The prediction of molecular phenotypes from DNA sequences remains a longstanding challenge in genomics, often driven by limited annotated data and the inability to transfer learnings between tasks. Here, we present an extensive study of foundation models pre-trained on DNA sequences, named Nucleotide Transformer, ranging from 50 million up to 2.5 billion parameters and integrating information from 3,202 human genomes and 850 genomes from diverse species. These transformer models yield context-specific representations of nucleotide sequences, which allow for accurate predictions even in low-data settings. We show that the developed models can be fine-tuned at low cost to solve a variety of genomics applications. Despite no supervision, the models learned to focus attention on key genomic elements and can be used to improve the prioritization of genetic variants. The training and application of foundational models in genomics provides a widely applicable approach for accurate molecular phenotype prediction from DNA sequence.

Closing the gap between measurable genetic information and observable traits is a longstanding challenge in genomics. Yet, the prediction of molecular phenotypes from DNA sequences alone remains limited and inaccurate, often driven by the scarcity of annotated data and the inability to transfer learnings between prediction tasks. Here, we present an extensive study of foundation models pre-trained on DNA sequences, named the Nucleotide Transformer, ranging from 50M up to 2.5B parameters and integrating information from 3,202 diverse human genomes, as well as 850 genomes selected across diverse phyla, including both model and non-model organisms. These transformer models yield transferable, context-specific representations of nucleotide sequences, which allow for accurate molecular phenotype prediction even in low-data settings. We show that the developed models can be fine-tuned at low cost and despite low available data regime to solve a variety of genomics applications. Despite no supervision, the transformer models learned to focus attention on key genomic elements, including those that regulate gene expression, such as enhancers. Lastly, we demonstrate that utilizing model representations can improve the prioritization of functional genetic variants. The training and application of foundational models in genomics explored in this study provide a widely applicable stepping stone to bridge the gap of accurate molecular phenotype prediction from DNA sequence. Code and weights available at: https://github.com/instadeepai/nucleotide-transformer in Jax and https://huggingface.co/InstaDeepAI in Pytorch. Example notebooks to apply these models to any downstream task are available on https://huggingface.co/docs/transformers/notebooks#pytorch-bio.

Centrosomes are the major microtubule organising centres of animal cells. Deregulation in their number occurs in cancer and was shown to trigger tumorigenesis in mice. However, the incidence, consequence and origins of this abnormality are poorly understood. Here, we screened the NCI-60 panel of human cancer cell lines to systematically analyse centriole number and structure. Our screen shows that centriole amplification is widespread in cancer cell lines and highly prevalent in aggressive breast carcinomas. Moreover, we identify another recurrent feature of cancer cells: centriole size deregulation. Further experiments demonstrate that severe centriole over-elongation can promote amplification through both centriole fragmentation and ectopic procentriole formation. Furthermore, we show that overly long centrioles form over-active centrosomes that nucleate more microtubules, a known cause of invasiveness, and perturb chromosome segregation. Our screen establishes centriole amplification and size deregulation as recurrent features of cancer cells and identifies novel causes and consequences of those abnormalities.

Breast cancer is a highly heterogeneous disease resulting in diverse clinical behaviours and therapeutic responses. DNA methylation is a major epigenetic alteration that is commonly perturbed in cancers. The aim of this study is to characterize the relationship between DNA methylation and aberrant gene expression in breast cancer. We analysed DNA methylation and gene expression profiles from breast cancer tissue and matched normal tissue in The Cancer Genome Atlas (TCGA). Genome-wide differential methylation analysis and methylation-gene expression correlation was performed. Gene expression changes were subsequently validated in the METABRIC dataset. The Oncoscore tool was used to identify genes that had previously been associated with cancer in the literature. A subset of genes that had not previously been studied in cancer was chosen for further analysis. We identified 368 CpGs that were differentially methylated between tumor and normal breast tissue (∆β > 0.4). Hypermethylated CpGs were overrepresented in tumor tissue and were found predominantly (56%) in upstream promoter regions. Conversely, hypomethylated CpG sites were found primarily in the gene body (66%). Expression analysis revealed that 209 of the differentially-methylated CpGs were located in 169 genes that were differently expressed between normal and breast tumor tissue. Methylation-expression correlations were predominantly negative (70%) for promoter CpG sites and positive (74%) for gene body CpG sites. Among these differentially-methylated and differentially-expressed genes, we identified 7 that had not previously been studied in any form of cancer. Three of these, TDRD10, PRAC2 and TMEM132C, contained CpG sites that showed diagnostic and prognostic value in breast cancer, particularly in estrogen-receptor (ER)-positive samples. A pan-cancer analysis confirmed differential expression of these genes together with diagnostic and prognostic value of their respective CpG sites in multiple cancer types. We have identified 368 DNA methylation changes that characterize breast cancer tumor tissue, of which 209 are associated with genes that are differentially-expressed in the same samples. Novel DNA methylation markers were identified, of which cg12374721 (PRAC2), cg18081940 (TDRD10) and cg04475027 (TMEM132C) show promise as diagnostic and prognostic markers in breast cancer as well as other cancer types.

Mutations causing aberrant splicing are frequently implicated in human diseases including cancer. Here, we establish a high-throughput screen of randomly mutated minigenes to decode thecis-regulatory landscape that determines alternative splicing of exon 11 in the proto-oncogeneMST1R(RON). Mathematical modelling of splicing kinetics enables us to identify more than 1000 mutations affectingRONexon 11 skipping, which corresponds to the pathological isoform RON∆165. Importantly, the effects correlate withRONalternative splicing in cancer patients bearing the same mutations. Moreover, we highlight heterogeneous nuclear ribonucleoprotein H (HNRNPH) as a key regulator ofRONsplicing in healthy tissues and cancer. Using iCLIP and synergy analysis, we pinpoint the functionally most relevant HNRNPH binding sites and demonstrate how cooperative HNRNPH binding facilitates a splicing switch ofRONexon 11. Our results thereby offer insights into splicing regulation and the impact of mutations on alternative splicing in cancer.