MOTIVATION:Nucleic-acid-binding proteins (NBPs) are crucial regulators of gene expression, recognizing specific RNA or DNA binding sites. While high-throughput experiments have generated vast amounts of binding data, a significant challenge remains in predicting binding affinities of a novel query NBP to any nucleic-acid sequence. Current computational methods often require prior experimental data for the query NBP or are limited to predictions over predefined short RNA or DNA sequences. RESULTS:We present New Protein Binding Intensity Predictor (NPBIP), a new method to predict the binding of a query NBP to any RNA or DNA sequence by integrating two complementary components: (i) a similarity-based method that computes a weighted mean of binding predictions over the training NBPs; and (ii) a deep-learning model that combines a large protein language model with a hybrid convolutional-transformer network to predict binding directly. We trained and evaluated NPBIP on 420 RNA-binding protein (RBP) experiments and 464 DNA-binding protein (DBP) experiments. NPBIP significantly outperformed each of its components and all competing baselines, achieving a mean Pearson correlation of 0.414±0.20 and 0.581±0.22 over the RNA- and DNA-binding experiments, respectively. This prediction performance was statistically comparable to an experimental k-mer upper bound over RBPs and statistically superior to an upper bound over DBPs. Furthermore, our interpretability analysis demonstrates that NPBIP recovers canonical binding motifs for both RBPs and DBPs, providing biological validation for the model's predictions. AVAILABILITY AND IMPLEMENTATION:Source code and datasets are publicly available at https://github.com/OrensteinLab/NPBIP.

Minimizers are sampling schemes which are ubiquitous in almost any high-throughput sequencing analysis. Assuming a fixed alphabet of size σ, a minimizer is defined by two positive integers k,w and a linear order ρ on k-mers. A sequence is processed by a sliding window algorithm that chooses in each window of length w+k-1 its minimum k-mer with respect to ρ. A key characteristic of a minimizer is its density, which is the expected frequency of chosen k-mers among all k-mers in a random infinite σ-ary sequence. Minimizers of smaller density are preferred as they produce smaller samples, which lead to reduced runtime and memory usage in downstream applications. While the hardness of finding a minimizer of minimum density for given input parameters (σ,k,w) is unknown, it has a huge search space of (σ^k)! and there is no known algorithm apart from a trivial brute-force search. In this paper, we tackle the minimum density problem for minimizers. We first formulate this problem as an ILP of size Θ(wσ^(w+k)), which has worst-case solution time that is doubly-exponential in (k+w) under standard complexity assumptions. Our experiments show that an ILP solver terminates with an optimal solution only for very small k and w. We then present our main method, called OptMini, which computes an optimal minimizer in O(w⋅2^{σ^k+O(k)}) time and thus is capable of processing large w values. In experiments, OptMini works much faster than the runtime predicts due to several additional tricks shrinking the search space without harming optimality. We use OptMini to compute minimum-density minimizers for (σ,k)∈{(2,2),(2,3),(2,4),(2,5),(2,6),(4,2)} and w∈[2,3σ^k], with the exception of certain w-ranges for k=6 and the single case of k=5, w=2. Finally, we derive conclusions and insights regarding the density values as a function of w, patterns in optimal minimizer orders, and the relation between minimum-size universal hitting sets and minimum-density minimizers. ### Competing Interest Statement The authors have declared no competing interest. Israel Science Foundation, 358/21

MOTIVATION:The CRISPR-Cas9 complex has revolutionized genome-editing technologies. By designing a 20 nt-long guide RNA, a Cas9 nuclease can be guided to cleave almost any genomic target site (followed by NGG). The cleavage induces double-stranded DNA breaks, which are then repaired by cellular pathways. Accurate CRISPR-Cas9 repair-outcome prediction is essential for designing guide RNAs with desired genomic effects, such as gene knockout. A central challenge is quantifying the rate of frameshifts, i.e. repair-outcomes that lead to a change in the local length that is not a multiple of three. Previous methods for frameshift-rate prediction were trained on only a few experimental or cellular contexts, mostly relied on manually defined microhomology features, and were limited by sparse features and class labels. RESULTS:We developed CROP, a feature-independent context-aware repair-outcome prediction method. By aggregating specific repair outcomes as Δlength classes, CROP overcomes class sparsity. We designed CROP to work with variable input sequence lengths and output classes to utilize multiple datasets simultaneously. We benchmarked CROP against state-of-the-art repair-outcome prediction methods over 18 datasets, which we curated and standardized from various studies. Across all datasets, CROP outperformed all competing methods in frameshift-rate prediction. We performed cross-experiment and cross-cellular frameshift-rate predictions to investigate the generalizability of repair mechanisms. Finally, we show that CROP learned microhomology principles from raw sequences without explicit feature engineering, establishing an end-to-end architecture for CRISPR-Cas9 repair-outcome prediction that learns from multiple datasets. AVAILABILITY AND IMPLEMENTATION:CROP is available at https://github.com/OrensteinLab/CROP.

Protein-protein interactions (PPIs) are shaped by evolutionary pressures that fine-tune binding affinities and drive the epistatic relationships that support functional outcomes. Here, we used the complex of bovine pancreatic trypsin inhibitor (BPTI) and chymotrypsin as a model system to study how mutations at one or two positions affect binding affinity and epistasis. To predict the binding affinity landscape of the BPTI-chymotrypsin complex, we combined deep sequencing data, obtained from a saturation scanning mutagenesis BPTI library, with a machine-learning (ML) model. Using this ML model, which was trained on a subset of experimental binding data, we predicted the binding affinities and epistatic interactions across thousands of single and double BPTI mutants, including those not observed in the library. Our predictive approach completed missing data points and enabled us to reveal global trends in affinity changes and mutation couplings within specific binding interface positions. Our analysis revealed that different mutations in the same position may have different effects on affinity, with most double mutations leading to increased epistasis, particularly at hotspot positions, thereby indicating a cooperative binding effect. In most cases, affinity and epistasis were inversely correlated, with affinity enhancement of double-mutant variants being associated with negative epistasis. Our approach can be readily generalized to predict mutation effects in larger combinatorial libraries and in proteins for which structural information is lacking.

Deep mutational scanning (DMS) has proven effective for mapping protein-protein interactions (PPIs), but it cannot provide complete coverage of the mutation landscape, particularly for multi-mutant variants. To address this limitation, we trained machine-learning (ML) models on previously generated DMS data for a stabilized amyloid precursor protein inhibitor (APPI) binding to either of two serine proteases, mesotrypsin and kallikrein-6 (KLK6), which are implicated in various human disorders. We combined the models to accurately predict the binding selectivity of APPI variants, including double-mutant variants, for the two serine proteases. We achieved a Pearson correlation of 0.937 between predicted log2 selectivity enrichment ratios and DMS-derived values. We further validated the predictions of our combined model by yeast-surface-display measurements and inhibition assays of purified APPI variants and revealed epistatic interactions that shape protease selectivity. Guided by binding selectivity predictions, we identified highly selective APPI variants, including the most selective mesotrypsin inhibitor reported to date. Together, these findings support the use of DMS and ML as a framework for predicting PPI selectivity and prioritizing selective therapeutic protein variants.

CRISPR/Cas9 technology has revolutionized geneediting technologies. However, the endogenous gene-editing efficiency at the target site, unintended off-target sites (OTSs), and the DNA repair outcomes are affected by various factors, among them are the sequence of the site, its flanking sequences, and the epigenetic marks harboring it. Researchers have developed computational methods to predict endogenous efficiency, OTSs, and repair outcomes. However, previous studies have failed to successfully incorporate the factors that affect on- and offtarget activity, and repair outcomes, and the methods were trained and evaluated on small datasets. A recent dataset of endogenous on-target efficiencies and repair outcomes of nearly 1,600 measurements in T cells, and a recently preprocessed dataset of 78 OTS experiments in T cells provide a unique opportunity to develop a method to predict endogenous on-target efficiency, OTSs, and repair outcomes. Here, we developed Epi- CRISPROn and EpiCRISPROff to improve on-target efficiency and OTSs prediction, respectively, by combining various inputs: the site and its flanking sequences, multiple epigenetic marks, and a high-throughput-based prediction. In EpiCRISPROn and EpiCRISPROff, the additional non-sequence features improved prediction performance from an average Spearman correlation of 0.31 to 0.51 in 5-fold cross-validation and from an AUPRC of 0.436 to 0.441 on a held-out test set, respectively. Moreover, EpiCRISPROn and EpiCRISPROff were trained on one cell type and successfully generalized to other cell types. In contrast, the inclusion of flanking sequences and epigenetic features did not improve repair-outcome prediction. Furthermore, by interrogating EpiCRISPROn and EpiCRISPROff trained models we shed light on on-target and OTS editing preferences. We expect EpiCRISPROn and EpiCRISPROff to advance the field of gene editing by providing improved cell-type-aware prediction of endogenous CRISPR/Cas9 on- and off-target activity.

MOTIVATION:Predicting the RNA-seq coverage of native and exogenous sequences is central to many molecular- and synthetic-biology applications. Substantial progress has been made in developing methods to predict the RNA-seq coverage of native genomic sequences, with the recently developed Shorkie achieving state-of-the-art performance in yeast. However, prediction performance of these methods over exogenous DNA is still unknown. Recent studies measured RNA-seq coverage of large exogenous genomes in yeast, providing a unique opportunity to train machine-learning models on a large exogenous sequence space and to improve both prediction performance and our understanding of regulatory mechanisms. RESULTS:We introduce ExoShorkie, a method we developed by extending Shorkie through transfer learning across multiple exogenous RNA-seq datasets. We demonstrate that ExoShorkie significantly improves prediction performance on held-out exogenous genomes and outperforms both a native-genome-trained Shorkie baseline and Yorzoi, the only competing method for predicting exogenous RNA-seq coverage in yeast, in cross-validation and in leave-one-genome-out evaluations. Furthermore, through interpretability analyses we reveal biologically meaningful regulatory motifs and distinct regulatory rules in exogenous genomes in yeast, providing new insights into transcriptional regulation. AVAILABILITY AND IMPLEMENTATION:ExoShorkie is available at https://github.com/OrensteinLab/ExoShorkie.

CRISPR/Cas9 technology has revolutionized gene-editing technologies. By designing a single guide RNA (sgRNA) of 2Ont, one can target any genomic loci followed by NGG downstream. However, the endogenous gene-editing efficiency at the target site and unintended off-target sites (OTSs) are affected by various factors, among them are the sequence of the site, its flanking sequences, and the epigenetic marks harbouring it. Since experimentally measuring endogenous efficiency and OTSs are time- and resource -consuming tasks, researchers have developed computational methods to predict them. However, previous studies have failed to successfully incorporate the factors that affect on- or off-target activity, and were trained and evaluated on small datasets. Recently, Leenay et al. produced the most comprehensive dataset of endogenous on-target efficiencies of nearly 1,600 measurements in T cells, and Yaish et al. reprocessed the largest dataset of 78 OTS experiments in T cells, providing a unique opportunity to develop a method to predict endogenous on-target efficiency and OTSs. Here, we developed EpiCRISPROn and EpiCRISPROff to improve on-target efficiency and OTSs prediction, respectively, by combining various inputs: the site and its flanking sequences, multiple epigenetic marks, and a high-throughput-based prediction. In EpiCRISROn and EpiCRISPROff, the additional non-sequence features improved prediction performance from an average Spearman correlation of 0.31 to 0.51 in 5 -fold cross -validation and from an AUPRC of 0.436 to 0.441 on a held-out test set, respectively. Moreover, EpiCRISPROn and EpiCRISPROff were trained on one cell type and successfully generalized to other cell types. Furthermore, we shed light on on-target and OTS preference by interrogating EpiCRISPROn and EpiCRISPROff trained models. We expect EpiCRISPROn and EpiCRISPROff to advance the field of gene editing by providing improved prediction of endogenous CRISPR/Cas9 on-target efficiency and off-target activity.

MOTIVATION:Minimizers are the most popular k-mer selection scheme in algorithms and data structures analyzing high-throughput sequencing (HTS) data. In a minimizer scheme, the smallest k-mer by some predefined order is selected as the representative of a sequence window containing w consecutive k-mers, which results in overlapping windows often selecting the same k-mer. Minimizers that achieve the lowest frequency of selected k-mers over a random DNA sequence, termed the expected density, are desired for improved performance of HTS analyses. Yet, no method to date exists to generate minimizers that achieve minimum expected density. Moreover, for k and w values used by common HTS algorithms and data structures, there is a gap between densities achieved by existing selection schemes and the theoretical lower bound. RESULTS:We developed GreedyMini, a toolkit of methods to generate minimizers with low expected or particular density, to improve minimizers, to extend minimizers to larger alphabets, k, and w, and to measure the expected density of a given minimizer efficiently. We demonstrate over various combinations of k and w values, including those of popular HTS methods, that GreedyMini can generate DNA minimizers that achieve expected densities very close to the lower bound, and both expected and particular densities much lower compared to existing selection schemes. Moreover, we show that GreedyMini's k-mer rank-retrieval time is comparable to common k-mer hash functions. We expect GreedyMini to improve the performance of many HTS algorithms and data structures and advance the research of k-mer selection schemes. AVAILABILITY AND IMPLEMENTATION:The toolkit, its source code, and precomputed minimizers for a variety of (k,w) pairs are available via https://github.com/OrensteinLab/GreedyMini.

Selective inhibitors are essential for targeted therapeutics and for probing enzyme functions in various biological systems. The two main challenges in identifying such protein-based inhibitors lie in the extensive experimental effort required, including the generation of large libraries, and in tailoring the selectivity of inhibitors to enzymes with homologous structures. To address these challenges, machine learning (ML) is being used to improve protein design by training on targeted libraries and identifying key interface mutations that enhance affinity and specificity. However, such ML-based methods are limited by inaccurate energy calculations and difficulties in predicting the structural impacts of multiple mutations. Here, we present an ML-based method that leverages HTS data to streamline the design of selective protease inhibitors. To demonstrate its utility, we applied our new method to find inhibitors of matrix metalloproteinases (MMPs), a family of homologous proteases involved in both physiological and pathological processes. By training ML models on binding data for three MMPs (MMP-1, MMP-3, and MMP-9), we successfully designed a novel N-TIMP2 variant with a differential specificity profile, namely, high affinity for MMP-9, moderate affinity for MMP-3, and low affinity for MMP-1. Our experimental validation showed that this novel variant exhibited a significant specificity shift and enhanced selectivity compared with wild-type N-TIMP2. Through molecular modeling and energy minimization, we obtained structural insights into the variant's enhanced selectivity. Our findings highlight the power of ML-based methods to reduce experimental workloads, facilitate the rational design of selective inhibitors, and advance the understanding of specific inhibitor-enzyme interactions in homologous enzyme systems.

Deep neural networks have been transforming the field of bioinformatics and computational biology in recent years, especially in genomic-related tasks. Trained neural networks have enabled unprecedented capabilities in predicting molecular and genomic phenotypes. A fundamental step following deep-neural-network training and performance evaluation is the interpretation of the trained neural networks to learn new biology and validate the trained models. Techniques like integrated gradients have been highly successful in local interpretability, i.e., attributing importance to a specific residue in a given DNA, RNA, or amino acid sequence. But, there still remains the challenge of finding the global patterns that are shared among many sequences to understand the biological mechanism. Currently, TF-MoDISco is the only available method for this task. However, TF-MoDISco takes hours to run on standard datasets, and it reports many redundant and false motifs. Here, we present F-MoDA (Fourier-based Motif Discovery in Attribution maps), a novel computational method for efficiently and accurately discovering shared sequence motifs in residue-level attribution maps. F-MoDA leverages signal processing techniques and a hierarchical clustering approach to identify recurring regulatory patterns. We evaluated F-MoDA against TF-MoDISco over an established motif-finding benchmark and found that F-MoDA reports motifs that are more similar to the ground truth, in addition to reporting fewer redundant motifs and fewer false motifs. Moreover, F-MoDA runs much faster and uses less memory. We expect F-MoDA to be utilized in many studies applying deep neural networks to genomics data. F-MoDA is publicly available at https://github.com/OrensteinLab/F-MoDA. [GRAPHICS] .

DNA motif discovery and, particularly, computational modeling of transcription factor binding motifs, has been a mecca of algorithmic bioinformatics for several decades. Here, we report the results of the largest open community challenge in Inferring BInding Specificities (IBIS), where participants all over the world were invited to construct binding specificity models from multi-assay experimental data for poorly studied human transcription factors. The submissions were rigorously tested against a rich held-out dataset. Benchmarking demonstrated a consistent advantage of properly designed deep learning models over traditional positional weight matrices and other machine learning methods. Yet, the positional weight matrices displayed a surprisingly strong performance out of the box, being only slightly behind the best deep learning models. A post-challenge assessment of a selection of other deep learning methods further solidified this finding. IBIS highlights the power of benchmarking in finding adequate DNA motif representations, emphasizes the pros and cons of various machine learning methods applied to DNA motif modeling, and establishes a rich dataset, benchmarking protocols, and computational framework for a fair cross-platform evaluation of future models of transcription factor binding motifs in DNA sequences.

RNA-binding proteins (RBP) play diverse roles in mRNA processing and function. However, from thousands of RBPs encoded in the human genome, a detailed molecular understanding of their interactions with RNA is available only for a small fraction. In most cases, our knowledge of the combination of RNA sequence and structure required for specific RBP binding is insufficient for accurately predicting binding sites transcriptome-wide. In this context, the rapidly expanding collection of transcriptomic data sets that map distinct, yet intertwined posttranscriptional marks, such as RNA structure and RBP binding, presents an opportunity for integrative analysis to better characterize RBP binding. A grand challenge faced by our community is that relatively little information on the secondary structure context within and near RBP-binding sites has been gleaned from integrating such data sets, partially due to lack of suitable computational methods. To engage scientists from diverse backgrounds in addressing this gap, the RNA Society organized the RBP Footprint Grand Challenge in 2021, an international community effort to develop new methods or leverage existing ones for predicting RBP-binding sites through analysis of a growing volume of sequence, structure, and binding data and to experimentally validate select predictions. Here, we report the initiative, analyses, and methods developed by the participants, validation results, and five new in vivo binding data sets generated for validation. We hope our work will inspire additional innovation in computational methods, further utilization of available data resources, and future endeavors to engage the community in collaborating toward closing other critical data-analysis gaps.

Studies determining protein-protein interactions (PPIs) by deep mutational scanning have focused mainly on a narrow range of affinities within complexes and thus include only partial coverage of the mutation space of given proteins. By inserting an affinity-reducing N-terminal alanine in the N-terminal domain of the tissue inhibitor of metalloproteinases 2 (N-TIMP2), we overcame the limitation of its narrow affinity range for matrix metalloproteinase 9 (MMP9CAT). We trained deep neural networks (DNNs) to quantitatively predict the binding affinity of unobserved wild-type variants and variants carrying an N-terminal alanine. Good correlation was obtained between predicted and observed log2 enrichment ratio (ER) values, which also correlated with the affinity of N-TIMP2 variants to MMP9CAT. Our ability to predict affinities of unobserved N-TIMP2 variants was confirmed on an independent dataset of experimentally validated N-TIMP2 proteins. This ability is of significant importance in the field of PPI prediction and for developing therapies targeting these interactions.

The CRISPR/Cas9 system is a highly accurate geneediting technique, but it can also lead to unintended off-target sites (OTS). Consequently, many high-throughput assays have been developed to measure OTS in a genome-wide manner, and their data was used to train machine-learning models to predict OTS. However, these models are inaccurate when considering OTS with bulges due to limited data compared to OTS without bulges. Recently, CHANGE-seq, a new in vitro technique to detect OTS, was used to produce a dataset of unprecedented scale and quality. In addition, the same study produced in cellula GUIDE-seq experiments, but none of these experiments included bulges. Here, we generated the most comprehensive GUIDE-seq dataset with bulges, and trained and evaluated state-of-the-art machine-learning models that consider OTS with bulges. We first reprocessed the publicly available experimental raw data of the CHANGE-seq study to gener-ate 20 new GUIDE-seq experiments, and hundreds of OTS with bulges among the original and new GUIDE-seq experiments. We then trained multiple machine-learning models, and demonstrated their state-of-the-art performance both in vitro and in cellula overall and when focusing on OTS with bulges. Last, we visualized the key features learned by our models on OTS with bulges in a unique representation. Graphical abstract

The interaction of sclerostin (Scl) with the low-density lipoprotein receptor-related protein 4 (LRP4) leads to a marked reduction in bone formation by inhibiting the Wnt/β-catenin pathway. To characterize the Scl-LRP4 binding interface, we sorted a combinatorial library of Scl variants and isolated variants with reduced affinity to LRP4. We identified Scl single-mutation variants enriched during the sorting process and verified their reduction in affinity toward LRP4-a reduction that was not a result of changes in the variants' secondary structure or stability. We found that Scl positions K75 (loop 1) and V136 (loop 3) are critical hotspots for binding to LRP4. Our findings establish the foundation for targeting these hotspots for developing novel therapeutic strategies to promote bone formation.

We demonstrate a transcriptional regulatory design algorithm that can boost expression in yeast and mammalian cell lines. The system consists of a simplified transcriptional architecture composed of a minimal core promoter and a synthetic upstream regulatory region (sURS) composed of up to three motifs selected from a list of 41 motifs conserved in the eukaryotic lineage. The sURS system was first characterized using an oligo-library containing 189,990 variants. We validate the resultant expression model using a set of 43 unseen sURS designs. The validation sURS experiments indicate that a generic set of grammar rules for boosting and attenuation may exist in yeast cells. Finally, we demonstrate that this generic set of grammar rules functions similarly in mammalian CHO-K1 and HeLa cells. Consequently, our work provides a design algorithm for boosting the expression of promoters used for expressing industrially relevant proteins in yeast and mammalian cell lines.

High-throughput sequencing (HTS) is the most established technique to measure transcript abundance. HTS reads often contain uncertain or low-quality base calls that introduce ambiguity in determining the underlying sequence. In many applications, these unresolved nucleotides are handled by looking at the consensus sequence of all HTS reads. However, this approach is not applicable where sequence heterogeneity is of biological relevance. To gauge the biological complexity of a set of HTS reads in face of unresolved base calls, one may apply the parsimony principle, i.e., find a smallest set of sequences that cover all ambiguous reads. But, no method to date solves this problem optimally. Here, we present FiSSC, a new method to find a smallest sequence cover of a set of ambiguous reads. We first prove that the problem is NP-hard. We then present filtering steps that preserve optimal solution size, and an integer-linear-programming formulation, which together form FiSSC. We tested FiSSC on A-to-I RNA editing datasets with binary ambiguities. FiSSC outperformed all baseline methods and achieved optimal results in all but one dataset. We expect FiSSC to advance the study of sequence variation and biological complexity of ambiguous reads in various biological domains.