Alan E. Murphy

The convergence of artificial intelligence (AI) and biotechnology is creating unprecedented capabilities for designing biological sequences, ranging from proteins and nucleic acids to entire genomes. However, significant barriers hinder the translation of generative AI models into real-world biological applications, including biased and fragmented training data, opaque model architectures that resist interpretation, and the increasing disparity between the speed of computational sequence generation and the throughput of experimental validation. As these barriers are overcome, the dual-use potential of these technologies is simultaneously expanded: the same advances that enable therapeutic design also lower the threshold for misuse, ranging from the generation of new pathogens to circumventing existing biosecurity screening. In order to address this issue, the Wübben Sandpit on Generative AI for Biological Sequences was convened, bringing together researchers from the fields of molecular biology, AI, biosecurity, ethics and policy. This white paper presents the resulting consensus recommendations, structured as a defence-in-depth framework spanning three domains. Firstly, the governance of sensitive training data should be achieved through risk-stratified access and funding-stage review. Secondly, the governance of models is addressed through lifecycle risk assessment, safety benchmarking and tiered deployment. Thirdly, the governance of nucleic acid synthesis is strengthened through screening, supply chain accountability and dedicated funding for countermeasure research. Rather than treating security as an impediment to innovation, the framework positions responsible governance as a prerequisite for realising the transformative potential of AI-driven biological design.

Anxiety is heritable and exists on a continuum, with symptoms ranging from adaptive threat response to clinical disorder. Here we performed a genome-wide association meta-analysis of generalized anxiety symptom severity in 693,869 individuals of European ancestry from 14 cohorts. We identified 80 independent genome-wide significant variants within 74 loci, 39 of which were newly associated with anxiety. SNP-based heritability was 5.9% (posterior s.d. = 0.15%). Polygenic scores were significantly associated with anxiety symptom severity and disorder in European, African and South Asian ancestry samples ( R ² = 1.2–2.9%). Significant genetic correlations ( r _g ) were estimated with mental and physical health traits, including case–control anxiety, neuroticism and depression ( r _g = 0.71–0.85), irritable bowel syndrome ( r _g = 0.57), coronary artery disease, endometriosis and migraine ( r _g = 0.20–0.27). Gene-based and pathway analyses implicated synaptic and axonal processes, with enriched expression in the brain. These findings highlight the discovery power gained from analysing a quantitative trait rather than a case–control phenotype in anxiety genetics.

Motivation: Mapping genes across identifier systems and species is a routine but critical step in bioinformatics workflows. Despite its ubiquity, gene mapping is frequently handled using bespoke, ad hoc solutions, increasing duplicated effort and introducing opportunities for error. These issues are exacerbated by the prevalence of non-one-to-one homolog relationships and inconsistent handling of gene identifiers across species and databases, which can compromise downstream analyses and reproducibility. Results: We present orthogene, an R/Bioconductor package that simplifies gene mapping within and across hundreds of species. orthogene provides a unified, workflow-oriented framework that integrates automated species and identifier standardization, homolog inference across multiple databases, flexible handling of ambiguous homolog relationships, and transformation of gene lists, tables, and high-dimensional matrices into analysis-ready formats. By abstracting common sources of technical complexity while retaining user control, orthogene enables transparent, reproducible, and scalable gene mapping across a wide range of biological contexts.

Introductory paragraph: Panic attacks, sudden episodes of intense fear accompanied by physical and psychological symptoms, affect approximately 23% of the population ^1,2 . Panic disorder, diagnosed in 2– 4% ³ , involves recurrent attacks followed by persistent worry about further attacks, leading to functional impairment and avoidance behaviours ^1,2 . We conducted genome-wide association meta-analyses of panic attacks and panic disorder (61,746 and 29,775 cases, respectively, and 105,814 controls), identifying the first genome-wide significant variants for both traits (16 for panic attacks; 7 for panic disorder). Geneset analysis using single-cell RNA sequencing data implicated afferent neurons from the eye, heart, and lung in panic attacks, suggesting a critical role for sensory processing and interoceptive awareness. Further analyses suggested that these associations generalize to other psychiatric disorders. These findings offer novel insights into the biological mechanisms underlying panic, the role of afferent neurons, and may inform the development of more targeted and effective interventions.

To understand the complex relationship between histone mark activity and gene expression, recent advances have used in silico predictions based on large-scale machine learning models. However, these approaches have omitted key contributing factors like cell state, histone mark function or distal effects, which impact the relationship, limiting their findings. Moreover, downstream use of these models for new biological insight is lacking. Here, we present the most comprehensive study of this relationship to date – investigating seven histone marks in eleven cell types across a diverse range of cell states. We used convolutional and attention-based models to predict transcription from histone mark activity at promoters and distal regulatory elements. Our work shows that histone mark function, genomic distance and cellular states collectively influence a histone mark’s relationship with transcription. We found that no individual histone mark is consistently the strongest predictor of gene expression across all genomic and cellular contexts. This highlights the need to consider all three factors when determining the effect of histone mark activity on transcriptional state. Furthermore, we conducted in silico histone mark perturbation assays, uncovering functional and disease related loci and highlighting frameworks for the use of chromatin deep learning models to uncover new biological insight.

Understanding how genetic variants affect the epigenome is key to interpreting GWAS, yet profiling these effects across the non-coding genome remains challenging due to experimental scalability. This necessitates accurate computational models. Existing machine learning approaches, while progressively improving, are confined to the cell types they were trained on, limiting their applicability. Here, we introduce Enformer Celltyping, a deep learning model which incorporates distal effects of DNA interactions, up to 100,000 base-pairs away, to predict epigenetic signals in previously unseen cell types. Using DNA and chromatin accessibility data for epigenetic imputation, Enformer Celltyping outperforms current best-in-class approaches and generalises across cell types and biological regions. Moreover, we propose a framework for evaluating models on genetic variant effect prediction using regulatory quantitative trait loci mapping studies, highlighting current limitations in genomic deep learning models. Despite this, Enformer Celltyping can also be used to study cell type-specific genetic enrichment of complex traits.

To understand the complex relationship between histone mark activity and gene expression, recent advances have used in silico predictions based on large-scale machine learning models. However, these approaches have omitted key contributing factors like cell state, histone mark function or distal effects, that impact the relationship, limiting their findings. Moreover, downstream use of these models for new biological insight is lacking. Here, we present the most comprehensive study of this relationship to date - investigating seven histone marks, in eleven cell types, across a diverse range of cell states. We used convolutional and attention-based models to predict transcription from histone mark activity at promoters and distal regulatory elements. Our work shows that histone mark function, genomic distance and cellular states collectively influence a histone mark’s relationship with transcription. We found that no individual histone mark is consistently the strongest predictor of gene expression across all genomic and cellular contexts. This highlights the need to consider all three factors when determining the effect of histone mark activity on transcriptional state. Furthermore, we conducted in silico histone mark perturbation assays, uncovering functional and disease related loci and highlighting frameworks for the use of chromatin deep learning models to uncover new biological insight. Graphical

Understanding genetic variants effects on the epigenome is crucial for interpreting genome-wide association studies (GWAS) results, yet profiling these effects across the non-coding genome remains challenging due to the scalability limits of experimental methods. This necessitates accurate computational models. Existing machine learning approaches, while progressively improving, are confined to the cell types they were trained on, limiting their applicability. Here, we propose the development of a deep learning model, Enformer Celltyping, which can both incorporate distal effects of DNA interactions, up to 100,000 base-pairs away, and predict epigenetic signals in a cell type-agnostic fashion. We demonstrate that Enformer Celltypings predictions of the epigenome out-perform the current best-in-class approach, have strong performance in a range of different cell types and biological regions and generalise to cell types assayed independently of the original training set. Finally, we propose an approach to test epigenetic models performance on genetic variant effect predictions using regulatory quantitative trait loci mapping studies, highlighting Enformer Celltypings and other genomic deep learning models limitations for this task. Our work introduces a new, accurate model for cell type-agnostic histone mark predictions and highlights the benefit of transfer learning for more efficient model development for deep learning in genomics. We make our customisable and efficient transfer learning approach for Enformer available to the community.