Biology subjects

Ozden, F.

Publications and source records attributed to Ozden, F..

7 recordsLinked to original sources

bioRxiv2026

Probabilistic Modelling of Base Editing Efficiency and Bystander Outcome Composition

Abstract. Base editors install single nucleotide substitutions without a double strand break or donor template, but editing at multiple susceptible positions can generate bystander alleles. Predicting the frequency of an intended outcome therefore requires both editing efficiency and the composition of edited products. Here we present crispAIBE, a probabilistic framework that uses separately trained transformer convolutional networks to predict Dirichlet distributions for editing efficiency and conditional allele composition. Split-conformal density-level regions are calibrated on data reserved from model selection, with composition thresholds fitted by outcome-set size. On held-out data, crispAIBE achieves Pearson correlations of 0.669 for efficiency and 0.896 for composition on measured allele sets, with empirical coverages of 0.902 and 0.906 at a nominal level of 0.90. On the published CRISPRon-BE benchmark, with test protospacers excluded from training, crispAIBE ranks first in seven of eight comparisons under the two-column Pearson statistic and first or second in seven of eight under the rank statistic. Within predicted-efficiency quintiles, the largest- and smallest-region deciles differ 1.9-fold in mean absolute error on average, compared with 1.7-fold for a boundary-distance ranking. When evaluated on U2OS, which was excluded from training, larger predictive regions were associated with larger absolute prediction errors. Local calibration with 100 U2OS protospacers yields median empirical coverage of 0.918 at the nominal 0.90 level. The crispAIBE repository is https://github.com/furkanozdenn/crispAIBE.

bioinformatics↗

bioRxiv2025

Probabilistic Modelling of Prime Editing Variant CorrectionEfficiency

Prime editing installs precise substitutions, insertions, and deletions without double-strand breaks or donor templates, but the efficiency of an individual pegRNA design is hard to predict in advance, and existing tools give only point estimates, leaving designers unable to judge which predictions to trust. We present crispAIPE, a transformer-based probabilistic framework that quantifies per pegRNA-target pair uncertainty by modelling the three competing prime-editing outcomes on the 2-simplex with a Dirichlet likelihood and pairing the posterior with split-conformal highest-density regions calibrated on a held-out fold, giving finite-sample coverage guarantees on the simplex without requiring the Dirichlet model itself to be well-calibrated. Trained on 92,423 PRIDICT Library-1 pegRNAs under mutation-level target-disjoint splitting, crispAIPE attains Spearman{rho} = 0.835, 0.843, and 0.693 on the edited, unedited, and indel fractions (Pearson r = 0.845, 0.855, 0.659), and its conformal regions match nominal coverage where region-construction baselines do not. pegRNA architecture and edit context, in particular GC content of the mutated reverse transcription template (RTT) and edit size for deletions, are associated with prediction uncertainty. Reusing the Library-1 calibration quantile unchanged on PRIDICT Library-2, the conformal region area generalises as an actionable filter for cross-cell-type design, identifying the pegRNAs whose Library-1 predictions transfer best while small-sample head fine-tuning further improves accuracy. Tool and trained models: https://github.com/furkanozdenn/pe-uncert.

bioinformatics↗

bioRxiv2023

RNAGEN: A generative adversarial network-based model to generate synthetic RNA sequences to target proteins

RNA - protein binding plays an important role in regulating protein activity by affecting localization and stability. While proteins are usually targeted via small molecules or other proteins, easy-to-design and synthesize small RNAs are a rather unexplored and promising venue. The problem is the lack of methods to generate RNA molecules that have the potential to bind to certain proteins. Here, we propose a method based on generative adversarial networks (GAN) that learn to generate short RNA sequences with natural RNA-like properties such as secondary structure and free energy. Using an optimization technique, we fine-tune these sequences to have them bind to a target protein. We use RNA-protein binding prediction models from the literature to guide the model. We show that even if there is no available guide model trained specifically for the target protein, we can use models trained for similar proteins, such as proteins from the same family, to successfully generate a binding RNA molecule to the target protein. Using this approach, we generated piRNAs that are tailored to bind to SOX2 protein using models trained for its relative (SOX10, SOX14, and SOX8) and experimentally validated in vitro that the top-2 molecules we generated specifically bind to SOX2.

bioinformatics↗

bioRxiv2023

Quantifying uncertainty in off-target activity for CRISPR guide RNAs

CRISPR-based genome editing technologies have revolutionised the field of molecular biology, offering unprecedented opportunities for precise genetic manipulation. However, off-target effects remain a significant challenge, potentially leading to unintended consequences and limiting the applicability of CRISPR-based genome editing technologies in clinical settings. Current literature predominantly focuses on point predictions for off-target activity, which may not fully capture the range of possible outcomes and associated risks. Here, we present crispAI, a neural network architecture-based approach for predicting uncertainty estimates for off-target cleavage activity, providing a more comprehensive risk assessment and facilitating improved decision-making in single guide RNA (sgRNA) design. Our approach makes use of the count noise model Zero Inflated Negative Binomial (ZINB) to model the uncertainty in the off-target cleavage activity data. In addition, we present the first-of-its-kind genome-wide sgRNA efficiency score, crispAI-aggregate, enabling prioritization among sgRNAs with similar point aggregate predictions by providing richer information compared to existing aggregate scores. We show that uncertainty estimates of our approach are calibrated and its predictive performance is superior to state-of-the-art in silico off-target cleavage activity prediction methods.

bioinformatics↗

bioRxiv2023

Learning to Generate 5' UTR Sequences for Optimized Ribosome Load and Gene Expression

The 5 untranslated region (5 UTR) of mRNA is crucial for the molecules translatability and stability, making it essential for designing synthetic biological circuits for high and stable protein expression. Several UTR sequences are patented and widely used in laboratories. This paper presents UTRGAN, a Generative Adversarial Network (GAN)-based model for generating 5 UTR sequences, coupled with an optimization procedure to ensure high expression for target gene sequences or high ribosome load and translation efficiency. The model generates sequences mimicking various properties of natural UTR sequences and optimizes them to achieve (i) up to 5-fold higher average expression on target genes, (ii) up to 2-fold higher mean ribosome load, and (iii) a 34-fold higher average translation efficiency compared to initial UTR sequences. UTRGAN-generated sequences also exhibit higher similarity to known regulatory motifs in regions such as internal ribosome entry sites, upstream open reading frames, G-quadruplexes, and Kozak and initiation start codon regions. In-vitro experiments show that the UTR sequences designed by UTRGAN result in a higher translation rate for the human TNF- protein compared to the human Beta Globin 5 UTR, a UTR with high production capacity.

bioinformatics↗

bioRxiv2022

ECOLE: Learning to call copy number variants on whole exome sequencing data

Copy number variants (CNV) are shown to contribute to the etiology of several genetic disor{-}ders. Accurate detection of CNVs on whole exome sequencing (WES) data has been a long sought-after goal for use in clinics. This was not possible despite recent improvements in performance because algo{-}rithms mostly suffer from low precision and even lower recall on expert-curated gold standard call sets. Here, we present a deep learning-based somatic and germline CNV caller for WES data, named ECOLE. Based on a variant of the transformer architecture, the model learns to call CNVs per exon, using high-confidence calls made on matched WGS samples. We further train and fine-tune the model with a small set of expert calls via transfer learning. We show that ECOLE achieves high performance on human ex{-}pert labeled data for the first time with 68.7% precision and 49.6% recall. This corresponds to precision and recall improvements of 18.7% and 30.8% over the next best-performing methods, respectively. We also show that the same fine-tuning strategy using tumor samples enables ECOLE to detect RT-qPCR validated variations in bladder cancer samples without the need for a control sample. ECOLE is available at https://github.com/ciceklab/ECOLE.

bioinformatics↗

bioRxiv2020

Polishing Copy Number Variant Calls on Exome Sequencing Data via Deep Learning

Accurate and efficient detection of copy number variants (CNVs) is of critical importance due to their significant association with complex genetic diseases. Although algorithms that use whole genome sequencing (WGS) data provide stable results with mostly-valid statistical assumptions, copy number detection on whole exome sequencing (WES) data shows comparatively lower accuracy. This is unfortunate as WES data is cost efficient, compact and is relatively ubiquitous. The bottleneck is primarily due to non-contiguous nature of the targeted capture: biases in targeted genomic hybridization, GC content, targeting probes, and sample batching during sequencing. Here, we present a novel deep learning model, DECoNT, which uses the matched WES and WGS data and learns to correct the copy number variations reported by any off-the-shelf WES-based germline CNV caller. We train DECoNT on the 1000 Genomes Project data, and we show that we can efficiently triple the duplication call precision and double the deletion call precision of the state-of-the-art algorithms. We also show that our model consistently improves the performance independent from (i) sequencing technology, (ii) exome capture kit and (iii) CNV caller. Using DECoNT as a universal exome CNV call polisher has the potential to improve the reliability of germline CNV detection on WES data sets.

bioinformatics↗