Biology subjects

Diaz-Navarro, A.

Publications and source records attributed to Diaz-Navarro, A..

4 recordsLinked to original sources

bioRxiv2025

Interpreting Attention Mechanisms in Genomic Transformer Models: A Framework for Biological Insights

Transformer models have shown strong performance on biological sequence prediction tasks, but the interpretability of their internal mechanisms remains underexplored. Given their application in biomedical research, understanding the mechanisms behind these models predictions is crucial for their widespread adoption. We introduce a method to interpret attention heads in genomic transformers by correlating per-token attention scores with curated biological annotations, and we use GPT-4 to summarize each heads focus. Applying this to DNABERT, Nucleotide Transformer, and scGPT, we find that attention heads learn biologically meaningful associations during self-supervised pre-training and that these associations shift with fine-tuning. We show that interpretability varies with tokenization scheme, and that context-dependence plays a key role in head behaviour. Through ablation, we demonstrate that heads strongly associated with biological features are more important for task performance than uninformative heads in the same layers. In DNABERT trained for TATA promoter prediction, we observe heads with positive and negative associations reflecting positive and negative learning dynamics. Our results offer a framework to trace how biological features are learned from random initialization to pre-training to fine-tuning, enabling insight into how genomic foundation models represent nucleotides, genes, and cells.

bioinformatics↗

bioRxiv2025

BRCA1 is a molecular correlate of cell proliferation in human brain development and in Group 3 and 4 medulloblastoma

The role of the BRCA1-mediated DNA damage repair pathway in regulating human brain development remains unknown, although it has been studied in mouse development. We report evidence for breast cancer type 1 susceptibility protein (BRCA1) being a molecular correlate of proliferation in human neural progenitor cells and in medulloblastoma (MB), a malignant pediatric hindbrain cancer whose cells resemble undifferentiated neural stem cells. In a computational search for molecules potentially keeping Group 3 (G3) and Group 4 (G4) MB tumour cells in a state of stalled differentiation, BRCA1 emerged as a leading candidate gene. We surveyed four independent transcriptomic datasets collectively spanning 142 human developing brain samples, multiple brain regions, and over 1.7 million single cells, and found that BRCA1 transcription is consistently enriched in human neural stem and progenitor cells, relative to differentiating or mature neurons. Across the human lifespan, BRCA1 expression is enriched in the brain during early development, particularly the first trimester of gestation. By analyzing 714 tumours, the largest transcriptomic survey of MB tumours to date, we found that BRCA1 expression is increased in carriers of isochromosome 17q (i17q) aberrations and in G4 MB tumours. Increased BRCA1 expression is associated with worse prognosis in G3 and G4 MB. In the developing cerebellum as well as in the cancer context, BRCA1 expression is correlated with transcription of the cell cycle and DNA damage repair pathways. When considered with previous mouse studies, our work is consistent with a model in which BRCA1 promotes genome surveillance in neural progenitors during human brain development and in G3 and G4 MB tumour growth, thus supporting proliferation.

genomics↗

bioRxiv2024

In silico generation of synthetic cancer genomes using generative AI

Cancer originates from alterations in the genome, and understanding how these changes lead to disease is crucial for achieving the goals of precision oncology. Connecting genomic alterations to health outcomes requires extensive computational analysis using accurate algorithms. Over the years, these algorithms have become increasingly sophisticated, but a severe shortage of open access gold-standard datasets presents a fundamental challenge. Since genomic data is considered personal health information, only an extremely limited number of deeply sequenced legacy cancer genomes can be shared and redistributed. As a result, tool benchmarking is often conducted on the same small set of genomes sequenced with older technologies and uncertain ground truths. This is a major obstacle to the development of improved analytic tools. To address this issue, we have developed OncoGAN, a novel generative AI tool that uses a combination of generative adversarial networks and tabular variational autoencoders to generate realistic but entirely synthetic cancer genomes based on training sets derived from large-scale genomic projects. Our results demonstrate that this approach accurately reproduces the scale, distribution, and characteristics of somatic point mutations, copy number alterations and structural variants across multiple common cancer types, while protecting donors privacy information. OncoGAN accurately recapitulates tumor type-specific mutational signatures as well as the positional distribution of somatic mutations. To evaluate the fidelity of the simulations, we tested the synthetic genomes using DeepTumour, a software capable of identifying tumor types based on mutational patterns, and demonstrated a high level of concordance between the synthetic genome tumor type and DeepTumours prediction of the type. We also showed that augmenting real donor data with OncoGAN-generated synthetic data could be used to train a more accurate version of DeepTumour. This tool will allow the generation of an extensive and realistic set of training and testing cancer genomes whose ground truth is known exactly. This advance provides computational biologists with the ability to develop realistic cancer genome benchmarking sets and make them available to the research community for the testing, development and enhancement of cancer genome analysis tools.

bioinformatics↗

bioRxiv2022

RFcaller: a machine learning approach combined with read-level features to detect somatic mutations

MotivationThe cost reduction in sequencing and the extensive genomic characterization of a wide variety of cancers is expanding the use of tumor sequencing approaches to a wide number of research groups and to the clinical practice. Although specific pipelines have been generated for the identification of somatic mutations, their results usually differ considerably, and a common approach in many projects is to use several callers to achieve a more reliable set of mutations. This procedure is computationally very expensive and time-consuming, and it suffers from the same limitations in sensitivity and specificity as other approaches. Expert revision of mutant calls is therefore required to verify calls that might be used for clinical diagnosis. Machine learning techniques provide a useful approach to incorporate expert-reviewed information for the identification of somatic mutations. ResultsWe have developed RFcaller, a pipeline based on machine learning algorithms, for the detection of somatic mutations in tumor-normal paired samples. RFcaller shows high accuracy for the detection of substitutions and indels from whole genome or exome data. It allows the detection of mutations in driver genes missed by other approaches, and has been validated by comparison to deep sequencing and Sanger sequencing. The pipeline is able to analyze a whole genome in a small period of time, and with a small computational footprint. Availability and implementationRFcaller is available at GitHub repository (https://github.com/xa-lab/RFcaller) and DockerHub (https://hub.docker.com/repository/docker/labxa/rfcaller). Contactxspuente@uniovi.es Supplementary informationSupplementary data is available online.

cancer biology↗