DNABERT: pre-trained Bidirectional Encoder Representations from Transformers model for DNA-language in genome

Genomic language model predicts protein co-regulation and function

Deciphering the relationship between a gene and its genomic context is fundamental to understanding and engineering biological systems. Machine learning has shown promise in learning latent relationships underlying the sequence-structure-function paradigm from massive protein sequence datasets. However, to date, limited attempts have been made in extending this continuum to include higher order genomic context information. Evolutionary processes dictate the specificity of genomic contexts in which a gene is found across phylogenetic distances, and these emergent genomic patterns can be leveraged to uncover functional relationships between gene products. Here, we train a genomic language model (gLM) on millions of metagenomic scaffolds to learn the latent functional and regulatory relationships between genes. gLM learns contextualized protein embeddings that capture the genomic context as well as the protein sequence itself, and encode biologically meaningful and functionally relevant information (e.g. enzymatic function, taxonomy). Our analysis of the attention patterns demonstrates that gLM is learning co-regulated functional modules (i.e. operons). Our findings illustrate that gLM's unsupervised deep learning of the metagenomic corpus is an effective and promising approach to encode functional semantics and regulatory syntax of genes in their genomic contexts and uncover complex relationships between genes in a genomic region.

Species-aware DNA language models capture regulatory elements and their evolution

BACKGROUND

The rise of large-scale multi-species genome sequencing projects promises to shed new light on how genomes encode gene regulatory instructions. To this end, new algorithms are needed that can leverage conservation to capture regulatory elements while accounting for their evolution.

RESULTS

Here, we introduce species-aware DNA language models, which we trained on more than 800 species spanning over 500 million years of evolution. Investigating their ability to predict masked nucleotides from context, we show that DNA language models distinguish transcription factor and RNA-binding protein motifs from background non-coding sequence. Owing to their flexibility, DNA language models capture conserved regulatory elements over much further evolutionary distances than sequence alignment would allow. Remarkably, DNA language models reconstruct motif instances bound in vivo better than unbound ones and account for the evolution of motif sequences and their positional constraints, showing that these models capture functional high-order sequence and evolutionary context. We further show that species-aware training yields improved sequence representations for endogenous and MPRA-based gene expression prediction, as well as motif discovery.

CONCLUSIONS

Collectively, these results demonstrate that species-aware DNA language models are a powerful, flexible, and scalable tool to integrate information from large compendia of highly diverged genomes.

A BERT-based model for the prediction of lncRNA subcellular localization in Homo sapiens

Understanding the subcellular localization of lncRNAs is crucial for comprehending their regulation activities. The conventional detection of lncRNA subcellular location usually uses in situ detection techniques, which are resource intensive. Some machine learning-based algorithms have been proposed for lncRNA subcellular location prediction in mammals. However, due to the low level of conservation of lncRNA sequence, the performance of cross-species models remains unsatisfactory. In this study, we curated a novel dataset containing subcellular location information of lncRNAs in Homo sapiens. Subsequently, based on the BERT pre-trained language algorithm, we developed a model for lncRNA subcellular location prediction. Our model achieved a micro-average area under the receiver operating characteristic (AUROC) of 0.791 on the training set and an AUROC of 0.700 on the testing nucleus set. Additionally, we conducted cross-species validation and motif discovery to further investigate underlying patterns. In summary, our study provides valuable guidance and computational analysis tools for exploring the mechanisms of lncRNA subcellular localization and the dynamic spatial changes of RNA in abnormal physiological states.

Effect of tokenization on transformers for biological sequences

MOTIVATION

Deep-learning models are transforming biological research, including many bioinformatics and comparative genomics algorithms, such as sequence alignments, phylogenetic tree inference, and automatic classification of protein functions. Among these deep-learning algorithms, models for processing natural languages, developed in the natural language processing (NLP) community, were recently applied to biological sequences. However, biological sequences are different from natural languages, such as English, and French, in which segmentation of the text to separate words is relatively straightforward. Moreover, biological sequences are characterized by extremely long sentences, which hamper their processing by current machine-learning models, notably the transformer architecture. In NLP, one of the first processing steps is to transform the raw text to a list of tokens. Deep-learning applications to biological sequence data mostly segment proteins and DNA to single characters. In this work, we study the effect of alternative tokenization algorithms on eight different tasks in biology, from predicting the function of proteins and their stability, through nucleotide sequence alignment, to classifying proteins to specific families.

RESULTS

We demonstrate that applying alternative tokenization algorithms can increase accuracy and at the same time, substantially reduce the input length compared to the trivial tokenizer in which each character is a token. Furthermore, applying these tokenization algorithms allows interpreting trained models, taking into account dependencies among positions. Finally, we trained these tokenizers on a large dataset of protein sequences containing more than 400 billion amino acids, which resulted in over a 3-fold decrease in the number of tokens. We then tested these tokenizers trained on large-scale data on the above specific tasks and showed that for some tasks it is highly beneficial to train database-specific tokenizers. Our study suggests that tokenizers are likely to be a critical component in future deep-network analysis of biological sequence data.

AVAILABILITY AND IMPLEMENTATION

Code, data, and trained tokenizers are available on https://github.com/technion-cs-nlp/BiologicalTokenizers.

Coding genomes with gapped pattern graph convolutional network

MOTIVATION

Genome sequencing technologies reveal a huge amount of genomic sequences. Neural network-based methods can be prime candidates for retrieving insights from these sequences because of their applicability to large and diverse datasets. However, the highly variable lengths of genome sequences severely impair the presentation of sequences as input to the neural network. Genetic variations further complicate tasks that involve sequence comparison or alignment.

RESULTS

Inspired by the theory and applications of "spaced seeds," we propose a graph representation of genome sequences called "gapped pattern graph." These graphs can be transformed through a Graph Convolutional Network to form lower-dimensional embeddings for downstream tasks. On the basis of the gapped pattern graphs, we implemented a neural network model and demonstrated its performance on diverse tasks involving microbe and mammalian genome data. Our method consistently outperformed all the other state-of-the-art methods across various metrics on all tasks, especially for the sequences with limited homology to the training data. In addition, our model was able to identify distinct gapped pattern signatures from the sequences.

AVAILABILITY AND IMPLEMENTATION

The framework is available at https://github.com/deepomicslab/GCNFrame.

Predicting TCR sequences for unseen antigen epitopes using structural and sequence features

T-cell receptor (TCR) recognition of antigens is fundamental to the adaptive immune response. With the expansion of experimental techniques, a substantial database of matched TCR-antigen pairs has emerged, presenting opportunities for computational prediction models. However, accurately forecasting the binding affinities of unseen antigen-TCR pairs remains a major challenge. Here, we present convolutional-self-attention TCR (CATCR), a novel framework tailored to enhance the prediction of epitope and TCR interactions. Our approach utilizes convolutional neural networks to extract peptide features from residue contact matrices, as generated by OpenFold, and a transformer to encode segment-based coded sequences. We introduce CATCR-D, a discriminator that can assess binding by analyzing the structural and sequence features of epitopes and CDR3-β regions. Additionally, the framework comprises CATCR-G, a generative module designed for CDR3-β sequences, which applies the pretrained encoder to deduce epitope characteristics and a transformer decoder for predicting matching CDR3-β sequences. CATCR-D achieved an AUROC of 0.89 on previously unseen epitope-TCR pairs and outperformed four benchmark models by a margin of 17.4%. CATCR-G has demonstrated high precision, recall and F1 scores, surpassing 95% in bidirectional encoder representations from transformers score assessments. Our results indicate that CATCR is an effective tool for predicting unseen epitope-TCR interactions. Incorporating structural insights enhances our understanding of the general rules governing TCR-epitope recognition significantly. The ability to predict TCRs for novel epitopes using structural and sequence information is promising, and broadening the repository of experimental TCR-epitope data could further improve the precision of epitope-TCR binding predictions.

From tradition to innovation: conventional and deep learning frameworks in genome annotation

Following the milestone success of the Human Genome Project, the 'Encyclopedia of DNA Elements (ENCODE)' initiative was launched in 2003 to unearth information about the numerous functional elements within the genome. This endeavor coincided with the emergence of numerous novel technologies, accompanied by the provision of vast amounts of whole-genome sequences, high-throughput data such as ChIP-Seq and RNA-Seq. Extracting biologically meaningful information from this massive dataset has become a critical aspect of many recent studies, particularly in annotating and predicting the functions of unknown genes. The core idea behind genome annotation is to identify genes and various functional elements within the genome sequence and infer their biological functions. Traditional wet-lab experimental methods still rely on extensive efforts for functional verification. However, early bioinformatics algorithms and software primarily employed shallow learning techniques; thus, the ability to characterize data and features learning was limited. With the widespread adoption of RNA-Seq technology, scientists from the biological community began to harness the potential of machine learning and deep learning approaches for gene structure prediction and functional annotation. In this context, we reviewed both conventional methods and contemporary deep learning frameworks, and highlighted novel perspectives on the challenges arising during annotation underscoring the dynamic nature of this evolving scientific landscape.

Understanding YTHDF2-mediated mRNA degradation by m6A-BERT-Deg

N6-methyladenosine (m6A) is the most abundant mRNA modification within mammalian cells, holding pivotal significance in the regulation of mRNA stability, translation and splicing. Furthermore, it plays a critical role in the regulation of RNA degradation by primarily recruiting the YTHDF2 reader protein. However, the selective regulation of mRNA decay of the m6A-methylated mRNA through YTHDF2 binding is poorly understood. To improve our understanding, we developed m6A-BERT-Deg, a BERT model adapted for predicting YTHDF2-mediated degradation of m6A-methylated mRNAs. We meticulously assembled a high-quality training dataset by integrating multiple data sources for the HeLa cell line. To overcome the limitation of small training samples, we employed a pre-training-fine-tuning strategy by first performing a self-supervised pre-training of the model on 427 760 unlabeled m6A site sequences. The test results demonstrated the importance of this pre-training strategy in enabling m6A-BERT-Deg to outperform other benchmark models. We further conducted a comprehensive model interpretation and revealed a surprising finding that the presence of co-factors in proximity to m6A sites may disrupt YTHDF2-mediated mRNA degradation, subsequently enhancing mRNA stability. We also extended our analyses to the HEK293 cell line, shedding light on the context-dependent YTHDF2-mediated mRNA degradation.

DSNetax: a deep learning species annotation method based on a deep-shallow parallel framework

Microbial community analysis is an important field to study the composition and function of microbial communities. Microbial species annotation is crucial to revealing microorganisms' complex ecological functions in environmental, ecological and host interactions. Currently, widely used methods can suffer from issues such as inaccurate species-level annotations and time and memory constraints, and as sequencing technology advances and sequencing costs decline, microbial species annotation methods with higher quality classification effectiveness become critical. Therefore, we processed 16S rRNA gene sequences into k-mers sets and then used a trained DNABERT model to generate word vectors. We also design a parallel network structure consisting of deep and shallow modules to extract the semantic and detailed features of 16S rRNA gene sequences. Our method can accurately and rapidly classify bacterial sequences at the SILVA database's genus and species level. The database is characterized by long sequence length (1500 base pairs), multiple sequences (428,748 reads) and high similarity. The results show that our method has better performance. The technique is nearly 20% more accurate at the species level than the currently popular naive Bayes-dominated QIIME 2 annotation method, and the top-5 results at the species level differ from BLAST methods by <2%. In summary, our approach combines a multi-module deep learning approach that overcomes the limitations of existing methods, providing an efficient and accurate solution for microbial species labeling and more reliable data support for microbiology research and application.

Self-supervised learning on millions of primary RNA sequences from 72 vertebrates improves sequence-based RNA splicing prediction

Language models pretrained by self-supervised learning (SSL) have been widely utilized to study protein sequences, while few models were developed for genomic sequences and were limited to single species. Due to the lack of genomes from different species, these models cannot effectively leverage evolutionary information. In this study, we have developed SpliceBERT, a language model pretrained on primary ribonucleic acids (RNA) sequences from 72 vertebrates by masked language modeling, and applied it to sequence-based modeling of RNA splicing. Pretraining SpliceBERT on diverse species enables effective identification of evolutionarily conserved elements. Meanwhile, the learned hidden states and attention weights can characterize the biological properties of splice sites. As a result, SpliceBERT was shown effective on several downstream tasks: zero-shot prediction of variant effects on splicing, prediction of branchpoints in humans, and cross-species prediction of splice sites. Our study highlighted the importance of pretraining genomic language models on a diverse range of species and suggested that SSL is a promising approach to enhance our understanding of the regulatory logic underlying genomic sequences.

DRANetSplicer: A Splice Site Prediction Model Based on Deep Residual Attention Networks

The precise identification of splice sites is essential for unraveling the structure and function of genes, constituting a pivotal step in the gene annotation process. In this study, we developed a novel deep learning model, DRANetSplicer, that integrates residual learning and attention mechanisms for enhanced accuracy in capturing the intricate features of splice sites. We constructed multiple datasets using the most recent versions of genomic data from three different organisms, Oryza sativa japonica, Arabidopsis thaliana and Homo sapiens. This approach allows us to train models with a richer set of high-quality data. DRANetSplicer outperformed benchmark methods on donor and acceptor splice site datasets, achieving an average accuracy of (96.57%, 95.82%) across the three organisms. Comparative analyses with benchmark methods, including SpliceFinder, Splice2Deep, Deep Splicer, EnsembleSplice, and DNABERT, revealed DRANetSplicer's superior predictive performance, resulting in at least a (4.2%, 11.6%) relative reduction in average error rate. We utilized the DRANetSplicer model trained on O. sativa japonica data to predict splice sites in A. thaliana, achieving accuracies for donor and acceptor sites of (94.89%, 94.25%). These results indicate that DRANetSplicer possesses excellent cross-organism predictive capabilities, with its performance in cross-organism predictions even surpassing that of benchmark methods in non-cross-organism predictions. Cross-organism validation showcased DRANetSplicer's excellence in predicting splice sites across similar organisms, supporting its applicability in gene annotation for understudied organisms. We employed multiple methods to visualize the decision-making process of the model. The visualization results indicate that DRANetSplicer can learn and interpret well-known biological features, further validating its overall performance. Our study systematically examined and confirmed the predictive ability of DRANetSplicer from various levels and perspectives, indicating that its practical application in gene annotation is justified.

Classification of Neisseria meningitidis genomes with a bag-of-words approach and machine learning

Whole genome sequencing of bacteria is important to enable strain classification. Using entire genomes as an input to machine learning (ML) models would allow rapid classification of strains while using information from multiple genetic elements. We developed a "bag-of-words" approach to encode, using SentencePiece or k-mer tokenization, entire bacterial genomes and analyze these with ML. Initial model selection identified SentencePiece with 8,000 and 32,000 words as the best approach for genome tokenization. We then classified in Neisseria meningitidis genomes the capsule B group genotype with 99.6% accuracy and the multifactor invasive phenotype with 90.2% accuracy, in an independent test set. Subsequently, in silico knockouts of 2,808 genes confirmed that the ML model predictions aligned with our current understanding of the underlying biology. To our knowledge, this is the first ML method using entire bacterial genomes to classify strains and identify genes considered relevant by the classifier.

GUANinE v1.0: Benchmark Datasets for Genomic AI Sequence-to-Function Models

Computational genomics increasingly relies on machine learning methods for genome interpretation, and the recent adoption of neural sequence-to-function models highlights the need for rigorous model specification and controlled evaluation, problems familiar to other fields of AI. Research strategies that have greatly benefited other fields - including benchmarking, auditing, and algorithmic fairness - are also needed to advance the field of genomic AI and to facilitate model development. Here we propose a genomic AI benchmark, GUANinE, for evaluating model generalization across a number of distinct genomic tasks. Compared to existing task formulations in computational genomics, GUANinE is large-scale, de-noised, and suitable for evaluating pretrained models. GUANinE v1.0 primarily focuses on functional genomics tasks such as functional element annotation and gene expression prediction, and it also draws upon connections to evolutionary biology through sequence conservation tasks. The current GUANinE tasks provide insight into the performance of existing genomic AI models and non-neural baselines, with opportunities to be refined, revisited, and broadened as the field matures. Finally, the GUANinE benchmark allows us to evaluate new self-supervised T5 models and explore the tradeoffs between tokenization and model performance, while showcasing the potential for self-supervision to complement existing pretraining procedures.

Computational design of mRNA vaccines

mRNA technology has emerged as a successful vaccine platform that offered a swift response to the COVID-19 pandemic. Accumulating evidence shows that vaccine efficacy, thermostability, and other important properties, are largely impacted by intrinsic properties of the mRNA molecule, such as RNA sequence and structure, both of which can be optimized. Designing mRNA sequence for vaccines presents a combinatorial problem due to an extremely large selection space. For instance, due to the degeneracy of the genetic code, there are over 10632 possible mRNA sequences that could encode the spike protein, the COVID-19 vaccines' target. Moreover, designing different elements of the mRNA sequence simultaneously against multiple objectives such as translational efficiency, reduced reactogenicity, and improved stability requires an efficient and sophisticated optimization strategy. Recently, there has been a growing interest in utilizing computational tools to redesign mRNA sequences to improve vaccine characteristics and expedite discovery timelines. In this review, we explore important biophysical features of mRNA to be considered for vaccine design and discuss how computational approaches can be applied to rapidly design mRNA sequences with desirable characteristics.

Machine Learning and Deep Learning in Synthetic Biology: Key Architectures, Applications, and Challenges

Machine learning (ML), particularly deep learning (DL), has made rapid and substantial progress in synthetic biology in recent years. Biotechnological applications of biosystems, including pathways, enzymes, and whole cells, are being probed frequently with time. The intricacy and interconnectedness of biosystems make it challenging to design them with the desired properties. ML and DL have a synergy with synthetic biology. Synthetic biology can be employed to produce large data sets for training models (for instance, by utilizing DNA synthesis), and ML/DL models can be employed to inform design (for example, by generating new parts or advising unrivaled experiments to perform). This potential has recently been brought to light by research at the intersection of engineering biology and ML/DL through achievements like the design of novel biological components, best experimental design, automated analysis of microscopy data, protein structure prediction, and biomolecular implementations of ANNs (Artificial Neural Networks). I have divided this review into three sections. In the first section, I describe predictive potential and basics of ML along with myriad applications in synthetic biology, especially in engineering cells, activity of proteins, and metabolic pathways. In the second section, I describe fundamental DL architectures and their applications in synthetic biology. Finally, I describe different challenges causing hurdles in the progress of ML/DL and synthetic biology along with their solutions.

Self-supervised Learning for DNA sequences with circular dilated convolutional networks

DNA molecules commonly exhibit wide interactions between the nucleobases. Modeling the interactions is important for obtaining accurate sequence-based inference. Although many deep learning methods have recently been developed for modeling DNA sequences, they still suffer from two major issues: 1) most existing methods can handle only short DNA fragments and fail to capture long-range information; 2) current methods always require massive supervised labels, which are hard to obtain in practice. We propose a new method to address both issues. Our neural network employs circular dilated convolutions as building blocks in the backbone. As a result, our network can take long DNA sequences as input without any condensation. We also incorporate the neural network into a self-supervised learning framework to capture inherent information in DNA without expensive supervised labeling. We have tested our model in two DNA inference tasks, the human variant effect and the open chromatin region of plants, where the experimental results show that our method outperforms five other deep learning models. Our code is available at https://github.com/wiedersehne/cdilDNA.

A comprehensive evaluation of large Language models on benchmark biomedical text processing tasks

Recently, Large Language Models (LLMs) have demonstrated impressive capability to solve a wide range of tasks. However, despite their success across various tasks, no prior work has investigated their capability in the biomedical domain yet. To this end, this paper aims to evaluate the performance of LLMs on benchmark biomedical tasks. For this purpose, a comprehensive evaluation of 4 popular LLMs in 6 diverse biomedical tasks across 26 datasets has been conducted. To the best of our knowledge, this is the first work that conducts an extensive evaluation and comparison of various LLMs in the biomedical domain. Interestingly, we find based on our evaluation that in biomedical datasets that have smaller training sets, zero-shot LLMs even outperform the current state-of-the-art models when they were fine-tuned only on the training set of these datasets. This suggests that pre-training on large text corpora makes LLMs quite specialized even in the biomedical domain. We also find that not a single LLM can outperform other LLMs in all tasks, with the performance of different LLMs may vary depending on the task. While their performance is still quite poor in comparison to the biomedical models that were fine-tuned on large training sets, our findings demonstrate that LLMs have the potential to be a valuable tool for various biomedical tasks that lack large annotated data.

Interpretable Multi-Scale Deep Learning for RNA Methylation Analysis across Multiple Species

RNA modification plays a crucial role in cellular regulation. However, traditional high-throughput sequencing methods for elucidating their functional mechanisms are time-consuming and labor-intensive, despite extensive research. Moreover, existing methods often limit their focus to specific species, neglecting the simultaneous exploration of RNA modifications across diverse species. Therefore, a versatile computational approach is necessary for interpretable analysis of RNA modifications across species. A multi-scale biological language-based deep learning model is proposed for interpretable, sequential-level prediction of diverse RNA modifications. Benchmark comparisons across species demonstrate the model's superiority in predicting various RNA methylation types over current state-of-the-art methods. The cross-species validation and attention weight visualization also highlight the model's capability to capture sequential and functional semantics from genomic backgrounds. Our analysis of RNA modifications helps us find the potential existence of "biological grammars" in each modification type, which could be effective for mapping methylation-related sequential patterns and understanding the underlying biological mechanisms of RNA modifications.

miTDS: Uncovering miRNA-mRNA interactions with deep learning for functional target prediction

MicroRNAs (miRNAs) are vital in regulating gene expression through binding to specific target sites on messenger RNAs (mRNAs), a process closely tied to cancer pathogenesis. Identifying miRNA functional targets is essential but challenging, due to incomplete genome annotation and an emphasis on known miRNA-mRNA interactions, restricting predictions of unknown ones. To address those challenges, we have developed a deep learning model based on miRNA functional target identification, named miTDS, to investigate miRNA-mRNA interactions. miTDS first employs a scoring mechanism to eliminate unstable sequence pairs and then utilizes a dynamic word embedding model based on the transformer architecture, enabling a comprehensive analysis of miRNA-mRNA interaction sites by harnessing the global contextual associations of each nucleotide. On this basis, miTDS fuses extended seed alignment representations learned in the multi-scale attention mechanism module with dynamic semantic representations extracted in the RNA-based dual-path module, which can further elucidate and predict miRNA and mRNA functions and interactions. To validate the effectiveness of miTDS, we conducted a thorough comparison with state-of-the-art miRNA-mRNA functional target prediction methods. The evaluation, performed on a dataset cross-referenced with entries from MirTarbase and Diana-TarBase, revealed that miTDS surpasses current methods in accurately predicting functional targets. In addition, our model exhibited proficiency in identifying A-to-I RNA editing sites, which represents an aberrant interaction that yields valuable insights into the suppression of cancerous processes.

scAuto as a comprehensive framework for single-cell chromatin accessibility data analysis

Interpreting single-cell chromatin accessibility data is crucial for understanding intercellular heterogeneity regulation. Despite the progress in computational methods for analyzing this data, there is still a lack of a comprehensive analytical framework and a user-friendly online analysis tool. To fill this gap, we developed a pre-trained deep learning-based framework, single-cell auto-correlation transformers (scAuto), to overcome the challenge. Following DNABERT's methodology of pre-training and fine-tuning, scAuto learns a general understanding of DNA sequence's grammar by being pre-trained on unlabeled human genome via self-supervision; it is then transferred to the single-cell chromatin accessibility analysis task of scATAC-seq data for supervised fine-tuning. We extensively validated scAuto on the Buenrostro2018 dataset, demonstrating its superior performance on chromatin accessibility prediction, single-cell clustering, and data denoising. Based on scAuto, we further developed an interactive web server for single-cell chromatin accessibility data analysis. It integrates tutorial-style interfaces for those with limited programming skills. The platform is accessible at http://zhanglab.icaup.cn. To our knowledge, this work is expected to help analyze single-cell chromatin accessibility data and facilitate the development of precision medicine.

PolyAMiner-Bulk is a deep learning-based algorithm that decodes alternative polyadenylation dynamics from bulk RNA-seq data

Alternative polyadenylation (APA) is a key post-transcriptional regulatory mechanism; yet, its regulation and impact on human diseases remain understudied. Existing bulk RNA sequencing (RNA-seq)-based APA methods predominantly rely on predefined annotations, severely impacting their ability to decode novel tissue- and disease-specific APA changes. Furthermore, they only account for the most proximal and distal cleavage and polyadenylation sites (C/PASs). Deconvoluting overlapping C/PASs and the inherent noisy 3' UTR coverage in bulk RNA-seq data pose additional challenges. To overcome these limitations, we introduce PolyAMiner-Bulk, an attention-based deep learning algorithm that accurately recapitulates C/PAS sequence grammar, resolves overlapping C/PASs, captures non-proximal-to-distal APA changes, and generates visualizations to illustrate APA dynamics. Evaluation on multiple datasets strongly evinces the performance merit of PolyAMiner-Bulk, accurately identifying more APA changes compared with other methods. With the growing importance of APA and the abundance of bulk RNA-seq data, PolyAMiner-Bulk establishes a robust paradigm of APA analysis.

Evaluation and optimization of sequence-based gene regulatory deep learning models

Neural networks have emerged as immensely powerful tools in predicting functional genomic regions, notably evidenced by recent successes in deciphering gene regulatory logic. However, a systematic evaluation of how model architectures and training strategies impact genomics model performance is lacking. To address this gap, we held a DREAM Challenge where competitors trained models on a dataset of millions of random promoter DNA sequences and corresponding expression levels, experimentally determined in yeast, to best capture the relationship between regulatory DNA and gene expression. For a robust evaluation of the models, we designed a comprehensive suite of benchmarks encompassing various sequence types. While some benchmarks produced similar results across the top-performing models, others differed substantially. All top-performing models used neural networks, but diverged in architectures and novel training strategies, tailored to genomics sequence data. To dissect how architectural and training choices impact performance, we developed the Prix Fixe framework to divide any given model into logically equivalent building blocks. We tested all possible combinations for the top three models and observed performance improvements for each. The DREAM Challenge models not only achieved state-of-the-art results on our comprehensive yeast dataset but also consistently surpassed existing benchmarks on Drosophila and human genomic datasets. Overall, we demonstrate that high-quality gold-standard genomics datasets can drive significant progress in model development.

Advancing Transcription Factor Binding Site Prediction Using DNA Breathing Dynamics and Sequence Transformers via Cross Attention

Understanding the impact of genomic variants on transcription factor binding and gene regulation remains a key area of research, with implications for unraveling the complex mechanisms underlying various functional effects. Our study delves into the role of DNA's biophysical properties, including thermodynamic stability, shape, and flexibility in transcription factor (TF) binding. We developed a multi-modal deep learning model integrating these properties with DNA sequence data. Trained on ChIP-Seq (chromatin immunoprecipitation sequencing) data in vivo involving 690 TF-DNA binding events in human genome, our model significantly improves prediction performance in over 660 binding events, with up to 9.6% increase in AUROC metric compared to the baseline model when using no DNA biophysical properties explicitly. Further, we expanded our analysis to in vitro high-throughput Systematic Evolution of Ligands by Exponential enrichment (SELEX) and Protein Binding Microarray (PBM) datasets, comparing our model with established frameworks. The inclusion of DNA breathing features consistently improved TF binding predictions across different cell lines in these datasets. Notably, for complex ChIP-Seq datasets, integrating DNABERT2 with a cross-attention mechanism provided greater predictive capabilities and insights into the mechanisms of disease-related non-coding variants found in genome-wide association studies. This work highlights the importance of DNA biophysical characteristics in TF binding and the effectiveness of multi-modal deep learning models in gene regulation studies.

NanoCon: contrastive learning-based deep hybrid network for nanopore methylation detection

MOTIVATION

5-Methylcytosine (5mC), a fundamental element of DNA methylation in eukaryotes, plays a vital role in gene expression regulation, embryonic development, and other biological processes. Although several computational methods have been proposed for detecting the base modifications in DNA like 5mC sites from Nanopore sequencing data, they face challenges including sensitivity to noise, and ignoring the imbalanced distribution of methylation sites in real-world scenarios.

RESULTS

Here, we develop NanoCon, a deep hybrid network coupled with contrastive learning strategy to detect 5mC methylation sites from Nanopore reads. In particular, we adopted a contrastive learning module to alleviate the issues caused by imbalanced data distribution in nanopore sequencing, offering a more accurate and robust detection of 5mC sites. Evaluation results demonstrate that NanoCon outperforms existing methods, highlighting its potential as a valuable tool in genomic sequencing and methylation prediction. In addition, we also verified the effectiveness of our representation learning ability on two datasets by visualizing the dimension reduction of the features of methylation and nonmethylation sites from our NanoCon. Furthermore, cross-species and cross-5mC methylation motifs experiments indicated the robustness and the ability to perform transfer learning of our model. We hope this work can contribute to the community by providing a powerful and reliable solution for 5mC site detection in genomic studies.

AVAILABILITY AND IMPLEMENTATION

The project code is available at https://github.com/Challis-yin/NanoCon.

HiTaxon: a hierarchical ensemble framework for taxonomic classification of short reads

Motivation

Whole microbiome DNA and RNA sequencing (metagenomics and metatranscriptomics) are pivotal to determining the functional roles of microbial communities. A key challenge in analyzing these complex datasets, typically composed of tens of millions of short reads, is accurately classifying reads to their taxa of origin. While still performing worse relative to reference-based short-read tools in species classification, ML algorithms have shown promising results in taxonomic classification at higher ranks. A recent approach exploited to enhance the performance of ML tools, which can be translated to reference-dependent classifiers, has been to integrate the hierarchical structure of taxonomy within the tool's predictive algorithm.

Results

Here, we introduce HiTaxon, an end-to-end hierarchical ensemble framework for taxonomic classification. HiTaxon facilitates data collection and processing, reference database construction and optional training of ML models to streamline ensemble creation. We show that databases created by HiTaxon improve the species-level performance of reference-dependent classifiers, while reducing their computational overhead. In addition, through exploring hierarchical methods for HiTaxon, we highlight that our custom approach to hierarchical ensembling improves species-level classification relative to traditional strategies. Finally, we demonstrate the improved performance of our hierarchical ensembles over current state-of-the-art classifiers in species classification using datasets comprised of either simulated or experimentally derived reads.

Availability and implementation

HiTaxon is available at: https://github.com/ParkinsonLab/HiTaxon.

Predicting circRNA-miRNA interactions utilizing transformer-based RNA sequential learning and high-order proximity preserved embedding

A key regulatory mechanism involves circular RNA (circRNA) acting as a sponge to modulate microRNA (miRNA), and thus, studying their interaction has significant medical implications. In this field, there are currently two pressing issues that remain unresolved. Firstly, due to the scarcity of verified interactions, we require a minimal amount of samples for training. Secondly, the current models lack interpretability. Therefore, we propose SPBCMI, a method that combines sequence features extracted using the Bidirectional Encoder Representations from Transformer (BERT) model and structural features of biological molecule networks extracted through graph embedding to train a GBDT (Gradient-boosted decision trees) classifier for prediction. Our method yielded an AUC of 0.9143, which is currently the best for this problem. Furthermore, in the case study, SPBCMI accurately predicted 7 out of 10 circRNA-miRNA interactions. These results show that our method provides an innovative and high-performing approach to understanding the interaction between circRNA and miRNA.

ProkBERT family: genomic language models for microbiome applications

Background

In the evolving landscape of microbiology and microbiome analysis, the integration of machine learning is crucial for understanding complex microbial interactions, and predicting and recognizing novel functionalities within extensive datasets. However, the effectiveness of these methods in microbiology faces challenges due to the complex and heterogeneous nature of microbial data, further complicated by low signal-to-noise ratios, context-dependency, and a significant shortage of appropriately labeled datasets. This study introduces the ProkBERT model family, a collection of large language models, designed for genomic tasks. It provides a generalizable sequence representation for nucleotide sequences, learned from unlabeled genome data. This approach helps overcome the above-mentioned limitations in the field, thereby improving our understanding of microbial ecosystems and their impact on health and disease.

Methods

ProkBERT models are based on transfer learning and self-supervised methodologies, enabling them to use the abundant yet complex microbial data effectively. The introduction of the novel Local Context-Aware (LCA) tokenization technique marks a significant advancement, allowing ProkBERT to overcome the contextual limitations of traditional transformer models. This methodology not only retains rich local context but also demonstrates remarkable adaptability across various bioinformatics tasks.

Results

In practical applications such as promoter prediction and phage identification, the ProkBERT models show superior performance. For promoter prediction tasks, the top-performing model achieved a Matthews Correlation Coefficient (MCC) of 0.74 for E. coli and 0.62 in mixed-species contexts. In phage identification, ProkBERT models consistently outperformed established tools like VirSorter2 and DeepVirFinder, achieving an MCC of 0.85. These results underscore the models' exceptional accuracy and generalizability in both supervised and unsupervised tasks.

Conclusions

The ProkBERT model family is a compact yet powerful tool in the field of microbiology and bioinformatics. Its capacity for rapid, accurate analyses and its adaptability across a spectrum of tasks marks a significant advancement in machine learning applications in microbiology. The models are available on GitHub (https://github.com/nbrg-ppcu/prokbert) and HuggingFace (https://huggingface.co/nerualbioinfo) providing an accessible tool for the community.

Multiple sequence alignment-based RNA language model and its application to structural inference

Compared with proteins, DNA and RNA are more difficult languages to interpret because four-letter coded DNA/RNA sequences have less information content than 20-letter coded protein sequences. While BERT (Bidirectional Encoder Representations from Transformers)-like language models have been developed for RNA, they are ineffective at capturing the evolutionary information from homologous sequences because unlike proteins, RNA sequences are less conserved. Here, we have developed an unsupervised multiple sequence alignment-based RNA language model (RNA-MSM) by utilizing homologous sequences from an automatic pipeline, RNAcmap, as it can provide significantly more homologous sequences than manually annotated Rfam. We demonstrate that the resulting unsupervised, two-dimensional attention maps and one-dimensional embeddings from RNA-MSM contain structural information. In fact, they can be directly mapped with high accuracy to 2D base pairing probabilities and 1D solvent accessibilities, respectively. Further fine-tuning led to significantly improved performance on these two downstream tasks compared with existing state-of-the-art techniques including SPOT-RNA2 and RNAsnap2. By comparison, RNA-FM, a BERT-based RNA language model, performs worse than one-hot encoding with its embedding in base pair and solvent-accessible surface area prediction. We anticipate that the pre-trained RNA-MSM model can be fine-tuned on many other tasks related to RNA structure and function.

Multiomics-integrated deep language model enables in silico genome-wide detection of transcription factor binding site in unexplored biosamples

MOTIVATION

Transcription factor binding sites (TFBS) are regulatory elements that have significant impact on transcription regulation and cell fate determination. Canonical motifs, biological experiments, and computational methods have made it possible to discover TFBS. However, most existing in silico TFBS prediction models are solely DNA-based, and are trained and utilized within the same biosample, which fail to infer TFBS in experimentally unexplored biosamples.

RESULTS

Here, we propose TFBS prediction by modified TransFormer (TFTF), a multimodal deep language architecture which integrates multiomics information in epigenetic studies. In comparison to existing computational techniques, TFTF has state-of-the-art accuracy, and is also the first approach to accurately perform genome-wide detection for cell-type and species-specific TFBS in experimentally unexplored biosamples. Compared to peak calling methods, TFTF consistently discovers true TFBS in threshold tuning-free way, with higher recalled rates. The underlying mechanism of TFTF reveals greater attention to the targeted TF's motif region in TFBS, and general attention to the entire peak region in non-TFBS. TFTF can benefit from the integration of broader and more diverse data for improvement and can be applied to multiple epigenetic scenarios.

AVAILABILITY AND IMPLEMENTATION

We provide a web server (https://tftf.ibreed.cn/) for users to utilize TFTF model. Users can train TFTF model and discover TFBS with their own data.

Advances in Protein-Ligand Binding Affinity Prediction via Deep Learning: A Comprehensive Study of Datasets, Data Preprocessing Techniques, and Model Architectures

BACKGROUND

Drug discovery is a complex and expensive procedure involving several timely and costly phases through which new potential pharmaceutical compounds must pass to get approved. One of these critical steps is the identification and optimization of lead compounds, which has been made more accessible by the introduction of computational methods, including deep learning (DL) techniques. Diverse DL model architectures have been put forward to learn the vast landscape of interaction between proteins and ligands and predict their affinity, helping in the identification of lead compounds.

OBJECTIVE

This survey fills a gap in previous research by comprehensively analyzing the most commonly used datasets and discussing their quality and limitations. It also offers a comprehensive classification of the most recent DL methods in the context of protein-ligand binding affinity prediction (BAP), providing a fresh perspective on this evolving field.

METHODS

We thoroughly examine commonly used datasets for BAP and their inherent characteristics. Our exploration extends to various preprocessing steps and DL techniques, including graph neural networks, convolutional neural networks, and transformers, which are found in the literature. We conducted extensive literature research to ensure that the most recent deep learning approaches for BAP were included by the time of writing this manuscript.

RESULTS

The systematic approach used for the present study highlighted inherent challenges to BAP via DL, such as data quality, model interpretability, and explainability, and proposed considerations for future research directions. We present valuable insights to accelerate the development of more effective and reliable DL models for BAP within the research community.

CONCLUSION

The present study can considerably enhance future research on predicting affinity between protein and ligand molecules, hence further improving the overall drug development process.

Prediction of SARS-CoV-2 Infection Phosphorylation Sites and Associations of these Modifications with Lung Cancer Development

INTRODUCTION

Since the emergence of SARS-CoV-2 viruses, multiple mutant strains have been identified. Infection with SARS-CoV-2 virus leads to alterations in host cell phosphorylation signal, which systematically modulates the immune response.

METHODS

Identification and analysis of SARS-CoV-2 virus infection phosphorylation sites enable insight into the mechanisms of viral infection and effects on host cells, providing important fundamental data for the study and development of potent drugs for the treatment of immune inflammatory diseases. In this paper, we have analyzed the SARS-CoV-2 virus-infected phosphorylation region and developed a transformer-based deep learning-assisted identification method for the specific identification of phosphorylation sites in SARS-CoV-2 virus-infected host cells.

RESULTS

Furthermore, through association analysis with lung cancer, we found that SARS-CoV-2 infection may affect the regulatory role of the immune system, leading to an abnormal increase or decrease in the immune inflammatory response, which may be associated with the development and progression of cancer.

CONCLUSION

We anticipate that this study will provide an important reference for SARS-CoV-2 virus evolution as well as immune-related studies and provide a reliable complementary screening tool for anti-SARS-CoV-2 virus drug and vaccine design.

promSEMBLE: Hard Pattern Mining and Ensemble Learning for Detecting DNA Promoter Sequences

Accurate identification of DNA promoter sequences is of crucial importance in unraveling the underlying mechanisms that regulate gene transcription. Initiation of transcription is controlled through regulatory transcription factors binding to promoter core regions in the DNA sequence. Detection of promoter regions is necessary if we are to build genetic regulatory networks for biomedical and clinical applications, and for identification of rarely expressed genes. We propose a novel ensemble learning technique using deep recurrent neural networks with convolutional feature extraction and hard negative pattern mining to detect several types of promoter sequences, including promoter sequences with the TATA-box and without the TATA-box, within DNA sequences of four different species. Using extensive independent tests and previously published results, we demonstrate that our method sets a new state-of-the-art of over 98% Matthews correlation coefficient in all eight organism categories for recognizing the stretch of base pairs that code for the promoter region within DNA sequences.

SetQuence & SetOmic: Deep set transformers for whole genome and exome tumour analysis

In oncology, Deep Learning has shown great potential to personalise tasks such as tumour type classification, based on per-patient omics data-sets. Being high dimensional, incorporation of such data in one model is a challenge, often leading to one-dimensional studies and, therefore, information loss. Instead, we first propose relying on non-fixed sets of whole genome or whole exome variant-associated sequences, which can be used for supervised learning of oncology-relevant tasks by our Set Transformer based Deep Neural Network, SetQuence. We optimise this architecture to improve its efficiency. This allows for exploration of not just coding but also non-coding variants, from large datasets. Second, we extend the model to incorporate these representations together with multiple other sources of omics data in a flexible way with SetOmic. Evaluation, using these representations, shows improved robustness and reduced information loss compared to previous approaches, while still being computationally tractable. By means of Explainable Artificial Intelligence methods, our models are able to recapitulate the biological contribution of highly attributed features in the tumours studied. This validation opens the door to novel directions in multi-faceted genome and exome wide biomarker discovery and personalised treatment among other presently clinically relevant tasks.

MRM-BERT: a novel deep neural network predictor of multiple RNA modifications by fusing BERT representation and sequence features

RNA modifications play crucial roles in various biological processes and diseases. Accurate prediction of RNA modification sites is essential for understanding their functions. In this study, we propose a hybrid approach that fuses a pre-trained sequence representation with various sequence features to predict multiple types of RNA modifications in one combined prediction framework. We developed MRM-BERT, a deep learning method that combined the pre-trained DNABERT deep sequence representation module and the convolutional neural network (CNN) exploiting four traditional sequence feature encodings to improve the prediction performance. MRM-BERT was evaluated on multiple datasets of 12 commonly occurring RNA modifications, including m6A, m5C, m1A and so on. The results demonstrate that our hybrid model outperforms other models in terms of area under receiver operating characteristic curve (AUC) for all 12 types of RNA modifications. MRM-BERT is available as an online tool (http://117.122.208.21:8501) or source code (https://github.com/abhhba999/MRM-BERT), which allows users to predict RNA modification sites and visualize the results. Overall, our study provides an effective and efficient approach to predict multiple RNA modifications, contributing to the understanding of RNA biology and the development of therapeutic strategies.

Machine Learning to Advance Human Genome-Wide Association Studies

Machine learning, including deep learning, reinforcement learning, and generative artificial intelligence are revolutionising every area of our lives when data are made available. With the help of these methods, we can decipher information from larger datasets while addressing the complex nature of biological systems in a more efficient way. Although machine learning methods have been introduced to human genetic epidemiological research as early as 2004, those were never used to their full capacity. In this review, we outline some of the main applications of machine learning to assigning human genetic loci to health outcomes. We summarise widely used methods and discuss their advantages and challenges. We also identify several tools, such as Combi, GenNet, and GMSTool, specifically designed to integrate these methods for hypothesis-free analysis of genetic variation data. We elaborate on the additional value and limitations of these tools from a geneticist's perspective. Finally, we discuss the fast-moving field of foundation models and large multi-modal omics biobank initiatives.

BERT-5mC: an interpretable model for predicting 5-methylcytosine sites of DNA based on BERT

DNA 5-methylcytosine (5mC) is widely present in multicellular eukaryotes, which plays important roles in various developmental and physiological processes and a wide range of human diseases. Thus, it is essential to accurately detect the 5mC sites. Although current sequencing technologies can map genome-wide 5mC sites, these experimental methods are both costly and time-consuming. To achieve a fast and accurate prediction of 5mC sites, we propose a new computational approach, BERT-5mC. First, we pre-trained a domain-specific BERT (bidirectional encoder representations from transformers) model by using human promoter sequences as language corpus. BERT is a deep two-way language representation model based on Transformer. Second, we fine-tuned the domain-specific BERT model based on the 5mC training dataset to build the model. The cross-validation results show that our model achieves an AUROC of 0.966 which is higher than other state-of-the-art methods such as iPromoter-5mC, 5mC_Pred, and BiLSTM-5mC. Furthermore, our model was evaluated on the independent test set, which shows that our model achieves an AUROC of 0.966 that is also higher than other state-of-the-art methods. Moreover, we analyzed the attention weights generated by BERT to identify a number of nucleotide distributions that are closely associated with 5mC modifications. To facilitate the use of our model, we built a webserver which can be freely accessed at: http://5mc-pred.zhulab.org.cn.

StableDNAm: towards a stable and efficient model for predicting DNA methylation based on adaptive feature correction learning

BACKGROUND

DNA methylation, instrumental in numerous life processes, underscores the paramount importance of its accurate prediction. Recent studies suggest that deep learning, due to its capacity to extract profound insights, provides a more precise DNA methylation prediction. However, issues related to the stability and generalization performance of these models persist.

RESULTS

In this study, we introduce an efficient and stable DNA methylation prediction model. This model incorporates a feature fusion approach, adaptive feature correction technology, and a contrastive learning strategy. The proposed model presents several advantages. First, DNA sequences are encoded at four levels to comprehensively capture intricate information across multi-scale and low-span features. Second, we design a sequence-specific feature correction module that adaptively adjusts the weights of sequence features. This improvement enhances the model's stability and scalability, or its generality. Third, our contrastive learning strategy mitigates the instability issues resulting from sparse data. To validate our model, we conducted multiple sets of experiments on commonly used datasets, demonstrating the model's robustness and stability. Simultaneously, we amalgamate various datasets into a single, unified dataset. The experimental outcomes from this combined dataset substantiate the model's robust adaptability.

CONCLUSIONS

Our research findings affirm that the StableDNAm model is a general, stable, and effective instrument for DNA methylation prediction. It holds substantial promise for providing invaluable assistance in future methylation-related research and analyses.

Transformer-based tool recommendation system in Galaxy

BACKGROUND

Galaxy is a web-based open-source platform for scientific analyses. Researchers use thousands of high-quality tools and workflows for their respective analyses in Galaxy. Tool recommender system predicts a collection of tools that can be used to extend an analysis. In this work, a tool recommender system is developed by training a transformer on workflows available on Galaxy Europe and its performance is compared to other neural networks such as recurrent, convolutional and dense neural networks.

RESULTS

The transformer neural network achieves two times faster convergence, has significantly lower model usage (model reconstruction and prediction) time and shows a better generalisation that goes beyond training workflows than the older tool recommender system created using RNN in Galaxy. In addition, the transformer also outperforms CNN and DNN on several key indicators. It achieves a faster convergence time, lower model usage time, and higher quality tool recommendations than CNN. Compared to DNN, it converges faster to a higher precision@k metric (approximately 0.98 by transformer compared to approximately 0.9 by DNN) and shows higher quality tool recommendations.

CONCLUSION

Our work shows a novel usage of transformers to recommend tools for extending scientific workflows. A more robust tool recommendation model, created using a transformer, having significantly lower usage time than RNN and CNN, higher precision@k than DNN, and higher quality tool recommendations than all three neural networks, will benefit researchers in creating scientifically significant workflows and exploratory data analysis in Galaxy. Additionally, the ability to train faster than all three neural networks imparts more scalability for training on larger datasets consisting of millions of tool sequences. Open-source scripts to create the recommendation model are available under MIT licence at https://github.com/anuprulez/galaxy_tool_recommendation_transformers.

AptaTrans: a deep neural network for predicting aptamer-protein interaction using pretrained encoders

BACKGROUND

Aptamers, which are biomaterials comprised of single-stranded DNA/RNA that form tertiary structures, have significant potential as next-generation materials, particularly for drug discovery. The systematic evolution of ligands by exponential enrichment (SELEX) method is a critical in vitro technique employed to identify aptamers that bind specifically to target proteins. While advanced SELEX-based methods such as Cell- and HT-SELEX are available, they often encounter issues such as extended time consumption and suboptimal accuracy. Several In silico aptamer discovery methods have been proposed to address these challenges. These methods are specifically designed to predict aptamer-protein interaction (API) using benchmark datasets. However, these methods often fail to consider the physicochemical interactions between aptamers and proteins within tertiary structures.

RESULTS

In this study, we propose AptaTrans, a pipeline for predicting API using deep learning techniques. AptaTrans uses transformer-based encoders to handle aptamer and protein sequences at the monomer level. Furthermore, pretrained encoders are utilized for the structural representation. After validation with a benchmark dataset, AptaTrans has been integrated into a comprehensive toolset. This pipeline synergistically combines with Apta-MCTS, a generative algorithm for recommending aptamer candidates.

CONCLUSION

The results show that AptaTrans outperforms existing models for predicting API, and the efficacy of the AptaTrans pipeline has been confirmed through various experimental tools. We expect AptaTrans will enhance the cost-effectiveness and efficiency of SELEX in drug discovery. The source code and benchmark dataset for AptaTrans are available at https://github.com/pnumlb/AptaTrans .

Integrative approaches based on genomic techniques in the functional studies on enhancers

With the development of sequencing technology and the dramatic drop in sequencing cost, the functions of noncoding genes are being characterized in a wide variety of fields (e.g. biomedicine). Enhancers are noncoding DNA elements with vital transcription regulation functions. Tens of thousands of enhancers have been identified in the human genome; however, the location, function, target genes and regulatory mechanisms of most enhancers have not been elucidated thus far. As high-throughput sequencing techniques have leapt forwards, omics approaches have been extensively employed in enhancer research. Multidimensional genomic data integration enables the full exploration of the data and provides novel perspectives for screening, identification and characterization of the function and regulatory mechanisms of unknown enhancers. However, multidimensional genomic data are still difficult to integrate genome wide due to complex varieties, massive amounts, high rarity, etc. To facilitate the appropriate methods for studying enhancers with high efficacy, we delineate the principles, data processing modes and progress of various omics approaches to study enhancers and summarize the applications of traditional machine learning and deep learning in multi-omics integration in the enhancer field. In addition, the challenges encountered during the integration of multiple omics data are addressed. Overall, this review provides a comprehensive foundation for enhancer analysis.

A transformer-based genomic prediction method fused with knowledge-guided module

Genomic prediction (GP) uses single nucleotide polymorphisms (SNPs) to establish associations between markers and phenotypes. Selection of early individuals by genomic estimated breeding value shortens the generation interval and speeds up the breeding process. Recently, methods based on deep learning (DL) have gained great attention in the field of GP. In this study, we explore the application of Transformer-based structures to GP and develop a novel deep-learning model named GPformer. GPformer obtains a global view by gleaning beneficial information from all relevant SNPs regardless of the physical distance between SNPs. Comprehensive experimental results on five different crop datasets show that GPformer outperforms ridge regression-based linear unbiased prediction (RR-BLUP), support vector regression (SVR), light gradient boosting machine (LightGBM) and deep neural network genomic prediction (DNNGP) in terms of mean absolute error, Pearson's correlation coefficient and the proposed metric consistent index. Furthermore, we introduce a knowledge-guided module (KGM) to extract genome-wide association studies-based information, which is fused into GPformer as prior knowledge. KGM is very flexible and can be plugged into any DL network. Ablation studies of KGM on three datasets illustrate the efficiency of KGM adequately. Moreover, GPformer is robust and stable to hyperparameters and can generalize to each phenotype of every dataset, which is suitable for practical application scenarios.

DNABERT-based explainable lncRNA identification in plant genome assemblies

Long non-coding ribonucleic acids (lncRNAs) have been shown to play an important role in plant gene regulation, involving both epigenetic and transcript regulation. LncRNAs are transcripts longer than 200 nucleotides that are not translated into functional proteins but can be translated into small peptides. Machine learning models have predominantly used transcriptome data with manually defined features to detect lncRNAs, however, they often underrepresent the abundance of lncRNAs and can be biased in their detection. Here we present a study using Natural Language Processing (NLP) models to identify plant lncRNAs from genomic sequences rather than transcriptomic data. The NLP models were trained to predict lncRNAs for seven model and crop species (Zea mays, Arabidopsis thaliana, Brassica napus, Brassica oleracea, Brassica rapa, Glycine max and Oryza sativa) using publicly available genomic references. We demonstrated that lncRNAs can be accurately predicted from genomic sequences with the highest accuracy of 83.4% for Z. mays and the lowest accuracy of 57.9% for B. rapa, revealing that genome assembly quality might affect the accuracy of lncRNA identification. Furthermore, we demonstrated the potential of using NLP models for cross-species prediction with an average of 63.1% accuracy using target species not previously seen by the model. As more species are incorporated into the training datasets, we expect the accuracy to increase, becoming a more reliable tool for uncovering novel lncRNAs. Finally, we show that the models can be interpreted using explainable artificial intelligence to identify motifs important to lncRNA prediction and that these motifs frequently flanked the lncRNA sequence.

Early detection of hepatocellular carcinoma via no end-repair enzymatic methylation sequencing of cell-free DNA and pre-trained neural network

BACKGROUND

Early detection of hepatocellular carcinoma (HCC) is important in order to improve patient prognosis and survival rate. Methylation sequencing combined with neural networks to identify cell-free DNA (cfDNA) carrying aberrant methylation offers an appealing and non-invasive approach for HCC detection. However, some limitations exist in traditional methylation detection technologies and models, which may impede their performance in the read-level detection of HCC.

METHODS

We developed a low DNA damage and high-fidelity methylation detection method called No End-repair Enzymatic Methyl-seq (NEEM-seq). We further developed a read-level neural detection model called DeepTrace that can better identify HCC-derived sequencing reads through a pre-trained and fine-tuned neural network. After pre-training on 11 million reads from NEEM-seq, DeepTrace was fine-tuned using 1.2 million HCC-derived reads from tumor tissue DNA after noise reduction, and 2.7 million non-tumor reads from non-tumor cfDNA. We validated the model using data from 130 individuals with cfDNA whole-genome NEEM-seq at around 1.6X depth.

RESULTS

NEEM-seq overcomes the drawbacks of traditional enzymatic methylation sequencing methods by avoiding the introduction of unmethylation errors in cfDNA. DeepTrace outperformed other models in identifying HCC-derived reads and detecting HCC individuals. Based on the whole-genome NEEM-seq data of cfDNA, our model showed high accuracy of 96.2%, sensitivity of 93.6%, and specificity of 98.5% in the validation cohort consisting of 62 HCC patients, 48 liver disease patients, and 20 healthy individuals. In the early stage of HCC (BCLC 0/A and TNM I), the sensitivity of DeepTrace was 89.6 and 89.5% respectively, outperforming Alpha Fetoprotein (AFP) which showed much lower sensitivity in both BCLC 0/A (50.5%) and TNM I (44.7%).

CONCLUSIONS

By combining high-fidelity methylation data from NEEM-seq with the DeepTrace model, our method has great potential for HCC early detection with high sensitivity and specificity, making it potentially suitable for clinical applications. DeepTrace: https://github.com/Bamrock/DeepTrace.

Deep Learning for Genomics: From Early Neural Nets to Modern Large Language Models

The data explosion driven by advancements in genomic research, such as high-throughput sequencing techniques, is constantly challenging conventional methods used in genomics. In parallel with the urgent demand for robust algorithms, deep learning has succeeded in various fields such as vision, speech, and text processing. Yet genomics entails unique challenges to deep learning, since we expect a superhuman intelligence that explores beyond our knowledge to interpret the genome from deep learning. A powerful deep learning model should rely on the insightful utilization of task-specific knowledge. In this paper, we briefly discuss the strengths of different deep learning models from a genomic perspective so as to fit each particular task with proper deep learning-based architecture, and we remark on practical considerations of developing deep learning architectures for genomics. We also provide a concise review of deep learning applications in various aspects of genomic research and point out current challenges and potential research directions for future genomics applications. We believe the collaborative use of ever-growing diverse data and the fast iteration of deep learning models will continue to contribute to the future of genomics.

Predictive analyses of regulatory sequences with EUGENe

Deep learning has become a popular tool to study cis-regulatory function. Yet efforts to design software for deep-learning analyses in regulatory genomics that are findable, accessible, interoperable and reusable (FAIR) have fallen short of fully meeting these criteria. Here we present elucidating the utility of genomic elements with neural nets (EUGENe), a FAIR toolkit for the analysis of genomic sequences with deep learning. EUGENe consists of a set of modules and subpackages for executing the key functionality of a genomics deep learning workflow: (1) extracting, transforming and loading sequence data from many common file formats; (2) instantiating, initializing and training diverse model architectures; and (3) evaluating and interpreting model behavior. We designed EUGENe as a simple, flexible and extensible interface for streamlining and customizing end-to-end deep-learning sequence analyses, and illustrate these principles through application of the toolkit to three predictive modeling tasks. We hope that EUGENe represents a springboard towards a collaborative ecosystem for deep-learning applications in genomics research.

DNA language models are powerful predictors of genome-wide variant effects

The expanding catalog of genome-wide association studies (GWAS) provides biological insights across a variety of species, but identifying the causal variants behind these associations remains a significant challenge. Experimental validation is both labor-intensive and costly, highlighting the need for accurate, scalable computational methods to predict the effects of genetic variants across the entire genome. Inspired by recent progress in natural language processing, unsupervised pretraining on large protein sequence databases has proven successful in extracting complex information related to proteins. These models showcase their ability to learn variant effects in coding regions using an unsupervised approach. Expanding on this idea, we here introduce the Genomic Pre-trained Network (GPN), a model designed to learn genome-wide variant effects through unsupervised pretraining on genomic DNA sequences. Our model also successfully learns gene structure and DNA motifs without any supervision. To demonstrate its utility, we train GPN on unaligned reference genomes of Arabidopsis thaliana and seven related species within the Brassicales order and evaluate its ability to predict the functional impact of genetic variants in A. thaliana by utilizing allele frequencies from the 1001 Genomes Project and a comprehensive database of GWAS. Notably, GPN outperforms predictors based on popular conservation scores such as phyloP and phastCons. Our predictions for A. thaliana can be visualized as sequence logos in the UCSC Genome Browser (https://genome.ucsc.edu/s/gbenegas/gpn-arabidopsis). We provide code (https://github.com/songlab-cal/gpn) to train GPN for any given species using its DNA sequence alone, enabling unsupervised prediction of variant effects across the entire genome.

Dynamic characterization and interpretation for protein-RNA interactions across diverse cellular conditions using HDRNet

RNA-binding proteins play crucial roles in the regulation of gene expression, and understanding the interactions between RNAs and RBPs in distinct cellular conditions forms the basis for comprehending the underlying RNA function. However, current computational methods pose challenges to the cross-prediction of RNA-protein binding events across diverse cell lines and tissue contexts. Here, we develop HDRNet, an end-to-end deep learning-based framework to precisely predict dynamic RBP binding events under diverse cellular conditions. Our results demonstrate that HDRNet can accurately and efficiently identify binding sites, particularly for dynamic prediction, outperforming other state-of-the-art models on 261 linear RNA datasets from both eCLIP and CLIP-seq, supplemented with additional tissue data. Moreover, we conduct motif and interpretation analyses to provide fresh insights into the pathological mechanisms underlying RNA-RBP interactions from various perspectives. Our functional genomic analysis further explores the gene-human disease associations, uncovering previously uncharacterized observations for a broad range of genetic disorders.

Exploration of DPP-IV Inhibitory Peptide Design Rules Assisted by the Deep Learning Pipeline That Identifies the Restriction Enzyme Cutting Site

The mining of antidiabetic dipeptidyl peptidase IV (DPP-IV) inhibitory peptides (DPP-IV-IPs) is currently a costly and laborious process. Due to the absence of rational peptide design rules, it relies on cumbersome screening of unknown enzyme hydrolysates. Here, we present an enhanced deep learning model called bidirectional encoder representation (BERT)-DPPIV, specifically designed to classify DPP-IV-IPs and explore their design rules to discover potent candidates. The end-to-end model utilizes a fine-tuned BERT architecture to extract structural/functional information from input peptides and accurately identify DPP-IV-Ips from input peptides. Experimental results in the benchmark data set showed BERT-DPPIV yielded state-of-the-art accuracy and MCC of 0.894 and 0.790, surpassing the 0.797 and 0.594 obtained by the sequence-feature model. Furthermore, we leveraged the attention mechanism to uncover that our model could recognize the restriction enzyme cutting site and specific residues that contribute to the inhibition of DPP-IV. Moreover, guided by BERT-DPPIV, proposed design rules for DPP-IV inhibitory tripeptides and pentapeptides were validated, and they can be used to screen potent DPP-IV-IPs.

Investigation of the BERT model on nucleotide sequences with non-standard pre-training and evaluation of different k-mer embeddings

MOTIVATION

In recent years, pre-training with the transformer architecture has gained significant attention. While this approach has led to notable performance improvements across a variety of downstream tasks, the underlying mechanisms by which pre-training models influence these tasks, particularly in the context of biological data, are not yet fully elucidated.

RESULTS

In this study, focusing on the pre-training on nucleotide sequences, we decompose a pre-training model of Bidirectional Encoder Representations from Transformers (BERT) into its embedding and encoding modules to analyze what a pre-trained model learns from nucleotide sequences. Through a comparative study of non-standard pre-training at both the data and model levels, we find that a typical BERT model learns to capture overlapping-consistent k-mer embeddings for its token representation within its embedding module. Interestingly, using the k-mer embeddings pre-trained on random data can yield similar performance in downstream tasks, when compared with those using the k-mer embeddings pre-trained on real biological sequences. We further compare the learned k-mer embeddings with other established k-mer representations in downstream tasks of sequence-based functional prediction. Our experimental results demonstrate that the dense representation of k-mers learned from pre-training can be used as a viable alternative to one-hot encoding for representing nucleotide sequences. Furthermore, integrating the pre-trained k-mer embeddings with simpler models can achieve competitive performance in two typical downstream tasks.

AVAILABILITY AND IMPLEMENTATION

The source code and associated data can be accessed at https://github.com/yaozhong/bert_investigation.

A systematic benchmark of machine learning methods for protein-RNA interaction prediction

RNA-binding proteins (RBPs) are central actors of RNA post-transcriptional regulation. Experiments to profile-binding sites of RBPs in vivo are limited to transcripts expressed in the experimental cell type, creating the need for computational methods to infer missing binding information. While numerous machine-learning based methods have been developed for this task, their use of heterogeneous training and evaluation datasets across different sets of RBPs and CLIP-seq protocols makes a direct comparison of their performance difficult. Here, we compile a set of 37 machine learning (primarily deep learning) methods for in vivo RBP-RNA interaction prediction and systematically benchmark a subset of 11 representative methods across hundreds of CLIP-seq datasets and RBPs. Using homogenized sample pre-processing and two negative-class sample generation strategies, we evaluate methods in terms of predictive performance and assess the impact of neural network architectures and input modalities on model performance. We believe that this study will not only enable researchers to choose the optimal prediction method for their tasks at hand, but also aid method developers in developing novel, high-performing methods by introducing a standardized framework for their evaluation.