr during night2. We always tracked one unique clutch in each ZebraBox run; therefore, the replicate experiment(s) provided both technical and biological replication. The data collected during nights or during days were analysed separately. The fixed effect was the group assignment (genotype or treatment). Random effects were intercepts for clutch assignment (experiment), larva number, and the larva’s age (dpf). Random effects of clutch assignment and larva number were modelled as nested, as data points between larvae of the same clutch were expected to be more similar than between larvae of different clutches; and within each clutch, the day 1 (or night 1) data point of one larva was expected to be similar to its day 2 (or night 2) data point. The command to create the day or night model was:

The model provided the slope and its standard error reported on top of the parameter plots (e.g. Fig. 3h). A model without the fixed effect group was then created, as:

This null model was compared with the full model created above with a likelihood-ratio test, which provided the p-value reported in the figures and legends.

The LME analysis was informed by a published tutorial¹⁰⁴. The FramebyFrame R package performs the above LME analysis when generating parameter plots with ggParameterGrid(…) or writing a report of LME statistics using LMEreport(…).

The main likelihood-ratio test (see above) tested the null hypothesis that group assignment had no effect on parameter values. In experiments with more than two groups, each group was also compared to the reference group using estimated marginal means implemented in the R package emmeans, which provided the p-value.

In Fig. 6c, the LME analysis was performed on data from both clutches, as described above. The colour for each parameter was based on two p-values obtained using estimated marginal means: the p-value when comparing DMSO-treated scrambled-injected controls with DMSO-treated psen2 F0 knockouts (“knockout p-value”), and the p-value when comparing DMSO-treated scrambled-injected controls with drug-treated psen2 F0 knockouts (“drug p-value”), such that:

• if knockout p-value < 0.05 and drug p-value > 0.05, parameter was a “rescue”;

• if knockout p-value < 0.05 and drug p-value < 0.05, parameter was a “missed rescue”;

• if knockout p-value > 0.05 and drug p-value < 0.05, parameter was a “side effect”;

• if knockout p-value > 0.05 and drug p-value > 0.05, parameter was a “no effect”.

One-minute vs. frame-by-frame

The one-minute analyses developed previously^41–43 use the middur parameter calculated by ZebraLab. The middur parameter reports, for each minute, the number of seconds each larva spent above the freeze threshold (set by the user, typically 3 Δ pixel) and below the burst threshold (set by the user, typically 200 Δ pixel). In Fig. 2–supplement 2b,c, to better focus the comparisons on the effect of the one-minute binning, we re-calculated the middur dataset of the experiment using the FramebyFrame R package (function rawToMiddur(…)) setting the freeze threshold to 0. Larvae plotted were the scrambled-injected larvae from sorl1 F0 knockout clutch 2 (Fig. 4–supplement 1).

Pilot long-term video-tracking in 96-well plate

For the experiment in Fig. 2b, we placed 4-dpf scrambled-injected larvae in the wells of a mesh-bottom plate and added to the water bath 50 mL of paramecia culture filtered through a 40-μm cell strainer nylon mesh (Fisherbrand) to remove the large debris which may be detected by the camera of the ZebraBox. Larvae were video-tracked for a total of 208 hr, first on a 14 hr:10 hr light:dark cycle for 63 hr then in constant dim light at 30 lux for 145 hr for the free-running segment. In the afternoon of the first day in constant dim conditions, we removed most of the water from the bath using a 10 mL Pasteur pipette, replaced with fresh fish water, and added another 50 mL of filtered paramecia. As can be seen in Fig. 2b, this created a sharp rise in activity. We did not intervene for the rest of the experiment. The activity trace was generated as above (see Behavioural data analysis).

Sleep latency in presence of paramecia

We video-tracked wild-type 6-dpf larvae as described above (see Behavioural video-tracking) for 24 hr, including a 14-hr night. Larvae were in the wells of a standard 96-square well plate (Whatman), and we added two drops of filtered paramecia in half of the wells. There was no water bath and the light level at night was 0 lux.

Zebrafish orthologues of Alzheimer’s risk genes

We used the GWAS meta-analysis and list of most likely causal genes from Schwartzentruber et al. (2021)²². We downloaded from Ensembl the list of all zebrafish orthologues of human genes (last updated 27/04/2023). More details about GWAS loci, causal genes, and zebrafish orthologues can be found in Supplementary file 1.

Single-cell RNA sequencing data

We downloaded single-cell RNA sequencing (scRNA-seq) data generated by Raj et al. (2020)²⁶ from GEO repository GSE158142 (.rds.gz files, Seurat data structure). In Fig. 1b, each gene was considered expressed if it was detected in at least three cells at this developmental stage. In Supplementary file 1, we manually grouped the clusters from Raj et al. (2020) in broad categories to add colours. The expression data plotted (e.g. Fig. 1c) were scaled from raw counts by the authors such that the mean expression across all cells was 0 and the variance was 1.

In situ hybridization chain reaction

In situ hybridization chain reaction (HCR) probes were designed as described by Choi et al. (2018)¹⁰⁵ using a custom Python script. HCR split initiator sequences B1, B3, and B5 were obtained from Choi et al. (2018). To generate probe sets, probe pairs were excluded if they fell below melting temperature and %GC thresholds. Probe pairs with strong sequence similarity to off-target transcripts were also excluded. For each target transcript, we generated sets of 15–25 probe pairs. Multiple transcripts were often stained in individual larvae. HCR probes were purchased as custom DNA oligos from Thermo Fisher Scientific. HCR amplifiers (B1-Alexa Fluor 488, B3-Alexa Fluor 546, and B5-Alexa Fluor 647) and buffers were purchased from Molecular Instruments. More details about HCR probes can be found in Supplementary file 1.

The HCR protocol we used was adapted from MI-Protocol-RNAFISH-Zebrafish (Rev10) (Molecular Instruments). Proteinase K and methanol permeabilisation steps from the protocol were skipped. HCR to label Alzheimer’s risk genes was performed in Tg(elavl3:H2b-GCaMP6s) larvae homozygous for the mitfa^w2 allele (nacre)^106,107. 6-dpf larvae were euthanised and fixed in 4% paraformaldehyde (PFA)/Dulbecco’s phosphate-buffered saline (DPBS) overnight at 4°C. Following fixation, larvae were washed three times 5 min in phosphate-buffered saline (PBS). Larvae were then transferred to a SYLGARD-coated Petri dish and the eyes were removed using forceps. Sample preparation was completed by two brief PBST (1× DPBS + 0.1% Tween 20) washes. For the probe hybridization stage, larvae were first incubated in hybridization buffer at 37°C for 30 min. Meanwhile, a probe solution was prepared by adding 4 µL of each 1 µM probe set (4 pmol) to 500 µL hybridization buffer. Hybridization buffer was replaced by the probe solution, and larvae were incubated at 37°C overnight (12–16 hr).

The next day, excess probe solution was removed by washing larvae four times 15 min with 500 µL of probe wash buffer preheated to 37°C. Larvae were then washed twice 5 min with 5× SSCT at room temperature. Samples were kept at room temperature for subsequent amplification steps. First, larvae were transferred to amplification buffer at room temperature for 30 min. 3 µM stocks of hairpin H1 and hairpin H2 were individually heated to 95°C for 90 sec then left to cool at room temperature in the dark for 30 min. For HCR on up to 8 larvae together, a hairpin solution was prepared by adding 4 µL (12 pmol) of hairpin H1 and 4 µL of hairpin H2 (12 pmol) to 200 µL amplification buffer. Finally, amplification buffer was removed and the hairpin solution was added to the larvae which were incubated overnight (12–16 hr) in the dark at room temperature. After overnight incubation, excess hairpins were removed by washing in 5× SSCT for twice 5 min, then twice 30 min and finally once 5 min at room temperature. Larvae were transferred to PBS and kept at 4°C protected from light for up to 3 days.

For a subset of transcripts (appa, appb, slc6a4a, slc6a4b, tph1a, tph1b, tph2), HCR was not carried out in nacre, elavl3:H2b-GCaMP6s larvae. Instead, gad1b was labelled as a reference channel. We proceeded as above, with some amendments at the sample preparation stage. Larvae were fixed in 4% PFA with 4% (w/v%) sucrose overnight at 4°C. Brains were dissected with forceps, removing the eyes and the skin that covers the brain. After dissection, there was a 20-min postfix in 4% PFA with 4% (w/v%) sucrose followed by three washes in PBST to remove the fixative. When performing HCR on sorl1 F0 knockout and scrambled-injected larvae (Fig. 5e and Fig. 5–supplement 1c), larvae from both conditions were pooled in a single Eppendorf tube prior to the hybridization steps. Using a scalpel, a portion of the tail was removed from the scrambled-injected larvae to differentiate the genotypes during imaging.

For imaging, larvae were mounted in 1% low melting point agarose (Sigma-Aldrich) in fish water and imaged with a ZEISS LSM 980 equipped with an Airyscan 2 detector (CO-8Y multiplex mode, confocal resolution) and a ZEISS C-Aprochromat 10×/0.45 W M27 objective. The whole brain was imaged without tiling. The image size was 844.29 µm (3188 pixels) × 846.41 µm (3196 pixels). Each pixel was 0.265 × 0.265 × 1.5 µm. The laser wavelength was 639 nm for B5-Alexa Fluor 647; 561 nm for B3-Alexa Fluor 546; 488 nm for GCaMP6s.

Samples in Fig. 1–supplement 3a,b, Fig. 5e,f, and Fig. 5–supplement 1c,d were imaged with a lightsheet ZEISS Z.1 microscope equipped with a W Plan-Aprochromat 10×/0.5 M27 75 mm objective. The image size was 889.56 µm (1920 pixels) × 889.56 µm (1920 pixels). Each pixel was 0.46 × 0.46 × 1.0 µm. The laser wavelength was 638 nm for B5-Alexa Fluor 647; 561 nm for B3-Alexa Fluor 546; 488 nm for GCaMP6s or B1-Alexa Fluor 488 to image gad1db.

Brains were registered to the reference brain from the Zebrafish Brain Browser^108,109 with elavl3:H2B-GCaMP6s or gad1b as reference channel using ANTs toolbox version 2.1.0¹¹⁰, as described previously¹¹¹.

We prepared Fig. 1d, and Fig. 1–supplement 3 and 4 using Fiji¹¹². The minimum/maximum and contrast of each stack/channel were adjusted before generating a maximum Z-projection. We created the sagittal views using Image > Stacks > Reslice before generating a maximum Z-projection. Whole-brain dorsal and sagittal outlines were from the Zebrafish Brain Browser.

In the sorl1 experiments, masks for anatomical regions were from the Zebrafish Brain Browser. To measure total signal intensity in different anatomical regions (Fig. 5f left and Fig. 5– supplement 1d left), we summed the grey pixel values of every voxel within each mask on the stacks after registration. To count the number of positive voxels in different anatomical regions (Fig. 5f right and Fig. 5–supplement 1d right), we first rescaled each stack to the maximum grey value of the stack. We then applied a threshold at grey pixel value 15 (8bit image, so grey pixel value 0–255) so that the grey value of any voxel with signal below 15 was turned to 0. The threshold value was decided using the multi-Otsu algorithm with two classes implemented in the scikit-image Python package¹¹³. We then counted for each anatomical region the number of voxels with grey pixel value > 0.

To prepare Fig. 5e and Fig. 5–supplement 1c, we calculated each median stack using MATLAB. The maximum of all four median stacks were then adjusted to the same value in Fiji to keep intensities comparable. The dorsal and sagittal Z-projections were prepared as above.

Amyloid beta measurements on psen1 and psen2 F0 knockouts

psen1, psen2, and double psen1/psen2 F0 knockout larvae were generated as described above (see Generation of F0 knockout larvae). Most double psen1/psen2 F0 knockout larvae died or were severely dysmorphic by 5 dpf. The other larvae were raised until 16 dpf to increase the amount of tissue available for the assay. Larvae were euthanised by an overdose of tricaine then decapitated with a scalpel. Heads were snap-frozen in liquid nitrogen then homogenized using a mechanical homogenizer in 100 μL TBS (50mM Tris-HCl, pH 8.0) containing 1:200 protease inhibitor cocktail set III (Calbiochem). Homogenates were centrifuged at 16,000 g at 4°C for 30 min. The supernatant was then collected and stored at –70°C. We measured total protein concentration of each sample using the Pierce Detergent Compatible Bradford Assay Kit (ThermoFisher). Aβ38, Aβ40, and Aβ42 concentrations were measured on the Meso Scale Discovery platform (Meso Scale Diagnostics) using the V-PLEX Plus Aβ Peptide Panel 1 (4G8) kit.

Each pool of larvae was measured in four technical replicates: two at the original concentration and two diluted 1:1 in TBS. Meso Scale returned concentrations of Aβ38, Aβ40, Aβ42 in pg/mL. All Aβ38 measurements were below the detection limit. We converted the Aβ concentrations to pg Aβ/µg total protein using the total protein content measured with the Bradford assay. The limit of detection (LOD) for each Aβ species was the concentration of the most diluted standard in pg/mL. We calculated a sample-specific LOD by calculating the Aβ concentration (pg Aβ/µg total protein) that would have given exactly the LOD in pg/mL. For example, the LOD for Aβ40 was 39.79 pg/mL. Therefore, for a sample whose total protein content was 3788 µg/mL, Aβ40 concentration should have been 0.0105 pg Aβ40/g total protein to give exactly the LOD (39.79 pg/mL). Each data point represents the mean of the four technical replicates, except if both diluted replicates were below the sample-specific LOD, in which case the data point represents the mean of the two undiluted technical replicates. If one undiluted sample was below the LOD but not the other, we replaced the below-LOD measurement by the sample-specific LOD itself. If all four technical replicates (undiluted and diluted) were below the sample-specific LOD, the sample was marked with a cross in the plot.

Annotation of the Rihel et al., 2010 small molecule behavioural database

From the Rihel et al., 2010 small molecule zebrafish behavioural database⁴², we extracted the name of each compound tested and the 5,756 behavioural fingerprints, which are the mean z-score for each sleep/wake parameter (see Fig. 6a for the one-minute parameters). The same compound could have alternate database names (e.g. “Dopamine hydrochloride” and “Dopamine HCl”), and could vary by its stereochemistry or salts (e.g. atropine sulfate vs. atropine methyl nitrate). Therefore, we first simplified as much as possible each compound name by removing various stereochemistry and salt information. This reduced the number of unique compound names from 5,149 to 4,731. We then used the webchem R package¹¹⁴ to automatically search each name in the PubChem database and return the PubChem CID. CID could be retrieved automatically for 4,365 (92%) of the names. We manually searched the remaining 366 compound names, eventually leaving only 12 of 4,731 compound names without a CID. The final number of unique CIDs was 3,674. The Therapeutic Target Database (TTD) uses a custom TTD Drug ID but provided a file for cross-matching PubChem CIDs to TTD Drug IDs (P1-03-TTD_crossmatching.txt). We found a TTD Drug ID for 1,738 of the 3,674 CIDs (47%). Using each TTD Drug ID as query, we then extracted indications and target proteins from databases provided by the TTD using TTD Drug IDs as queries (P1-05-Drug_disease.txt and P1-01-TTD_target_download.txt). Proteins are given a target ID by the TTD, which we then used as query to extract the KEGG pathways each protein was associated with (P4-01-Target-KEGGpathway_all.txt).

As TTD Drug IDs were used to extract annotations but that only 47% of compounds were assigned one, the ZOLTAR app currently uses less than half of the dataset by Rihel et al., 2010 for prediction of disrupted pathways. In the future, other sources could be used to enrich the annotations, for example using machine learning to predict target proteins¹¹⁵.

Behavioural pharmacology from sorl1 F0 knockout’s fingerprint

We measured the cosine similarity (range −1.0–1.0) between the mean fingerprint of the two sorl1 F0 knockout clutches and each of the 5,756 small molecule fingerprints from the Rihel et al., 2010 database. We then ranked the 5,756 small molecule fingerprints from the most positive cosine (SU6656: cos = 0.83) to the most negative cosine (flumazenil: cos = −0.80).

At this stage, the general goal of the analysis is to detect whether compounds with a given annotation (indication, target, or KEGG pathway) are found more towards the top and/or bottom of the ranked list than expected by chance. For example, two cases suggest that inhibition of a pathway is what causes the knockout behavioural phenotype: all compounds inhibiting this pathway are found towards the top of the list (positive cosines); or all compounds activating this pathway are found towards the bottom of the list (negative cosines). The database may include both agonists and antagonists for a given pathway, so we also want to detect cases where compounds interacting with the pathway are found both towards the top and towards the bottom of the ranked list. Therefore, the null hypothesis for a given annotation is specifically that fingerprints with this annotation are found across the ranked list in a pattern that can be explained by chance, or that they are mostly found around the centre of the list (cos ∼ 0 position).

In practice, we counted fingerprint positions starting at the centre (cos ∼ 0 position) and counting in both directions. Namely, the fingerprint with the smallest positive cosine (6,7-dichloroquinoxaline-2,3-dione, cos = 0.0003) was position #1, then position #2 for the next smallest positive cosine, etc. We did not set negative positions, so the fingerprint with the smallest negative cosine (L-Histidine hydrochloride, cos = −0.0003) was also position #1, then position #2, etc. The fingerprint with the smallest positive cosine was not always at the centre of the list, as there could be more fingerprints with positive cosines than fingerprints with negative cosines, or vice-versa. In a case where there were more fingerprints with positive cosines, using these positions directly as ranks would have biased the analysis against detecting negative enrichment (i.e. more fingerprints towards the bottom of the list than expected by chance) as the random draws (see below) would have mostly picked fingerprints with large ranks on the positive-cosine side simply because there were more of them. We controlled for this by forcing ranks to be on the same range (same maximum rank) on either side of the list. Namely, we set the ranks of the fingerprints with positive cosines as:

The same was done for the ranks of the fingerprints with negative cosines (i.e. no negative ranks). Using 1000 as the maximum rank was arbitrary; any number could have been used.

For example, there were 3,201 positive-cosine fingerprints for sorl1. Therefore, the rank of the fingerprint with the smallest positive cosine was 1 × 1000/3201 = #0.31; the next fingerprint was rank 2 × 1000/3201 = #0.62, etc., until the fingerprint with the largest positive cosine which had rank 3201 × 1000/3201 = #1000.00. On the opposite side, there were 2,555 negative-cosine fingerprints. Therefore, the rank of the fingerprint with the smallest negative cosine was 1 × 1000/2555 = #0.39; the next fingerprint was rank 2 × 1000/2555 = #0.78, etc., until the fingerprint with the largest negative cosine which had rank 2555 × 1000/2555 = #1000.00.

At this stage, we had a list of fingerprints from largest positive cosine to largest negative cosine, each with a rank representing its distance from the cosine ∼ 0 position. We swapped each compound for its annotations, for example its target(s). If a compound had no annotated target (or indication or KEGG pathway), it was simply deleted without editing any of the ranks. If a compound had multiple annotated targets (or indications or KEGG pathways), all targets received the same rank. For example, aspirin had rank #874.36 (cos = −0.50) and two annotated targets: HMG-CoA reductase and Prostaglandin G/H synthase. Therefore, swapping aspirin for its targets generated two rows with the same rank #874.36. The same compound could appear multiple times in the list, in which case its targets (or indications or KEGG pathways) were simply repeated.

To measure enrichment of a given annotation towards the top and/or bottom of the list, we summed the ranks of all its instances. For example, the serotonin transporter (SLC6A4) appeared 59 times in the list for sorl1. The sum of those 59 ranks was 34,958. Was this sum of ranks surprising or could it be explained by chance? To test this, we randomly drew 59 positions and measured the sum of ranks 100,000 times to generate a null distribution of sum of ranks (e.g. Fig. 5b). We counted how many random draws gave a larger sum of ranks than the observed sum of ranks (34,958) to calculate a simulated p-value. In the case of SLC6A4, 1,127 random draws gave a larger sum of ranks; therefore, the probability that a sum of ranks as high as 34,958 was obtained by chance was estimated at 1,127/100,000 = 0.01127, which we reported as the simulated p-value.

A simpler alternative for this analysis may be to sum absolute cosines directly, instead of ranks, as a measure of enrichment. In theory, this option would consider more directly how well the knockout fingerprint and the small molecule fingerprints match.

Behavioural pharmacology from psen2 F0 knockout’s fingerprint

We measured the cosine similarity between the mean fingerprint of the two psen2 F0 knockout clutches and each of the 5,756 small molecule fingerprints from the Rihel et al., 2010 database. We then ranked the 5,756 small molecule fingerprints from the most positive cosine (3-3-Acetoxypregn-16-en-12,20-dione: cos = 0.82) to the most negative cosine (isogedunin: cos = −0.87). We searched within the ∼100 fingerprints with the most negative cosines (range of cosines: −0.72–0.87) for compounds which Rihel et al., 2010 labelled as shortlisted because it affected one behavioural parameter with a large effect size and/or affected the same parameter in the same direction across the two days/nights. From these compounds, we selected fenoprofen and betamethasone because they both had a replicate fingerprint with cosine <−0.60. Tinidazole was selected based on a preliminary experiment in wild-type larvae (data not shown), although it was not shortlisted by Rihel et al., 2010.

To simplify this selection process in the ZOLTAR app (francoiskroll.shinyapps.io/zoltar/), the shortlisted compounds are labelled, and clicking on a candidate compound automatically plots all the fingerprints from this compound (same PubChem CID).

Protein alignment plots

This refers to Fig. 3a,b; Fig. 3–supplement 3a,b; Fig. 4a; Fig. 4–supplement 2a,b; Fig. 4–supplement 3a; and Fig. 4–supplement 4a. Amino acid sequences were obtained from UniProt. The ClustalOmega algorithm, implemented in R by the msa package, was used to align the sequences. Definitions of “highly similar” and “weakly similar” groups of amino acids were taken from the ClustalOmega online documentation. The following groups of amino acids were considered highly similar: STA, NEQK, NHQK, NDEQ, QHRK, MILV, MILF, HY, FYW. The following groups of amino acids were considered weakly similar: CSA, ATV, SAG, STNK, STPA, SGND, SNDEQK, NDEQHK, NEQHRK, FVLIM, HFY. Identity was the number of positions where the zebrafish amino acid matched the human amino acid divided by the total length of the alignment. Similarity was the number of positions where the zebrafish amino acid matched or was highly similar to the human amino acid divided by the total length of the alignment. UniProt IDs of the amino acid sequences used were (human vs. zebrafish): appa, P05067 vs. Q6NUZ1; appb, P05067 vs. B0V0E5; psen1, P49768 vs. Q9W6T7; psen2, P49810 vs. Q90ZE4; apoea, P02649 vs. Q503V2; apoeb, P02649 vs. O42364; cd2ap, Q9Y5K6 vs. F1R1N9; clu, P10909 vs. Q6PBL3; sorl1, Q92673 vs. Q90ZE4.

Gene schematics

This refers to Fig. 3c,d; Fig. 3–supplement 3c,d; Fig. 4b; Fig. 4–supplement 2c,d; Fig. 4–supplement 3b; and Fig. 4–supplement 4b. Coordinates of 5′-UTRs, 3′-UTRs, and exon boundaries were obtained from Ensembl accessed through R using the biomaRt package v2.54.1¹¹⁶. For each gene, we drew the transcript that was used as reference when selecting the crRNAs, usually the longest coding transcript annotated in Ensembl. Ensembl IDs of transcripts drawn were: appa, ENSDART00000166786; appb, ENSDART00000077908; psen1, ENSDART00000149864; psen2, ENSDART00000006381; apoea, ENSDART00000172219; apoeb, ENSDART00000058965; cd2ap, ENSDART00000102611; clu, ENSDART00000127173; sorl1, ENSDART00000156995.

Pictures

For the pictures of the psen2 F0 knockout larvae and scrambled-injected siblings (Fig. 3f and Fig. 3–supplement 2a), larvae were anaesthetised and mounted in 1% low melting point agarose (Sigma-Aldrich) in fish water. Pictures were then taken with a Nikon SMZ1500 brightfield microscope with illumination from above the sample.

Pictures of the psen1/psen2 double F0 knockout larvae were taken with an Olympus MVX10 microscope connected to a computer with the software cellSens (Olympus).

Statistics

For statistics on behavioural parameters, see Statistics on behavioural parameters. For the permutations procedure used in Fig. 5a,b and Fig. 5–supplement 1a,b, see Behavioural pharmacology from sorl1 F0 knockout’s fingerprint.

Threshold for statistical significance was α = 0.05. In figures, ns refers to p > 0.05, * to p ≤ 0.05, ** to p ≤ 0.01, and *** to p ≤ 0.001. In text, data distributions are reported as mean ± standard deviation, unless stated otherwise.

In text, estimates of behavioural parameter effect sizes are often reported in % vs. scrambled-injected controls. To calculate those, we first calculated, for each clutch and day or night, the mean of all control data points for this parameter, typically returning four averages (e.g. clutch1 controls, night1 and night2; clutch2 controls, night1 and night2). To each average, the slope from the LME model (see Statistics on behavioural parameters) was added/subtracted to estimate the knockout averages. An effect size in % vs. controls (ef%) was then calculated for each clutch and day or night as:

If the slope was positive; or

If the slope was negative; where ko and con is the knockout or control average for one clutch and one day or night. The effect sizes were then averaged to return one effect size representing the estimated % increase or decrease vs. scrambled controls.

In Fig. 2–supplement 2d, the data were 47 observations, one per larva. Each observation was the time in hours when the larva had its first sleep bout (status 2). Every larva slept at least once during the experiment, so there were no censored (status 1) observations. We calculated the hazard ratio with a Cox proportional-hazards model and the p-value with a likelihood-ratio test. The functions we used were implemented in the R package survival v3.4.0. The FramebyFrame R package automatically performs this survival analysis when generating sleep latency survival plots with the function ggSleepLatencyGrid(…).

In Fig. 5f and Fig. 5–supplement 1d, we compared values between groups (sorl1 F0 knockouts vs. scrambled-injected) in each anatomical region using Welch’s t-test. To test the null hypothesis that group assignment had no effect on values across regions, we used LME modelling implemented in the lmer function of the R package lme4 v1.1.31¹⁰³. Taking total signal intensity as an example, the command to create the LME model was:

The model provided the slope and its standard error reported in the figures. We then created a model without the fixed effect group, as:

We compared this null model with the full model created above with a likelihood-ratio test, which provided the p-value reported in the figures.

Software

Data analysis was performed in R v4.2.2 ran through RStudio 2023.06.0+421. Analysis of HCR on sorl1 F0 knockouts used Jupyter notebooks written in Python 3. Figures were prepared with Adobe Illustrator 2019. Videos were prepared with Adobe Premiere Pro 2020.

Code availability

Source code of the FramebyFrame R package is available at github.com/francoiskroll/FramebyFrame. The package can be installed directly into R. The GitHub repository includes installation instructions, tutorial, and documentation.

The ZOLTAR online app (francoiskroll.shinyapps.io/zoltar/) was written in R using the Shiny package. Source code is available at github.com/francoiskroll/ZOLTAR.

The online app illustrating the simplified model of knockout by frameshift (francoiskroll.shinyapps.io/frameshiftmodel/) was written in R using the Shiny package. Source code is available github.com/francoiskroll/frameshiftShiny.

Other code used for analysis will be made available at github.com/francoiskroll/ZFAD.

Data availability

Data will be made available on Zenodo.

Acknowledgements

We thank the members of the Rihel lab and other zebrafish groups at University College London (UCL) for helpful discussions. We thank Daniel J. Stein and Douglas A. Lauffenburger (MIT) for inspiration. We thank Alexandra Lubin for the Illumina MiSeq runs; Chintan Trivedi for code used to design HCR probes and advice on analysis of the HCR images; Tom Ryan for help raising the psen1/psen2 knockout larvae used for the Aβ measurements; Alizée Kastler for the slc6a4b HCR probes; Stephen Carter for his assessment of the psen1/psen2 double knockout eye phenotype; Declan Lyons for the measurements of the 3D-printed plate’s mesh. We thank all supporting staff at UCL including Fish Facility staff for fish care and husbandry. FK thanks Filippo Del Bene for the time spent in his lab working on the manuscript. Most of the behavioural small molecule dataset was generated by JR and David Prober in Alexander F Schier’s lab at Harvard University. FK was supported and funded by the Leonard Wolfson PhD Programme in Neurodegeneration. The work was also funded by a Wellcome Trust Investigator Award awarded to JR (217150/Z/19/Z).