Partial loss of CFIm25 causes aberrant alternative polyadenylation and learning deficits

We previously showed that NUDT21-spanning copy-number variations (CNVs) are associated with intellectual disability (Gennarino et al., 2015). However, the patients’ CNVs also included other genes. To determine if reduced NUDT21 function alone can cause disease, we generated Nudt21+/- mice to mimic the human state of decreased expression. We found that although these mice have 50% reduced Nudt21 mRNA, they only have 30% less of its cognate protein, CFIm25. Despite this partial protein-level compensation, the Nudt21+/- mice have learning deficits and cortical hyperexcitability. Further, to determine the molecular mechanism driving neural dysfunction, we partially inhibited NUDT21 in human stem cell-derived neurons to reduce CFIm25 by 30%. This reduction in CFIm25 was sufficient to induce misregulated alternative polyadenylation (APA) and protein levels in hundreds of genes, dozens of which cause intellectual disability when mutated. Altogether, these results indicate that disruption of NUDT21-regulated APA events in the brain can cause intellectual disability.


INTRODUCTION 43
The brain is acutely sensitive to the dose of numerous proteins, such that even small 44 perturbations in their levels can cause neurological disease. Proteins that affect the expression of 45 8 spatial learning deficits require more time to find the platform throughout training. For an 166 additional measurement, the platform is removed after the training and the mice are placed in a 167 novel location on the perimeter of the pool. We then record how much time, out of a minute, 168 they spend searching for the platform in the quadrant of the pool where it had been. Mice with 169 spatial learning and memory deficits spend relatively less time in the platform quadrant, 170 indicating that they do not remember its location as well. We found that the Nudt21 +/mice 171 indeed have spatial learning deficits. On average, they required more time to locate the platform 172 during their training trials, despite swimming slightly faster and covering a greater distance 173 (Figure 2Ci & ii). Moreover, when the platform was removed, the Nudt21 +/mice spent ~30% 174 less time in the correct quadrant (Figure 2Ciii & iv). Thus, the conditioned fear and Morris water 175 maze results show that Nudt21 +/mice have learning and memory deficits in multiple domains. 176 177 Nudt21 +/mice have increased cerebral spike activity. In addition to intellectual disability, the 178 patients with NUDT21-spanning deletions had seizures. In general, though, mice are much less 179 sensitive to seizure-causing mutations than humans, but seizure susceptibility can be assessed 180 using electroencephalography (EEG) (Amendola et al., 2014;Jiang et al., 1998;Kriscenski-Perry 181 et al., 2002;Miura et al., 2002). Despite a lack of detectable seizures in Nudt21 +/mice, we found 182 significantly more EEG spikes in their frontal cortex relative to their wild-type littermates. This 183 result indicates that Nudt21 haploinsufficiency is alone sufficient to cause cortical 184 hyperexcitability, and suggests that NUDT21 loss of function might increase seizure risk ( Figure  185 3). 186 187 9 NUDT21 depletion induces aberrant alternative polyadenylation and altered protein levels. 188 The observations that individuals with NUDT21-spanning CNVs have intellectual disability (ID) 189 and that Nudt21 +/mice have learning deficits provides strong evidence that NUDT21 loss of 190 function causes disease, but do not reveal the molecular pathology. Therefore, to understand how 191 partial loss of NUDT21 function might sicken neurons and cause ID, we used shRNA to inhibit 192  (Figure 4-figure supplement 1A & B). 199 Like in the whole-brain extracts from Nudt21 +/mice, the human neurons had a 30% 200 reduction of CFIm25 protein, despite a greater reduction of NUDT21 mRNA (Figure 4A & B). 201 This difference shows that there is also homeostatic stabilization of CFIm25 in human neurons. 202 Intriguingly, in addition to reduced NUDT21 mRNA levels as expected, the PAC-seq data also 203 showed that NUDT21 itself undergoes aberrant alternative polyadenylation following its 204 depletion (Figure 4Ci). These shortened NUDT21 mRNAs have no stop codon, and are likely to 205 undergo non-stop decay, an mRNA decay mechanism that degrades mRNAs lacking an in-frame 206 stop codon van Hoof et al., 2002). 207 As in the NUDT21 CNV patient-derived lymphoblasts, NUDT21 loss of function causes 208 3′ UTR shortening of MECP2 in human neurons, as well as other genes strongly regulated by 209 NUDT21 in cancer-cell-line assays, such as VMA21, LAMC1, and PAK1 (Figure 4Cii  Notwithstanding disrupted RNA localization, the most apparent consequence of 219 misregulated APA is altered protein levels (Tian and Manley, 2016). As expected, for many of 220 the genes with significantly misregulated APA, we also see concordant changes in protein levels 221 ( Figure 4E). Most commonly, NUDT21 loss results in shorter mRNAs and increased protein, 222 such as with MeCP2 and VMA21, which respectively had protein-level increases of 50% and 223 almost 300% (Figure 4Cii-iii & E). However, there are exceptions that further clarify the 224 relationship between NUDT21 loss and protein levels. Occasionally, the shortened mRNA will 225 lose its stop codon and so likely undergo non-stop decay, resulting in reduced protein, such as 226 with ZCCHC6 (Figure 4Civ & E). Conversely, as with KIF9, NUDT21 reduction occasionally 227 induces synthesis of longer mRNAs, and this could paradoxically increase protein levels if the 228 longer mRNAs gain a stop codon and cease to be degraded by non-stop decay (Figure 4Cv & E). 229 Intriguingly, the degradation of KIF9 and others through non-stop decay in the control neurons 230 shows an example of non-stop decay being used for normal gene regulation, rather than just as a 231 quality control mechanism. We observed that ~15% of the genes with altered APA have a 232 change in the percentage of mRNA isoforms undergoing non-stop decay. However, the majority 233 11 follow the trends seen in cancer cell lines: NUDT21 loss of function in neurons causes 234 widespread 3′ UTR shortening and increased protein levels (Figure 4D & E). 235

DISCUSSION 237
Here we show that a partial loss of Nudt21 function causes learning deficits and cortical 238 hyperexcitability in mice. Further, we show that partial loss of NUDT21 function broadly 239 disrupts gene expression in human neurons via widespread misregulated alternative 240 polyadenylation and protein levels. Alongside our previous discovery that patients with 241 NUDT21-spanning deletions have intellectual disability and gnomAD data that NUDT21 is a 242 highly constrained gene, our results provide strong evidence that partial loss of NUDT21 243 function causes intellectual disability (Gennarino et al., 2015;Lek et al., 2016). 244 Because a relatively small reduction in CFIm25 protein was sufficient to cause deficits, 245 these data suggest that individuals with missense variants in NUDT21 that affect its function may 246 also have intellectual disability. Moreover, duplications of NUDT21, which we previously 247 showed are associated with ID, should lead to comparable dysregulation of alternative 248 polyadenylation and protein levels, including reduced MeCP2, and thus cause disease 249 (Gennarino et al., 2015). 250 Beyond providing insight into NUDT21-associated disease, these data provide useful 251 perspectives on the broader field of pediatric neurodevelopmental disease research. They 252 illustrate the importance of protein-level homeostasis. A mere 30% reduction in CFIm25 protein 253 was sufficient to cause learning deficits in mice, and mice are neurologically less sensitive to 254 genetic insult than humans (Tan and Zoghbi, 2018). Further, many genes associated with 255 neurodevelopmental diseases encode proteins that regulate or affect transcription (Rubeis et al., We incubated the lysates on ice for ten minutes, then rotated them for 20 minutes at room 441 temperature. We spun the samples at top speed for 20 minutes to remove membrane, then 442 quantified the protein levels with the Pierce Protein BCA Assay kit (ThermoFisher Scientific). 443 For the ESC-derived neurons infected with shRNA targeting NUDT21, our extraction protocol 444 was similar, except we lysed them directly in their 12-well plate and rocked them for 20 minutes 445 at room temperature. 446 447 Western blot: We diluted the protein to 1 µg/µL in reducing buffer (LDS and sample reducing 448 agent (ThermoFisher Scientific)) and ran 10 µg/sample. We imaged the membranes with the LI-449 COR Odyssey and analyzed the data with LI-COR Image Studio (LI-COR Biosciences), 450 comparing H3-normalized CFIm25 levels. We confirmed the specificity of the CFIm25 antibody 451 by shNUDT21 knockdown in HEK293T cells. We present the data from human neurons infected 452 with shRNA clone V2LHS_253272 because we used it for the PAC-seq experiment. We 453 analyzed the data by two-tailed, unpaired t-test, and present it as mean ± SEM. *, **, ***, and 454 **** denote P<0.05, P<0.01, P<001, P<0.0001.

Behavioral tests 482
Because neurodevelopmental disease loss-of-function mouse models typically have more 483 pronounced phenotypes when they are older, we started our behavioral battery on 30-week-old 484 mice. We performed the assays on two cohorts. We tested the mice during their light cycle, 485 typically between 11 AM and 5 PM. Prior to each assay, we habituated the mice to the testing 486 facility for 30-60 minutes. The investigators were blind to the mice's genotypes during all 487 assays. 488 489

Conditioned fear 490
We first habituate the mice for 30 minutes in an adjacent room on each day of the test. On day 491 one, we conditioned the mice by placing them in the habitest operant cage (Coulbourn) for a 492 training session. The training consists of two minutes habituation, then a 30 second 85 dB tone 493 followed by a foot shock of 1.0 mA for two seconds. After another two minutes, the 30 second 494 85 dB tone is played again (the training day cue). Throughout the experiment, except for the two 495 seconds during the foot shock, the FreezeFrame3 system (Coulbourn/Actimetrics) recorded the 496 28 mice's movement and freezing episodes. On day two, we performed the contextual and cued fear 497 assays. For contextual fear, we returned the mice to the test chambers precisely as we had done 498 during the training, and recorded their freezing in the chamber for five minutes. We waited two 499 hours before beginning the cued fear assay. We first changed the holding cages and test chamber 500 shape, color, texture, scent, and lighting to make the experience as unrecognizable as possible to 501 the mice. We then placed them in the modified chamber, and after three minutes played the 30 502 second 85 dB tone, then recorded their freezing for the following three minutes. We analyzed the 503 difference in freezing between the two groups after the second sound cue (training cue) on day 504 one, throughout the contextual fear test, and after the sound cue in the cued fear test, each by 505 unpaired, two-tailed t-test. We excluded data for all the mice from one contextual fear trial. We 506 suspect there was a technical error in the collection of those data: the FreezeFrame3 system 507 recorded them as freezing far more than the mice in any other trial (50-80%) and two of the three 508 were significant outliers in the Grubbs test. 509 510

Open field 511
We lit the room to 200 lux and set the ambient white noise to 60 dB during habituation and 512 throughout the test. We placed each mouse in the open field, a 40 x 40 x 30 cm chamber 513 equipped with photobeams (Accuscan Instruments), and recorded their activity for 30 minutes. 514 We analyzed total distance and center tolerance (center distance/total distance) by unpaired, two-515 tailed t-test. 516 517 Morris water maze 518 29 Our Morris water maze experiment took pl ace in a 120 cm diameter pool of water. We hid a 10 519 cm X 10 cm platform 0.5-1 cm underwater in the Southeast quadrant. On each wall of the testing 520 room, we taped brightly colored shapes that the mice can use for orientation. Our experiment 521 spanned four days. Each day, we set the lighting to 60 lux and habituated the mice in the testing 522 room in their home cages for 30 minutes, then ten minutes in holding cages. Prior to the first 523 day's experiment, we introduced the mice to the invisible platform by placing them on it for ten 524 seconds. We next pulled the mice into the water and let them swim for ten seconds to ensure they 525 could swim, then placed them directly in front of the platform to confirm they could climb back 526 on to it. We tested the mice in two training blocks per day for the four days. Each training block 527 consisted of four trials. For each trial, we placed the mice in a different quadrant of the pool 528 (North, South, East, or West); the quadrant order was the same for every mouse in the trial, but 529 different for every trial. We removed the mice after they found the platform, or if they did not 530 find that platform, we guided them to the platform to rest on it for ten seconds before removing 531 them. We used an EthoVision XT automated video tracking system (Noldus Information 532 Technology) to track the mice's location, speed, and latency to find the platform. After the 533 second training block on the fourth day, we immediately performed the probe trial: we removed 534 the platform and placed the mice in a new location, the Northwest quadrant, and tracked them for 535 60 seconds. We analyzed their speed by unpaired, two-tailed t-test; their latency to find the 536 platform by two-way, repeated measures ANOVA (genotype*block); and their time in each 537 quadrant in the probe trial by two-way, repeated-measures ANOVA (genotype*quadrant). 538 539

Video electroencephalography (EEG) and spike counting 548
Surgery and data recordings: The Baylor College of Medicine Institutional Animal Care and 549 Use Committee approved all research and animal care procedures. We tested eight Nudt21 +/-550 mice and six wild-type littermate controls. Experimenters were blind to the mouse genotype. We 551 secured 54-week-old mice on a stereotaxic frame (David Kopf) under 1-2% isoflurane 552 anesthesia. Each mouse was prepared under aseptic condition for the following recordings: the 553 cortical EEG recording electrodes of Channels 1 and 2 were made of Teflon-coated silver wires 554 (bare diameter 127 µm, A-M systems) and implanted in the subdural space of the parietal cortex 555 and frontal cortex, respectively, with reference at the midline over the cerebellum. The electrode 556 of the third channel, made of Teflon-coated tungsten wire (bare diameter 50 µm, A-M systems) 557 was stereotaxically aimed at the hippocampal dentate gyrus (1.9 mm posterior, 1.7 mm lateral, 558 and 1.8 mm below the bregma) with reference in the ipsilateral corpus callosum (Paxinos and 559 Franklin, 2001). In addition, Teflon-coated silver wires were used to record the electromyogram 560 (EMG) in the neck muscles to monitor mouse activity. All of the electrode wires together with 561 the attached miniature connector sockets were fixed on the skull by dental cement. After two 562 weeks of post-surgical recovery, mice received three two-hour sessions of EEG/EMG recordings 563 31 over a week. Signals were amplified (100x) and filtered (bandpass, 0.1 Hz -1 kHz) with the 564 1700 Differential AC Amplifier (A-M Systems), then digitized at two kHz and stored on disk for 565 off-line analysis (DigiData 1440A and pClamp10, Molecular Devices). The time-locked mouse 566 behavior was recorded by ANY-maze tracking system (Stoelting Co.). 567 568 EEG data analysis: Abnormal synchronous discharges were manually identified when the sharp 569 positive deflections exceeding twice the baseline and lasting 25-100 ms (Roberson et al., 2011). 570 We counted the number of abnormal spikes over the recording period using Clampfit 10 software 571 (Molecular Devices, LLC) and averaged the spike numbers across sessions for each animal. 572 Since the data follow a lognormal distribution, we statistically compared the genotypes with the 573 Mann-Whitney test using Prism 7 (Graphpad). The measure of central tendency is the geometric 574 mean and * indicates P<0.05. 575 576 shRNA lentivirus production and titer assessment 577 Virus production. We made several viruses that express shRNAs targeting NUDT21. We 578 transfected 45 ug DNA into 80%-90% confluent, low-passage HEK293T cells (ATCC CRL-579 3216; RRID:CVCL_0063) in 150 mm dishes at a 4:3:1 ratio of pGIPz, psPAX2, pMD2.G with 580 TransIT-293 transfection reagent (Mirus, MIR 2706). The following day, we changed the media 581 to 10 mL. At 48 and 72 hours, we collected and pooled their media, then centrifuged at 4000 x g 582 for ten minutes and filtered the supernatant through a Poly-ethersulfone filter (VWR, 28145-505) 583 to remove cellular debris. We concentrated the virus 100-fold with Lenti-X concentrator 584 (Clontech, 631231) following the manufacturer's recommendations before aliquoting and 585 freezing at -80C. 586 32 587 Titer assessment. 588 We measured the viral titer using Open Biosystems pGIPZ method (Thermo Fisher Scientific). 589 We plated 5 X 10 4 HEK293T cells in a 24-well plate. The following day, we made a serial 590 dilution of the virus in a 96-well plate, and when the HEK293T cells reached ~50% confluency, 591 we infected them with the diluted virus. We cultured the cells for two days, then counted the 592 tGFP colonies with Axiovert 25 microscope (Zeiss) microscope X-cite 120 lamp (ExFo) to 593 determine the viral titer. Our viruses had 10 9 transducing units/mL. We lysed the neurons in the tissue-culture plate with TRIzol Reagent (ThermoFisher Scientific) 621 and immediately transferred the lysate to microfuge tubes for trituration. We then isolated the 622 RNA by chloroform phase separation, precipitation with 2-propanol, washing with 75% ethanol, 623 and eluting in water. 624 625 Library preparation and sequencing 626 We prepared sequencing libraries as previously described (Routh et al., 2017). We reverse 627 transcribed 1 ug of total RNA with the partial P7 adapter (Illumina_4N_21T) and dNTPs with 628 the addition of spiked-in azido-nucleotides (AzVTPs) at 5:1. We click-ligated the p5 adapter 629 (IDT) to the 5′ end of the cDNA with CuAAC. We then amplified the cDNA for 21 cycles with 630 Universal primer and 3′ indexing primer and purified it on a 2% agarose gel by extracting 631