High-Throughput Investigation of Diverse Junction Elements in RNA Tertiary Folding.

RNAs fold into defined tertiary structures to function in critical biological processes. While quantitative models can predict RNA secondary structure stability, we are still unable to predict the thermodynamic stability of RNA tertiary structure. Here, we probe conformational preferences of diverse RNA two-way junctions to develop a predictive model for the formation of RNA tertiary structure. We quantitatively measured tertiary assembly energetics of >1,000 of RNA junctions inserted in multiple structural scaffolds to generate a "thermodynamic fingerprint" for each junction. Thermodynamic fingerprints enabled comparison of junction conformational preferences, revealing principles for how sequence influences 3-dimensional conformations. Utilizing fingerprints of junctions with known crystal structures, we generated ensembles for related junctions that predicted their thermodynamic effects on assembly formation. This work reveals sequence-structure-energetic relationships in RNA, demonstrates the capacity for diverse compensation strategies within tertiary structures, and provides a path to quantitative modeling of RNA folding energetics based on "ensemble modularity."

DNA mismatch and damage patterns revealed by single-molecule sequencing.

Mutations accumulate in the genome of every cell of the body throughout life, causing cancer and other diseases1,2. Most mutations begin as nucleotide mismatches or damage in one of the two strands of the DNA before becoming double-strand mutations if unrepaired or misrepaired3,4. However, current DNA-sequencing technologies cannot accurately resolve these initial single-strand events. Here we develop a single-molecule, long-read sequencing method (Hairpin Duplex Enhanced Fidelity sequencing (HiDEF-seq)) that achieves single-molecule fidelity for base substitutions when present in either one or both DNA strands. HiDEF-seq also detects cytosine deamination-a common type of DNA damage-with single-molecule fidelity. We profiled 134 samples from diverse tissues, including from individuals with cancer predisposition syndromes, and derive from them single-strand mismatch and damage signatures. We find correspondences between these single-strand signatures and known double-strand mutational signatures, which resolves the identity of the initiating lesions. Tumours deficient in both mismatch repair and replicative polymerase proofreading show distinct single-strand mismatch patterns compared to samples that are deficient in only polymerase proofreading. We also define a single-strand damage signature for APOBEC3A. In the mitochondrial genome, our findings support a mutagenic mechanism occurring primarily during replication. As double-strand DNA mutations are only the end point of the mutation process, our approach to detect the initiating single-strand events at single-molecule resolution will enable studies of how mutations arise in a variety of contexts, especially in cancer and ageing.

Observing a DNA polymerase choose right from wrong.

DNA polymerase (pol) ß is a model polymerase involved in gap-filling DNA synthesis utilizing two metals to facilitate nucleotidyl transfer. Previous structural studies have trapped catalytic intermediates by utilizing substrate analogs (dideoxy-terminated primer or nonhydrolysable incoming nucleotide). To identify additional intermediates during catalysis, we now employ natural substrates (correct and incorrect nucleotides) and follow product formation in real time with 15 different crystal structures. We are able to observe molecular adjustments at the active site that hasten correct nucleotide insertion and deter incorrect insertion not appreciated previously. A third metal binding site is transiently formed during correct, but not incorrect, nucleotide insertion. Additionally, long incubations indicate that pyrophosphate more easily dissociates after incorrect, compared to correct, nucleotide insertion. This appears to be coupled to subdomain repositioning that is required for catalytic activation/deactivation. The structures provide insights into a fundamental chemical reaction that impacts polymerase fidelity and genome stability.

Defining the influence of Rad51 and Dmc1 lineage-specific amino acids on genetic recombination.

The vast majority of eukaryotes possess two DNA recombinases: Rad51, which is ubiquitously expressed, and Dmc1, which is meiosis-specific. The evolutionary origins of this two-recombinase system remain poorly understood. Interestingly, Dmc1 can stabilize mismatch-containing base triplets, whereas Rad51 cannot. Here, we demonstrate that this difference can be attributed to three amino acids conserved only within the Dmc1 lineage of the Rad51/RecA family. Chimeric Rad51 mutants harboring Dmc1-specific amino acids gain the ability to stabilize heteroduplex DNA joints with mismatch-containing base triplets, whereas Dmc1 mutants with Rad51-specific amino acids lose this ability. Remarkably, RAD-51 from Caenorhabditis elegans, an organism without Dmc1, has acquired "Dmc1-like" amino acids. Chimeric C. elegans RAD-51 harboring "canonical" Rad51 amino acids gives rise to toxic recombination intermediates, which must be actively dismantled to permit normal meiotic progression. We propose that Dmc1 lineage-specific amino acids involved in the stabilization of heteroduplex DNA joints with mismatch-containing base triplets may contribute to normal meiotic recombination.

Correction of non-random mutational biases along a linear bacterial chromosome by the mismatch repair endonuclease NucS.

The linear chromosome of Streptomyces exhibits a highly compartmentalized structure with a conserved central region flanked by variable arms. As double strand break (DSB) repair mechanisms play a crucial role in shaping the genome plasticity of Streptomyces, we investigated the role of EndoMS/NucS, a recently characterized endonuclease involved in a non-canonical mismatch repair (MMR) mechanism in archaea and actinobacteria, that singularly corrects mismatches by creating a DSB. We showed that Streptomyces mutants lacking NucS display a marked colonial phenotype and a drastic increase in spontaneous mutation rate. In vitro biochemical assays revealed that NucS cooperates with the replication clamp to efficiently cleave G/T, G/G and T/T mismatched DNA by producing DSBs. These findings are consistent with the transition-shifted mutational spectrum observed in the mutant strains and reveal that NucS-dependent MMR specific task is to eliminate G/T mismatches generated by the DNA polymerase during replication. Interestingly, our data unveil a crescent-shaped distribution of the transition frequency from the replication origin towards the chromosomal ends, shedding light on a possible link between NucS-mediated DSBs and Streptomyces genome evolution.

Structure of an internal loop motif with three consecutive U•U mismatches from stem-loop 1 in the 3'-UTR of the SARS-CoV-2 genomic RNA.

The single-stranded RNA genome of SARS-CoV-2 is highly structured. Numerous helical stem-loop structures interrupted by mismatch motifs are present in the functionally important 5'- and 3'-UTRs. These mismatches modulate local helical geometries and feature unusual arrays of hydrogen bonding donor and acceptor groups. However, their conformational and dynamical properties cannot be directly inferred from chemical probing and are difficult to predict theoretically. A mismatch motif (SL1-motif) consisting of three consecutive U•U base pairs is located in stem-loop 1 of the 3'-UTR. We combined NMR-spectroscopy and MD-simulations to investigate its structure and dynamics. All three U•U base pairs feature two direct hydrogen bonds and are as stable as Watson-Crick A:U base pairs. Plasmodium falciparum 25S rRNA contains a triple U•U mismatch motif (Pf-motif) differing from SL1-motif only with respect to the orientation of the two closing base pairs. Interestingly, while the geometry of the outer two U•U mismatches was identical in both motifs the preferred orientation of the central U•U mismatch was different. MD simulations and potassium ion titrations revealed that the potassium ion-binding mode to the major groove is connected to the different preferred geometries of the central base pair in the two motifs.

The clade-specific target recognition mechanisms of plant RISCs.

Eukaryotic Argonaut proteins (AGOs) assemble RNA-induced silencing complexes (RISCs) with guide RNAs that allow binding to complementary RNA sequences and subsequent silencing of target genes. The model plant Arabidopsis thaliana encodes 10 different AGOs, categorized into three distinct clades based on amino acid sequence similarity. While clade 1 and 2 RISCs are known for their roles in post-transcriptional gene silencing, and clade 3 RISCs are associated with transcriptional gene silencing in the nucleus, the specific mechanisms of how RISCs from each clade recognize their targets remain unclear. In this study, I conducted quantitative binding analyses between RISCs and target nucleic acids with mismatches at various positions, unveiling distinct target binding characteristics unique to each clade. Clade 1 and 2 RISCs require base pairing not only in the seed region but also in the 3' supplementary region for stable target RNA binding, with clade 1 exhibiting a higher stringency. Conversely, clade 3 RISCs tolerate dinucleotide mismatches beyond the seed region. Strikingly, they bind to DNA targets with an affinity equal to or surpassing that of RNA, like prokaryotic AGO complexes. These insights challenge existing views on plant RNA silencing and open avenues for exploring new functions of eukaryotic AGOs.

Nucleosomes around a mismatched base pair are excluded via an Msh2-dependent reaction with the aid of SNF2 family ATPase Smarcad1.

Post-replicative correction of replication errors by the mismatch repair (MMR) system is critical for suppression of mutations. Although the MMR system may need to handle nucleosomes at the site of chromatin replication, how MMR occurs in the chromatin environment remains unclear. Here, we show that nucleosomes are excluded from a >1-kb region surrounding a mismatched base pair in Xenopus egg extracts. The exclusion was dependent on the Msh2-Msh6 mismatch recognition complex but not the Mlh1-containing MutL homologs and counteracts both the HIRA- and CAF-1 (chromatin assembly factor 1)-mediated chromatin assembly pathways. We further found that the Smarcad1 chromatin remodeling ATPase is recruited to mismatch-carrying DNA in an Msh2-dependent but Mlh1-independent manner to assist nucleosome exclusion and that Smarcad1 facilitates the repair of mismatches when nucleosomes are preassembled on DNA. In budding yeast, deletion of FUN30, the homolog of Smarcad1, showed a synergistic increase of spontaneous mutations in combination with MSH6 or MSH3 deletion but no significant increase with MSH2 deletion. Genetic analyses also suggested that the function of Fun30 in MMR is to counteract CAF-1. Our study uncovers that the eukaryotic MMR system has an ability to exclude local nucleosomes and identifies Smarcad1/Fun30 as an accessory factor for the MMR reaction.

A mechanistic study on the tolerance of PAM distal end mismatch by SpCas9.

The therapeutic application of CRISPR-Cas9 is limited due to its off-target activity. To have a better understanding of this off-target effect, we focused on its mismatch-prone PAM distal end. The off-target activity of SpCas9 depends directly on the nature of mismatches, which in turn results in deviation of the active site of SpCas9 due to structural instability in the RNA-DNA duplex strand. In order to test the hypothesis, we designed an array of mismatched target sites at the PAM distal end and performed in vitro and cell line-based experiments, which showed a strong correlation for Cas9 activity. We found that target sites having multiple mismatches in the 18th to 15th position upstream of the PAM showed no to little activity. For further mechanistic validation, Molecular Dynamics simulations were performed, which revealed that certain mismatches showed elevated root mean square deviation values that can be attributed to conformational instability within the RNA-DNA duplex. Therefore, for successful prediction of the off-target effect of SpCas9, along with complementation-derived energy, the RNA-DNA duplex stability should be taken into account.

A Seed Mismatch Enhances Argonaute2-Catalyzed Cleavage and Partially Rescues Severely Impaired Cleavage Found in Fish.

The RNAi pathway provides both innate immunity and efficient gene-knockdown tools in many eukaryotic species, but curiously not in zebrafish. We discovered that RNAi is less effective in zebrafish at least partly because Argonaute2-catalyzed mRNA slicing is impaired. This defect is due to two mutations that arose in an ancestor of most teleost fish, implying that most fish lack effective RNAi. Despite lacking efficient slicing activity, these fish have retained the ability to produce miR-451, a microRNA generated by a cleavage reaction analogous to slicing. This ability is due to a G-G mismatch within the fish miR-451 precursor, which substantially enhances its cleavage. An analogous G-G mismatch (or sometimes also a G-A mismatch) enhances target slicing, despite disrupting seed pairing important for target binding. These results provide a strategy for restoring RNAi to zebrafish and reveal unanticipated opposing effects of a seed mismatch with implications for mechanism and guide-RNA design.

DNA mismatch repair protects the genome from oxygen-induced replicative mutagenesis.

DNA mismatch repair (MMR) corrects mismatched DNA bases arising from multiple sources including polymerase errors and base damage. By detecting spontaneous mutagenesis using whole genome sequencing of cultured MMR deficient human cell lines, we show that a primary role of MMR is the repair of oxygen-induced mismatches. We found an approximately twofold higher mutation rate in MSH6 deficient DLD-1 cells or MHL1 deficient HCT116 cells exposed to atmospheric conditions as opposed to mild hypoxia, which correlated with oxidant levels measured using electron paramagnetic resonance spectroscopy. The oxygen-induced mutations were dominated by T to C base substitutions and single T deletions found primarily on the lagging strand. A broad sequence context preference, dependence on replication timing and a lack of transcriptional strand bias further suggested that oxygen-induced mutations arise from polymerase errors rather than oxidative base damage. We defined separate low and high oxygen-specific MMR deficiency mutation signatures common to the two cell lines and showed that the effect of oxygen is observable in MMR deficient cancer genomes, where it best correlates with the contribution of mutation signature SBS21. Our results imply that MMR corrects oxygen-induced genomic mismatches introduced by a replicative process in proliferating cells.

Role of helical structure and dynamics in oligoadenylate synthetase 1 (OAS1) mismatch tolerance and activation by short dsRNAs.

The 2'-5'-oligoadenylate synthetases (OAS) are innate immune sensors of cytosolic double-stranded RNA (dsRNA) that play a critical role in limiting viral infection. How these proteins are able to avoid aberrant activation by cellular RNAs is not fully understood, but adenosine-to-inosine (A-to-I) editing has been proposed to limit accumulation of endogenous RNAs that might otherwise cause stimulation of the OAS/RNase L pathway. Here, we aim to uncover whether and how such sequence modifications can restrict the ability of short, defined dsRNAs to activate the single-domain form of OAS, OAS1. Unexpectedly, we find that all tested inosine-containing dsRNAs have an increased capacity to activate OAS1, whether in a destabilizing (I•U) or standard Watson-Crick-like base pairing (I-C) context. Additional variants with strongly destabilizing A•C mismatches or stabilizing G-C pairs also exhibit increased capacity to activate OAS1, eliminating helical stability as a factor in the relative ability of the dsRNAs to activate OAS1. Using thermal difference spectra and molecular dynamics simulations, we identify both increased helical dynamics and specific local changes in helical structure as important factors in the capacity of short dsRNAs to activate OAS1. These helical features may facilitate more ready adoption of the distorted OAS1-bound conformation or stabilize important structures to predispose the dsRNA for optimal binding and activation of OAS1. These studies thus reveal the molecular basis for the greater capacity of some short dsRNAs to activate OAS1 in a sequence-independent manner.

Structural basis for reduced ribosomal A-site fidelity in response to P-site codon-anticodon mismatches.

Rapid and accurate translation is essential in all organisms to produce properly folded and functional proteins. mRNA codons that define the protein-coding sequences are decoded by tRNAs on the ribosome in the aminoacyl (A) binding site. The mRNA codon and the tRNA anticodon interaction is extensively monitored by the ribosome to ensure accuracy in tRNA selection. While other polymerases that synthesize DNA and RNA can correct for misincorporations, the ribosome is unable to correct mistakes. Instead, when a misincorporation occurs, the mismatched tRNA-mRNA pair moves to the peptidyl (P) site and, from this location, causes a reduction in the fidelity at the A site, triggering post-peptidyl transfer quality control. This reduced fidelity allows for additional incorrect tRNAs to be accepted and for release factor 2 (RF2) to recognize sense codons, leading to hydrolysis of the aberrant peptide. Here, we present crystal structures of the ribosome containing a tRNALys in the P site with a U•U mismatch with the mRNA codon. We find that when the mismatch occurs in the second position of the P-site codon-anticodon interaction, the first nucleotide of the A-site codon flips from the mRNA path to engage highly conserved 16S rRNA nucleotide A1493 in the decoding center. We propose that this mRNA nucleotide mispositioning leads to reduced fidelity at the A site. Further, this state may provide an opportunity for RF2 to initiate premature termination before erroneous nascent chains disrupt the cellular proteome.

Role of condensates in modulating DNA repair pathways and its implication for chemoresistance.

For cells, it is important to repair DNA damage, such as double-strand and single-strand DNA breaks, because unrepaired DNA can compromise genetic integrity, potentially leading to cell death or cancer. Cells have multiple DNA damage repair pathways that have been the subject of detailed genetic, biochemical, and structural studies. Recently, the scientific community has started to gain evidence that the repair of DNA double-strand breaks may occur within biomolecular condensates and that condensates may also contribute to DNA damage through concentrating genotoxic agents used to treat various cancers. Here, we summarize key features of biomolecular condensates and note where they have been implicated in the repair of DNA double-strand breaks. We also describe evidence suggesting that condensates may be involved in the repair of other types of DNA damage, including single-strand DNA breaks, nucleotide modifications (e.g., mismatch and oxidized bases), and bulky lesions, among others. Finally, we discuss old and new mysteries that could now be addressed considering the properties of condensates, including chemoresistance mechanisms.

Rationally Designed Dual Base Pair Mismatch Enables Toehold-Mediated Strand Displacement to Efficiently Recognize Single-Nucleotide Polymorphism without Enzymes.

The efficiency of the enzyme-free toehold-mediated strand displacement (TMSD) technique is often insufficient to detect single-nucleotide polymorphism (SNP) that possesses only single base pair mismatch discrimination. Here, we report a novel dual base pair mismatch strategy enabling TMSD biosensing for SNP detection under enzyme-free conditions when coupled with catalytic hairpin assembly (CHA) and fluorescence resonance energy transfer (FRET). The strategy is based on a competitive strand displacement reaction mechanism, affected by the thermodynamic stability originating from rationally designed dual base pair mismatch, for the specific recognition of mutant-type DNA. In particular, enzyme-free nucleic acid circuits, such as CHA, emerge as a powerful method for signal amplification. Eventually, the signal transduction of this proposed biosensor was determined by FRET between streptavidin-coated 605 nm emission quantum dots (605QDs, donor) and Cy5/biotin hybridization (acceptor, from CHA) when incubated with each other. The proposed biosensor displayed high sensitivity to the mutant target (MT) with a detection concentration down to 4.3 fM and led to high discrimination factors for all types of mismatches in multiple sequence contexts. As such, the application of this proposed biosensor to investigate mechanisms of the competitive strand displacement reaction further illustrates the versatility of our dual base pair mismatch strategy, which can be utilized for the creation of a new class of biosensors.

7-Deazapurine and Pyrimidine Nucleoside and Oligonucleotide Cycloadducts Formed by Inverse Diels-Alder Reactions with 3,6-Di(pyrid-2-yl)-1,2,4,5-tetrazine: Ethynylated and Vinylated Nucleobases for Functionalization and Impact of Pyridazine Adducts on DNA Base Pair Stability and Mismatch Discrimination.

The manuscript reports on 7-deazapurine and pyrimidine nucleoside and oligonucleotide cycloadducts formed by the inverse electron demand Diels-Alder (iEDDA) reaction with 3,6-di(pyrid-2-yl)-1,2,4,5-tetrazine. Cycloadducts were constructed from ethynylated and vinylated nucleobases. Oligonucleotides were synthesized containing iEDDA modifications, and the impact on duplex stability was investigated. iEDDA reactions were performed on nucleoside triple bond side chains. Oxidation was not required in these cases as dihydropyridazine intermediates are not formed. In contrast, oxidation is necessary for reactions performed on alkenyl compounds. This was verified on 5-vinyl-2'-deoxyuridine. A diastereomeric mixture of 1,2-dihydropyridazine cycloadduct intermediates was isolated, characterized, and later oxidized. 12-mer oligonucleotides containing 1,2-pyridazine inverse Diels-Alder cycloadducts and their precursors were hybridized to short DNA duplexes. For that, a series of phosphoramidites was prepared. DNA duplexes with 7-functionalized 7-deazaadenines and 5-functionalized pyrimidines display high duplex stability when spacer units are present between nucleobases and pyridazine cycloadducts. A direct connectivity of the pyridazine moiety to nucleobases as reported for metabolic labeling of vinyl nucleosides reduced duplex stability strongly. Oligonucleotides bearing linkers with and without pyridazine cycloadducts attached to the 7-deazaadenine nucleobase significantly reduced mismatch formation with dC and dG.

Isomerization Pathways of a Mismatched Base Pair of A:8OG in Free Duplex DNA.

The A:8OG base pair (bp) is the outcome of DNA replication of the mismatched C:8OG bp. A high A:8OG bp population increases the C/G to A/T transversion mutation, which is responsible for various diseases. MutY is an important enzyme in the error-proof cycle and reverts A:8OG to C:8OG bp by cleaving adenine from the A:8OG bp. Several X-ray crystallography studies have determined the structure of MutY during the lesion scanning and lesion recognition stages. Interestingly, glycosidic bond (χ) angles of A:8OG bp in those two lesion recognition structures were found to differ, which implies that χ-torsion isomerization should occur during the lesion recognition process. In this study, as a first step to understanding this isomerization process, we characterized the intrinsic dynamic features of A:8OG in free DNAs by a free energy landscape simulation at the all-atom level. In this study, four isomerization states were assigned in the order of abundance: Aanti:8OGsyn > Aanti:8OGanti > Asyn:8OGanti ≈ Asyn:8OGsyn. Of these bp states, only 8OG in Asyn:8OGanti was located in the extrahelical space, whereas the purine bases (A and 8OG) in the other bp states remained inside the DNA helix. Also, free energy landscapes showed that the isomerization processes connecting these four bp states proceeded mostly in the intrahelical space via successive single glycosidic bond rotations of either A or 8OG.

Dynamic basis for dG•dT misincorporation via tautomerization and ionization.

Tautomeric and anionic Watson-Crick-like mismatches have important roles in replication and translation errors through mechanisms that are not fully understood. Here, using NMR relaxation dispersion, we resolve a sequence-dependent kinetic network connecting G•T/U wobbles with three distinct Watson-Crick mismatches: two rapidly exchanging tautomeric species (Genol•T/UG•Tenol/Uenol; population less than 0.4%) and one anionic species (G•T-/U-; population around 0.001% at neutral pH). The sequence-dependent tautomerization or ionization step was inserted into a minimal kinetic mechanism for correct incorporation during replication after the initial binding of the nucleotide, leading to accurate predictions of the probability of dG•dT misincorporation across different polymerases and pH conditions and for a chemically modified nucleotide, and providing mechanisms for sequence-dependent misincorporation. Our results indicate that the energetic penalty for tautomerization and/or ionization accounts for an approximately 10-2 to 10-3-fold discrimination against misincorporation, which proceeds primarily via tautomeric dGenol•dT and dG•dTenol, with contributions from anionic dG•dT- dominant at pH 8.4 and above or for some mutagenic nucleotides.

Double base mismatches mediated catalytic hairpin assembly for enzyme-free single-base mutation detection: integrating signal recognition and amplification in one.

Single nucleotide polymorphism (SNP) biosensors are emerging rapidly for their promising applications in human disease prevention diagnosis, treatment, and prognosis. However, it remains a bottleneck in equipping simple and stable biosensors with the traits of high sensitivity, non-enzyme, and low cost. Double base mismatches mediated chain displacement reactions have attracted fascinating advantages of tailorable thermodynamics stability, non-enzyme, and excellent assembly compliance to involvement in SNP identification. As the base mismatch position and amount in DNA sequence can be artificially adjusted, it provides plenty of selectivity and specificity for exploring perfect biosensors. Herein, a biosensor with double base mismatches mediated catalytic hairpin assembly (CHA) is designed via one base mismatch in the toehold domain and the other base mismatch in the stem sequence of hairpin 1 (H1) by triggering CHA reaction to achieve selective amplification of the mutation target (MT) and fluorescence resonance energy transfer (FRET) effect that is composed of Cy3 and Cy5 terminally attached H1 and hairpin 2 (H2). Depending on the rationally designed base mismatch position and toehold length, the fabricated biosensors show superior SNP detection performance, exhibiting a good linearity with high sensitivity of 6.6 fM detection limit and a broad detection abundance of 1%. The proposed biosensor can be used to detect the KRAS mutation gene in real samples and obtain good recoveries between 106 and 116.99%. Remarkably, these extendible designs of base mismatches can be used for more types of SNP detection, providing flexible adjustment based on base mismatch position and toehold length variations, especially for their thermodynamic model for DNA-strand displacement reactions.

Crispr-SGRU: Prediction of CRISPR/Cas9 Off-Target Activities with Mismatches and Indels Using Stacked BiGRU.

CRISPR/Cas9 is a popular genome editing technology, yet its clinical application is hindered by off-target effects. Many deep learning-based methods are available for off-target prediction. However, few can predict off-target activities with insertions or deletions (indels) between single guide RNA and DNA sequence pairs. Additionally, the analysis of off-target data is challenged due to a data imbalance issue. Moreover, the prediction accuracy and interpretability remain to be improved. Here, we introduce a deep learning-based framework, named Crispr-SGRU, to predict off-target activities with mismatches and indels. This model is based on Inception and stacked BiGRU. It adopts a dice loss function to solve the inherent imbalance issue. Experimental results show our model outperforms existing methods for off-target prediction in terms of accuracy and robustness. Finally, we study the interpretability of this model through Deep SHAP and teacher-student-based knowledge distillation, and find it can provide meaningful explanations for sequence patterns regarding off-target activity.