Biology subjects

Bao, G.

Publications and source records attributed to Bao, G..

At least 19 recordsLinked to original sources

bioRxiv2026

Proteome-wide quantification of protein turnover in frog and fly embryos reveals divergent strategies of maternal inheritance

Every embryo inherits a maternal proteome that it must remodel with zygotic proteins to build its many cell types. The fate of the maternal proteome remains contested because indirect measurements cannot resolve it. Here, we combine 18O-water labeling with multiplexed proteomics to quantify protein turnover proteome-wide in frog and fly embryos. Through hatching, the frog preserves the bulk of its maternal proteome, confining rapid degradation to a small regulatory module. The fly cannot meet its synthesis demand from yolk alone and instead degrades nearly all maternal proteins, including housekeeping proteins long assumed stable, recycling them into new protein. Yet the turnover hierarchy is conserved, with disordered and regulatory proteins degrading fastest, while the fly rescales the whole proteome ~eightfold faster. These results recast the developmental proteome as both informational inheritance and metabolic reserve, establish 18O-water labeling as a turnover method for non-feeding organisms, and provide a resource of embryonic half-lives.

developmental biology↗

bioRxiv2026

Genetically-encoded discovery and development of peptide-macrocycle imaging agents for PD-L1

The unique cell surface composition of tumor cells forms the molecular basis for many targeting and cell-based therapies. Here, we describe the development of novel peptide-based targeting agents for programmed death ligand 1 (PD-L1). Molecular imaging by peptide agents, coupled with therapeutic intervention using the same modality, represents a critical advancement in cancer management. Whole-body PET imaging of PD-L1 expression offers a superior alternative to traditional immuno-histochemistry, making PD-L1 radiodiagnostic imaging a highly sought-after modality. PD-L1 targeting modalities developed for clinical imaging to date can be divided into antibodies, protein domains, and small macrocyclic peptides with fewer than 20 amino acids. The latter modalities can address many challenges seen in antibody-based targeting vectors. All potent PD-L1 targeting peptide modalities reported to date rely extensively on non-canonical amino acids (ncAAs). Here, we report a comprehensive structure-activity relationship (SAR) analysis of a family of macrocycles discovered from an Sx2Cx8Cx2 phage-display library composed entirely of natural amino acids (x represents 19 natural amino acids excluding Cys). Using >10,000 variants in ''focused'' phage-display libraries, we optimized these macrocycles to achieve single-digit-nanomolar potency in protein- and cell-based assays. En route to this optimization, the activity of 216 synthetic macrocycles towards PD-L1 was measured in five distinct assays; two leads have been evaluated by imaging in tumor xenografts in mice, and the X-ray structure of one advanced lead in complex with PD-L1 has been determined at 2.78 [A] resolution. This publication demonstrates the development potential of PD-L1-targeting macrocycles that do not require extensive incorporation of ncAAs and the democratization of discovery by mapping the optimization path to single-digit-nanomolar assets for targeted radiopharmaceuticals via canonical phage-display technology.

pharmacology and toxicology↗

bioRxiv2026

Neuromodulation of Foxp2+ hypothalamic neurons induces therapeutic hypothermia

The inability to selectively trigger therapeutic hypothermia independent of environmental cooling has hindered causal analysis of its broader physiological benefits. Here, we employed P57, a natural product that pharmacologically induces therapeutic hypothermia circumventing external cold stress, to identify the neuronal substrates underlying hypothermia-induced antitumor effects. Integrating functional ultrasound imaging, activity-dependent neuronal labeling, and single-nucleus RNA sequencing, we identify a previously unrecognized population of Foxp2 neurons in the medial preoptic area (MPA) that are selectively activated by P57. Inhibition of MPA Foxp2+ neurons abolishes P57-induced hypothermia, whereas the paraventricular hypothalamus (PVH) serves as a critical downstream node. Importantly, targeted activation of MPA Foxp2 neurons is sufficient to induce sustained hypothermia, suppress systemic metabolism, and inhibit tumor growth, providing causal evidence that neuronally driven reductions in core body temperature can exert antitumor effects. Together, these findings establish MPA Foxp2 neurons as a controllable node for therapeutic hypothermia induction and demonstrate that neuromodulation of defined neuronal populations can achieve therapeutic benefit by directly controlling physiological states.

neuroscience↗

bioRxiv2026

Hemodynamic phenotypes linked to high-altitude subclinical organ damage

BackgroundChronic exposure to high-altitude hypoxia imposes sustained cardiovascular stress, yet hemodynamic adaptation among healthy high-altitude dwellers is heterogeneous and remains poorly characterized. This study aimed to identify distinct hemodynamic phenotypes in a healthy high-altitude population using unsupervised machine learning and to evaluate their association with multi-system subclinical target organ damage. MethodsThis cross-sectional study enrolled 694 healthy adults permanently residing at [≥]3300 m on the Qinghai-Tibet Plateau. Unsupervised K-means clustering was performed on nine hemodynamic variables, including peripheral and central blood pressures, augmentation index (AIx), pulse pressure amplification ratio (pPP/cPP), and systolic pressure amplification (pSBP-cSBP). Differences across phenotypes in carotid intima-media thickness (IMT), estimated glomerular filtration rate (eGFR), left ventricular mass index (LVMI), and pulse wave velocity (PWV) were assessed using one-way ANOVA with Bonferroni-corrected post-hoc tests. ResultsThree distinct hemodynamic phenotypes were successfully identified. The C2 (Balanced Adaptation) phenotype (n = 245) demonstrated the most favorable hemodynamic profile, characterized by the lowest blood pressure and augmentation index (AIx) values, along with the highest peripheral-to-central pulse pressure ratio (pPP/cPP). The C1 (Vascular Stress) phenotype (n = 267) presented with normal peripheral systolic blood pressure (125.9 {+/-} 11.3 mmHg) but exhibited markedly elevated wave reflection indices, including the highest heart rate-adjusted augmentation index (AIx@HR75: 31.9 {+/-} 9.7%) and the lowest pPP/cPP ratio (1.29 {+/-} 0.08). The C3 (High-Load Decompensation) phenotype (n = 182) displayed significantly elevated blood pressures and the greatest overall hemodynamic load. Regarding target organ damage, a clear gradient was observed across the three phenotypes. The C3 phenotype showed the highest carotid intima-media thickness (IMT: 1.162 {+/-} 0.23 mm) and left ventricular mass index (LVMI: 69.18 {+/-} 40.73 g/m{superscript 2}). Conversely, the C2 phenotype exhibited the highest estimated glomerular filtration rate (eGFR: 97.38 {+/-} 16.38 mL/min/1.73m{superscript 2}) and the lowest IMT (0.994 {+/-} 0.26 mm). The C1 phenotype consistently displayed intermediate values for all organ damage indicators. After Bonferroni correction, all pairwise comparisons for LVMI and pulse wave velocity (PWV) reached statistical significance (all P < 0.05). ConclusionsHealthy high-altitude individuals manifest three distinct hemodynamic phenotypes arrayed along a cardiovascular risk continuum. The novel Vascular Stress (C1) phenotype represents a "masked" high-risk state characterized by normal peripheral blood pressure but elevated arterial stiffness and wave reflection, challenging sole reliance on brachial pressure for risk assessment. This phenotype-based stratification provides a framework for precision prevention and early intervention in high-altitude populations.

physiology↗

bioRxiv2026

Universal Baseline for in vitro Selection of Genetically Encoded Libraries

Genetically encoded (GE) libraries enable identification of high-affinity ligands for diverse molecular targets through iterative in vitro selection and DNA sequencing or next-generation sequencing (NGS). Despite their impact in therapeutic development, a systematic framework for evaluating reproducibility in GE-molecular discoveries remains limited. To aid such analysis, we introduce the concept of baseline response, which reproducibly partitions active and inactive members of in vitro selection. The baseline response is provided by spiking a random DNA-barcoded population. We calibrated the baseline concept using Bioconductor EdgeR differential enrichment (DE) analysis of NGS of phage-displayed selection on oligosaccharide chitin and hepatitis virus NS3a* protease as model targets. We further show that mixing discovery campaigns also offers an effective baseline: chitin-enriched peptides serve as a baseline for DE-analysis of NS3a* selection and NS3a*-enriched peptides serve as a baseline for chitin binders. We applied baseline-stratified DE-analysis to 66 parallel selections performed in 3-5 replicates across 22 extracellular targets, including HER1-3, EpCAM, CAIX, PD-L1, and eight integrin receptors. Automated DE-analysis across hundreds of NGS files produced hits validated in a secondary screen and yielded synthetic macrocyclic ligands with mid-nanomolar affinity confirmed in 2-3 biophysical assays. For PD-L1, we further demonstrated how baseline-calibrated NGS data provide decision-enabling information for optimization of peptide macrocycles to yield potent single-digit nanomolar ligands for the cell-surface receptor. We anticipate that baseline-based analyses of NGS data from in vitro selection procedures will offer a scalable framework for reproducible hit discovery and standardized analysis across diverse in vitro selection campaigns. Significance StatementGenetically encoded selection technologies such as phage, mRNA and ribosome display, have produced FDA-approved therapeutics and numerous clinical candidates. Yet reproducibility in such in vitro discovery systems is rarely evaluated against a defined experimental baseline. Here, we establish a universal baseline by spiking unrelated, DNA-barcoded peptide sequences into selection libraries and quantifying their binding alongside target-enriched populations. This composition-agnostic strategy enables rigorous normalization, confidence assessment, and cross-target comparison of molecular discovery outcomes. Our framework introduces practical standards for reproducibility and statistical benchmarking across genetically encoded display platforms.

biochemistry↗

bioRxiv2025

High-Power Dual-Channel Chamber for High-Frequency Magnetic Neuromodulation

ObjectiveSeveral novel methods, including magnetogenetics and magnetoelectric stimulation, use high frequency alternating magnetic fields to precisely manipulate neural activity. To quantify the behavioral effects of such interventions in a freely moving mouse, we developed a dual-channel magnetic chamber, specifically designed for rate-sensitive magnetothermal-genetic stimulation, and adaptable for other uses of alternating magnetic fields. ApproachThrough an optimized coil design, the system allows independent control of two spatially orthogonal uniform magnetic fields delivered at different frequencies within a 10 x 10 x 6 cm3 chamber suitable for mouse studies. The two channels have nominal frequencies of 50 and 550 kHz with peak magnetic field strengths of 88 and 12.5 mT, achieved with resonant coil drives having peak voltages of 1.6 and 1.8 kV and currents of 1.0 and 0.26 kA, respectively. Additionally, a liquid cooling system enables magnetic field generation for second-level durations, and an observation port and camera allow video capture of the animals behavior within the chamber. Main ResultsThe system generates high-amplitude magnetic fields across two widely separated frequency channels with negligible interference (< 1%). Relatively uniform magnetic field distribution ({+/-}10% across 94% of the chamber volume) is maintained throughout the chamber, and temperature increase of the inner side of the coil enclosure during the operation is limited to < 0.35 {degrees}C/s to ensure in vivo safety. Using cobalt-doped and undoped iron oxide nanoparticles, we demonstrate channel-specific heating rates of 3.5 {degrees}C/s and 1.5 {degrees}C/s, respectively, validating frequency-selectivity. Both channels can run continuously for 4 seconds stably. SignificanceWe present a novel magnetic stimulation platform that combines high-frequency, high-power capability with two independently-controlled channels generating different frequencies, along with a real-time behavioral observation system for freely moving animals. The system supports frequency-multiplexed stimulation strategies for precise modulation of neural activity, making it a versatile tool for advancing magnetogenetics, neural circuit interrogation, and noninvasive stimulation approaches in neuroscience and bioengineering.

bioengineering↗

bioRxiv2025

Synergistic Arg-C Ultra and Lys-C Digestion for Quantitative Proteomics

Shotgun proteomics hinges on complete enzymatic digestion of proteins into peptides. Incomplete digestion narrows proteome coverage and inflates variability in quantitative workflows, whether label-free DIA or multiplexing with isobaric tags. Sequential Lys-C/Trypsin digestions mitigate missed cleavages at lysine residues, but arginine sites remain a persistent challenge. Arg-C Ultra, a recently released cysteine protease, efficiently targets arginine residues but requires reducing conditions that inactivate Lys-C activity and compromise NHS- ester labeling in multiplexed workflows. Here, we systematically characterized Arg-C Ultra and Lys-C with chromogenic substrates that mimic arginine- and lysine-containing peptides, as well as shotgun proteomics. Arg-C Ultra operates optimally at room temperature, pH 7.5-8.5, under reducing conditions, whereas Lys-C is most active at 37 {degrees}C, pH 7.5-8.5, yet rapidly loses activity when exposed to common reductants. Among tested reducing agents, 1 mM TCEP uniquely preserved TMTpro integrity while sustaining Arg-C Ultra activity. Guided by these insights, we established a sequential digestion workflow that is fully compatible with both label- free DIA and TMTpro multiplexing. Proteins are first digested overnight with Lys-C at 37 {degrees}C (pH 8.5), then treated with 1 mM TCEP and Arg-C Ultra at room temperature (pH 8.5). The resulting peptides can be analyzed directly by label-free DIA or subjected to TMTpro labeling for multiplexed quantification. Applied to HeLa cell lysates, this protocol achieved >99% arginine and 95% lysine cleavage efficiencies, boosting the number of quantified proteins by 6% in label- free DIA and 11% in TMTproC experiments. Replicate measurements displayed reproducibility that approached the limits set by ion statistics. Thus, the introduced synergistic Lys-C/Arg-C Ultra digestion strategy enhances proteome coverage with excellent quantitative reproducibility across both label-free and multiplexed platforms.

systems biology↗

bioRxiv2025

SarcAsM: AI-based multiscale analysis of sarcomere organization and contractility in cardiomyocytes

Cardiomyocyte function critically depends on sarcomere dynamics and their organization in myofibrils. To uncover how cardiomyocyte function emerges from individual sarcomere dynamics, comprehensive analysis of Z-bands (nanometer scale), sarcomeres ([~]2 {micro}m scale), and myofibrils ([~]10 to 100 {micro}m scale) is required. Tools for such multiscale analyses are presently lacking. Here we introduce the Sarcomere Analysis Multitool (SarcAsM), which combines deep learning and graph-based methods for automated, fast, and unbiased structural assessment of sarcomeric Z-bands, sarcomeres, and myofibrils as well as their organization in larger myofibril domains. SarcAsM features a generalist deep learning model, pre-trained on a broad range of experimental and published images, ensuring its versatile application and immediate usability across diverse datasets. Finally, we demonstrate the versatile utility of SarcAsM in analyzing sarcomere structure and dynamics under acute and chronic drug exposure in hiPSC-derived cardiomyocytes with fluorescently tagged Z-bands. SarcAsM is available as open-source Python package and stand-alone application.

bioinformatics↗

bioRxiv2025

Haplotype rather than single causal variants effects contribute to regulatory gene expression associations in human myeloid cells

Genome-wide association studies typically identify hundreds to thousands of loci, many of which harbor multiple independent peaks, each parsimoniously assumed to be due to the activity of a single causal variant. Fine-mapping of such variants has become a priority and since most associations are located within regulatory regions, it is also assumed that they colocalize with regulatory variants that influence the expression of nearby genes. Here we examine these assumptions by using a moderate throughput expression CROPseq protocol in which Cas9 nuclease is used to induce small insertions and deletions across the credible set of SNPs that may account for expression quantitative trait loci (eQTL) for genes associated with inflammatory bowel disease (IBD). Of the 4,384 SNPs targeted in 88 loci (an average of 50 per locus), 439 were significant and further examined for validation. From these, 98 significantly altered target gene expression in HL-60 myeloid cell line, 74 in induced macrophages from these HL-60 cells, and 78 in induced neutrophils for a total of 201 validated effects (46%), 43 of which were observed in at least two of the cell types. Considering the observed sensitivity and specificity of the controls, we estimate that there are at least 150 true positives per cell type, an average of almost 2.4 for each of the 64 eQTL for which putative causal variants have been fine-mapped. This implies that haplotype effects are likely to explain many of the associations. We also demonstrate that the same approach can be used to investigate the activity of very rare variants in regulatory regions for 89 genes, providing a rapid strategy for establishing clinical relevance of non-coding mutations.

genomics↗

bioRxiv2024

Covalently linked adenovirus-AAV complexes as a novel platform technology for gene therapy

Adeno-associated virus (AAV) has found immense success as a delivery system for gene therapy, yet the small 4.7 kb packaging capacity of the AAV sharply limits the scope of its application. In addition, high doses of AAV are frequently required to facilitate therapeutic effects, leading to acute toxicity issues. While dual and triple AAV approaches have been developed to mitigate the packaging capacity problem, these necessitate even higher doses to ensure that co-infection occurs at sufficient frequency. To address these challenges, we herein describe a novel delivery system consisting of adenovirus (Ad) covalently linked to multiple adeno-associated virus capsids as a new way of more efficiently co-infecting cells with lower overall amounts of AAVs. We utilize the DogTag-DogCatcher (DgT-DgC) molecular glue system to construct our AdAAVs and we demonstrate that these hybrid virus complexes achieve enhanced co-transduction of cultured cells, including physiologically relevant primary cells. On this basis, AdAAV technology may eventually facilitate therapeutic co-delivery of multiple transgenes at low virus doses for treating complex ailments.

synthetic biology↗

bioRxiv2024

SKP2-mediated FBXO2 proteasomal degradation drives hepatocellular carcinoma progression via stabilizing Hsp47

Accumulating studies highlight that dysregulated E3 ubiquitin ligases are associated with the onset and advancement of cancers. Nevertheless, the impact and mechanism of most E3 ubiquitin ligases on tumorigenesis and tumor metastasis remain poorly understood. Here, we show that loss of FBXO2, an E3 ubiquitin ligase, accelerates hepatocellular carcinoma (HCC) tumor growth and metastasis to the lung through stabilizing heat shock protein 47 (Hsp47). Downregulation of FBXO2, caused by DNA-PKcs-medicated phosphorylation at serine 17 and E3 ligase SKP2-mediated ubiquitination at lysine 79 and subsequent proteasomal degradation, is observed in tumor tissues compared to their parallel non-tumor tissues resected from patients with HCC. Patients whose tumors are enriched for SKP2 or Hsp47 or express low levels of FBXO2 have poor median survival compared to those whose tumors have reversed levels of SKP2, FBXO2 and Hsp47. Together, FBXO2 acts as a tumor suppressor in HCC development. The components of the SKP2-FBXO2-Hsp47 axis provide newly prognostic and therapeutic factors for anti-HCC.

cancer biology↗

bioRxiv2023

Electroadhesive hydrogel interface for prolonged mucosal theranostics

Establishing a robust and intimate mucosal interface that allows medical devices to remain within lumen-confined organs for extended periods has valuable applications, particularly for gastrointestinal (GI) theranostics. Here, we report the development of e-GLUE, an electroadhesive hydrogel interface for robust and prolonged mucosal retention following electrical activation. Notably, this novel mucosal adhesion mechanism can increase the adhesion energy of hydrogels on the mucosa by up to 30-fold and enable in vivo GI retention of e-GLUE devices for up to 30 days. Strong mucosal adhesion occurs within one minute of electrical activation, despite the presence of luminal fluid, mucus exposure, and organ motility, thereby ensuring compatibility with complex in vivo environments. In swine studies, we demonstrate the utility of e-GLUE for mucosal hemostasis, sustained local delivery of therapeutics, and intimate biosensing in the GI tract. This system can enable improved treatments for various health conditions, including gastrointestinal bleeding, inflammatory bowel disease, and diagnostic applications in the GI tract and beyond.

bioengineering↗

bioRxiv2023

Encapsulation of AAV into protein vault nanoparticles as a novel solution to gene therapy's neutralizing antibody problem

Although adeno-associated virus (AAV) has enjoyed enormous success as a delivery modality for gene therapy, it continues to suffer from the high prevalence of preexisting neutralizing antibodies in human populations, limiting who can receive potentially life-saving treatments. In this regard, AAV therapies generally also must be administered as a single dose since neutralizing antibodies develop in patients who receive the virus. Strategies for circumventing these issues remain limited. As a novel solution, we employed SpyTag-SpyCatcher molecular glue technology to facilitate packaging of AAVs inside of recombinant protein vault nanoparticles. Vaults are endogenous particles produced by mammalian cells. We therefore hypothesized that they may shield packaged molecules from neutralizing antibodies. Vaults have previously been utilized to deliver drugs and proteins into cells, but our study represents the first time anyone has packaged an entire virus inside of a vault. We showed that our vaultAAV (VAAV) delivery vehicle transduces cells in the presence of anti-AAV neutralizing serum. VAAV is positioned as a new gene therapy delivery platform with potential to overcome the neutralizing antibody problem and perhaps even allow administration of multiple doses, expanding the scope of AAV treatments.

bioengineering↗

bioRxiv2023

No trade-off between ovary activation and immune protein expression in female bumble bees (Bombus impatiens)

Evidence for a trade-off between reproduction and immunity has manifested in many animal species, including social insects. However, investigations in social insect queens present a conundrum: new gynes of many social hymenopterans, such as bumble bees and ants, must first mate, then transition from being solitary to social as they establish their nests, thus experiencing confounding shifts in environmental conditions. Worker bumble bees offer an opportunity to investigate patterns of immune protein expression associated with ovary activation while minimizing extraneous environmental factors and genetic differences. Here, we use proteomics to interrogate the patterns of immune protein expression of female bumble bees (Bombus impatiens) by 1) sampling queens at different stages of their life cycle, then 2) by sampling workers with different degrees of ovary activation. Patterns of immune protein expression in the hemolymph of queens are consistent with a reproduction-immunity trade-off, but equivalent samples from workers are not. This brings into question whether queen bumble bees really experience a reproduction-immunity trade-off, or if patterns of immune protein expression may actually be due to the selective pressure of the different environmental conditions they are exposed to during their life cycle.

physiology↗

bioRxiv2023

Bioengineering of a human innervated cardiac muscle model

Cardiac autonomic neurons control cardiac contractility. Dysregulation of the autonomic nervous system can lead to sympathetic overdrive resulting in heart failure and an increased incidence of fatal arrhythmias. Here, we introduce innervated engineered human myocardium (iEHM), a novel model of neuro-cardiac junctions, constructed by fusion of a bioengineered neural organoid (BENO) patterned to autonomic nervous system and engineered human myocardium (EHM). Projections of sympathetic neurons into engineered human myocardium formed presynaptic terminals in close proximity to cardiomyocytes and an extensive vascular network co-developing in the tissues. Contractile responses to optogenetic stimulation of the accordingly engineered neuronal component demonstrated functionality of neuro-cardiac junctions in iEHM. This model will serve as a human surrogate system to delineate neuron and cardiac cell contribution to brain and heart diseases and is an important step towards engineering a human brain to heart axis in a dish.

bioengineering↗

bioRxiv2023

In vivo expansion of gene-targeted hepatocytes through transient inhibition of an essential gene

Homology Directed Repair (HDR)-based genome editing is an approach that could permanently correct a broad range of genetic diseases. However, its utility is limited by inefficient and imprecise DNA repair mechanisms in terminally differentiated tissues. Here, we tested "Repair Drive", a novel method for improving targeted gene insertion in the liver by selectively expanding correctly repaired hepatocytes in vivo. Our system consists of transient conditioning of the liver by knocking down an essential gene, and delivery of an untargetable version of the essential gene in cis with a therapeutic transgene. We show that Repair Drive dramatically increases the percentage of correctly targeted hepatocytes, up to 25%. This resulted in a five-fold increased expression of a therapeutic transgene. Repair Drive was well-tolerated and did not induce toxicity or tumorigenesis in long term follow up. This approach will broaden the range of liver diseases that can be treated with somatic genome editing.

genetics↗

bioRxiv2022

Protein nanobarcodes enable single-step multiplexed fluorescence imaging

Multiplexed cellular imaging typically relies on the sequential application of detection probes, such as antibodies or DNA barcodes, which is complex and time-consuming. To address this, we developed here protein nanobarcodes, composed of combinations of epitopes recognized by specific sets of nanobodies. The nanobarcodes are read in a single imaging step, relying on nanobodies conjugated to distinct fluorophores, which enables a precise analysis of large numbers of protein combinations.

cell biology↗

bioRxiv2021

The Need for Transfer Learning in CRISPR-Cas Off-Target Scoring

MotivationThe scalable design of safe guide RNA sequences for CRISPR gene editing depends on the computational "scoring" of DNA locations that may be edited. As there is no widely accepted benchmark dataset to compare scoring models, we present a curated "TrueOT" dataset that contains thoroughly validated datapoints to best reflect the properties of in vivo editing. Many existing models are trained on data from high throughput assays. We hypothesize that such models may suboptimally transfer to the low throughput data in TrueOT due to fundamental biological differences between proxy assays and in vivo behavior. We developed new Siamese convolutional neural networks, trained them on a proxy dataset, and compared their performance against existing models on TrueOT. ResultsOur simplest model with a single convolutional and pooling layer surprisingly exhibits state-of-the-art performance on TrueOT. Adding subsequent layers improved performance on a proxy dataset while compromising performance on TrueOT. We demonstrate improved generalization on TrueOT with a Siamese model of higher complexity when we apply transfer learning techniques. These results suggest an urgent need for the CRISPR community to agree upon a benchmark dataset such as TrueOT and highlight that various sources of CRISPR data cannot be assumed to be equivalent. Availability and ImplementationOur code base and datasets are available on GitHub at github.com/baolab-rice/CRISPR_OT_scoring.

bioinformatics↗