Biology subjects

Egan, R.

Publications and source records attributed to Egan, R..

12 recordsLinked to original sources

bioRxiv2026

FP8 Inference in Genomic Foundation Models: Theoretical vs. Realized Speedups on GenomeOcean

Genomic Foundation Models (GFMs) are increasingly used for large-scale sequence analysis and generation. Compared with frontier language models, GFMs are typically smaller and frequently operate on long genomic sequences, with evaluation often requiring preservation of biologically meaningful structure and sequence-level relationships. Although low-precision post-training quantization (PTQ) has shown substantial memory and throughput benefits for general-purpose language models, it remains unclear whether these benefits transfer to GFMs given their distinct model scales, sequence characteristics, and evaluation requirements. We present an empirical case study of FP8 post-training quantization applied to GenomeOcean, a computationally efficient genomic foundation model with strong reported performance across diverse genomics tasks [Zhou et al., 2025]. Its range of model scales, from 100M to 4B parameters, provides a useful setting for examining how quantization effects vary with model size. We evaluate FP8 across two primary GFM inference regimes--embedding extraction and autoregressive generation--and assess its impact along two dimensions: biological fidelity relative to BF16 baselines and system-level efficiency in terms of throughput, memory usage, and energy efficiency. We find that FP8 largely preserves biological fidelity across the evaluated scales and inference regimes, while reducing GPU memory footprint at 4B scale and improving energy efficiency during autoregressive generation. However, realized throughput gains remain substantially below FP8s theoretical 2x hardware ceiling, with a best-case improvement of 19.3% in autoregressive generation and benefits varying strongly by model scale and workload. Autoregressive generation shows the clearest gains, driven largely by KV-cache compression, whereas embedding extraction provides limited or negative throughput benefits at smaller model scales. We attribute this theory-practice gap to the interaction of model-scale effects, memory-system bottlenecks, and software-stack limitations. These findings highlight the need for workload-specific empirical evaluation before adopting low-precision inference in scientific foundation models. Code availabilityhttps://github.com/jgi-genomeocean/genomeocean_efficiency

bioinformatics↗

bioRxiv2025

GenomeOcean: An Efficient Genome Foundation Model Trained on Large-Scale Metagenomic Assemblies

Functional genomic sequences occupy an infinitesimal fraction of the astronomically large DNA sequence space, implying that evolution explores a low-dimensional manifold shaped by universal biochemical and evolutionary constraints. To investigate this underexplored Genomic Manifold Hypothesis using scalable, data-driven evidence, we developed GenomeOcean, a 4-billion-parameter generative model trained on 219 TB of diverse, co-assembled global environmental metagenomic data using an optimized transformer. The model captures both universal phylogenetic syntax and protein-coding fidelity. Furthermore, we validate the embeddings intrinsic low dimensionality. Crucially, comparisons with the independently trained Evo 2 model demonstrate a strong linear correspondence between their embedding spaces and convergent generative behaviors. These results suggest that the genomic manifold represents a fundamental biological principle, offering a unified framework for understanding evolutionary constraints and guiding synthetic biology.

bioinformatics↗

bioRxiv2024

GenomeFace: a deep learning-based metagenome binner trained on 43,000 microbial genomes

Metagenomic binning, the process of grouping DNA sequences into taxonomic units, is critical for understanding the functions, interactions, and evolutionary dynamics of microbial communities. We propose a deep learning approach to binning using two neural networks, one based on composition and another on environmental abundance, dynamically weighting the contribution of each based on characteristics of the input data. Trained on over 43,000 prokaryotic genomes, our network for composition-based binning is inspired by metric learning techniques used for facial recognition. Using a task-specific, multi-GPU accelerated algorithm to cluster the embeddings produced by our network, our binner leverages marker genes observed to be universally present in nearly all taxa to grade and select optimal clusters of sequences from a hierarchy of candidates. We evaluate our approach on four simulated datasets with known ground truth. Our linear time integration of marker genes recovers more near complete genomes than state of the art but computationally infeasible solutions using them, while being over an order of magnitude faster. Finally, we demonstrate the scalability and acuity of our approach by testing it on three of the largest metagenome assemblies ever performed. Compared to other binners, we produced 47%-183% more near complete genomes. From these datasets, we find over the genomes of over 3000 new candidate species which have never been previously cataloged, representing a potential 4% expansion of the known bacterial tree of life.

bioinformatics↗

bioRxiv2022

Genomic features predict bacterial life history strategies in soil, as identified by metagenomic stable isotope probing

Bacteria catalyze the formation and destruction of soil organic matter, but the bacterial dynamics in soil that govern carbon (C) cycling are not well understood. Life history strategies explain the complex dynamics of bacterial populations and activities based on tradeoffs in energy allocation to growth, resource acquisition, and survival. Such tradeoffs influence the fate of soil C, but their genomic basis remains poorly characterized. We used multi-substrate metagenomic DNA stable isotope probing to link genomic features of bacteria to their C acquisition and growth dynamics. We identify several genomic features associated with patterns of bacterial C acquisition and growth, notably genomic investment in resource acquisition and regulatory flexibility. Moreover, we identify genomic tradeoffs defined by numbers of transcription factors, membrane transporters, and secreted products, which match predictions from life history theory. We further show that genomic investment in resource acquisition and regulatory flexibility can predict bacterial ecological strategies in soil.

microbiology↗

bioRxiv2022

Persistent Memory as an Effective Alternative to Random Access Memory in Metagenome Assembly

The assembly of metagenomes decomposes members of complex microbe communities and allows the characterization of these genomes without laborious cultivation or single-cell metagenomics. Metagenome assembly is a process that is memory intensive and time consuming. Multi-terabyte sequences can become too large to be assembled on a single computer node, and there is no reliable method to predict the memory requirement due to data-specific memory consumption pattern. Currently, out-ofmemory (OOM) is one of the most prevalent factors that accounts for metagenome assembly failures. In this study, we explored the possibility of using Persistent Memory (PMem) as a less expensive substitute for dynamic random access memory (DRAM) to reduce OOM and increase the scalability of metagenome assemblers. We evaluated the execution time and memory usage of three popular metagenome assemblers (MetaSPAdes, MEGAHIT, and MetaHipMer2) in datasets up to one terabase. We found that PMem can enable metagenome assemblers on terabyte-sized datasets by partially or fully substituting DRAM at a cost of longer running times. In addition, different assemblers displayed distinct memory/speed trade-offs in the same hardware/software environment. Because PMem was provided directly without any application-specific code modification, these findings are likely to be generalized to other memory-intensive bioinformatics applications.

bioinformatics↗

bioRxiv2021

Critical Assessment of Metagenome Interpretation - the second round of challenges

Evaluating metagenomic software is key for optimizing metagenome interpretation and focus of the community-driven initiative for the Critical Assessment of Metagenome Interpretation (CAMI). In its second challenge, CAMI engaged the community to assess their methods on realistic and complex metagenomic datasets with long and short reads, created from [~]1,700 novel and known microbial genomes, as well as [~]600 novel plasmids and viruses. Altogether 5,002 results by 76 program versions were analyzed, representing a 22x increase in results. Substantial improvements were seen in metagenome assembly, some due to using long-read data. The presence of related strains still was challenging for assembly and genome binning, as was assembly quality for the latter. Taxon profilers demonstrated a marked maturation, with taxon profilers and binners excelling at higher bacterial taxonomic ranks, but underperforming for viruses and archaea. Assessment of clinical pathogen detection techniques revealed a need to improve reproducibility. Analysis of program runtimes and memory usage identified highly efficient programs, including some top performers with other metrics. The CAMI II results identify current challenges, but also guide researchers in selecting methods for specific analyses.

bioinformatics↗

bioRxiv2019

A new method for rapid genome classification, clustering, visualization, and novel taxa discovery from metagenome

Current supervised phylogeny-based methods fall short on recognizing species assembled from metagenomic datasets from under-investigated habitats, as they are often incomplete or lack closely known relatives. Here, we report an efficient software suite, \"Genome Constellation\", that estimates similarities between genomes based on their k-mer matches, and subsequently uses these similarities for classification, clustering, and visualization. The clusters of reference genomes formed by Genome Constellation closely resemble known phylogenetic relationships while simultaneously revealing unexpected connections. In a dataset containing 1,693 draft genomes assembled from the Antarctic lake communities where only 40% could be placed in a phylogenetic tree, Genome Constellation improves taxa assignment to 61%. It revealed six clusters derived from new bacterial phyla and 63 new giant viruses, 3 of which missed by the traditional marker-based approach. In summary, we demonstrate that Genome Constellation can tackle the computational and algorithmic challenges in large-scale taxonomy analyses in metagenomics.

bioinformatics↗

bioRxiv2018

MiniScrub: de novo long read scrubbing using approximate alignment and deep learning

Long read sequencing technologies such as Oxford Nanopore can greatly de-crease the complexity of de novo genome assembly and large structural variation iden-tification. Currently Nanopore reads have high error rates, and the errors often cluster into low-quality segments within the reads. Many methods for resolving these errors require access to reference genomes, high-fidelity short reads, or reference genomes, which are often not available. De novo error correction modules are available, often as part of assembly tools, but large-scale errors still remain in resulting assemblies, motivating further innovation in this area. We developed a novel Convolutional Neu-ral Network (CNN) based method, called MiniScrub, for de novo identification and subsequent \"scrubbing\" (removal) of low-quality Nanopore read segments. MiniScrub first generates read-to-read alignments by MiniMap, then encodes the alignments into images, and finally builds CNN models to predict low-quality segments that could be scrubbed based on a customized quality cutoff. Applying MiniScrub to real world con-trol datasets under several different parameters, we show that it robustly improves read quality. Compared to raw reads, de novo genome assembly with scrubbed reads pro-duces many fewer mis-assemblies and large indel errors. We propose MiniScrub as a tool for preprocessing Nanopore reads for downstream analyses. MiniScrub is open-source software and is available at https://bitbucket.org/berkeleylab/jgi-miniscrub

bioinformatics↗

bioRxiv2018

Head-to-Head Comparison of Three Methods of Quantifying Competitive Fitness in C. elegans

Organismal fitness is relevant in many contexts in biology. The most meaningful experimental measure of fitness is competitive fitness, when two or more entities (e.g., genotypes) are allowed to compete directly. In theory, competitive fitness is simple to measure: an experimental population is initiated with the different types in known proportions and allowed to evolve under experimental conditions to a predefined endpoint. In practice, there are several obstacles to obtaining robust estimates of competitive fitness in multicellular organisms, the most pervasive of which is simply the time it takes to count many individuals of different types from many replicate populations. Methods by which counting can be automated in high throughput are desirable, but for automated methods to be useful, the bias and technical variance associated with the method must be (a) known, and (b) sufficiently small relative to other sources of bias and variance to make the effort worthwhile.\n\nThe nematode Caenorhabditis elegans is an important model organism, and the fitness effects of genotype and environmental conditions are often of interest. We report a comparison of three experimental methods of quantifying competitive fitness, in which wild-type strains are competed against GFP-marked competitors under standard laboratory conditions. Population samples were split into three replicates and counted (1) \"by eye\" from a saved image, (2) from the same image using CellProfiler image analysis software, and (3) with a large particle flow cytometer (a \"worm sorter\"). From 720 replicate samples, neither the frequency of wild-type worms nor the among-sample variance differed significantly between the three methods. CellProfiler and the worm sorter provide at least a tenfold increase in sample handling speed with little (if any) bias or increase in variance.

evolutionary biology↗

bioRxiv2018

Genome-wide prediction of synthetic rescue mediators of resistance to targeted and immunotherapy

Most patients with advanced cancer eventually acquire resistance to targeted therapies, spurring extensive efforts to identify molecular events mediating therapy resistance. Many of these events involve synthetic rescue (SR) interactions, where the reduction in cancer cell viability caused by targeted gene inactivation is rescued by an adaptive alteration of another gene (the rescuer). Here we perform a genome-wide prediction of SR rescuer genes by analyzing tumor transcriptomics and survival data of 10,000 TCGA cancer patients. Predicted SR interactions are validated in new experimental screens. We show that SR interactions can successfully predict cancer patients response and emerging resistance. Inhibiting predicted rescuer genes sensitizes resistant cancer cells to therapies synergistically, providing initial leads for developing combinatorial approaches to overcome resistance proactively. Finally, we show that the SR analysis of melanoma patients successfully identifies known mediators of resistance to immunotherapy and predicts novel rescuers.

cancer biology↗

bioRxiv2017

Critical Assessment of Metagenome Interpretation - a benchmark of computational metagenomics software

In metagenome analysis, computational methods for assembly, taxonomic profiling and binning are key components facilitating downstream biological data interpretation. However, a lack of consensus about benchmarking datasets and evaluation metrics complicates proper performance assessment. The Critical Assessment of Metagenome Interpretation (CAMI) challenge has engaged the global developer community to benchmark their programs on datasets of unprecedented complexity and realism. Benchmark metagenomes were generated from ~700 newly sequenced microorganisms and ~600 novel viruses and plasmids, including genomes with varying degrees of relatedness to each other and to publicly available ones and representing common experimental setups. Across all datasets, assembly and genome binning programs performed well for species represented by individual genomes, while performance was substantially affected by the presence of related strains. Taxonomic profiling and binning programs were proficient at high taxonomic ranks, with a notable performance decrease below the family level. Parameter settings substantially impacted performances, underscoring the importance of program reproducibility. While highlighting current challenges in computational metagenomics, the CAMI results provide a roadmap for software selection to answer specific research questions.

bioinformatics↗

bioRxiv2016

De novo Identification of DNA Modifications Enabled by Genome-Guided Nanopore Signal Processing

Advances in nanopore sequencing technology have enabled investigation of the full catalogue of covalent DNA modifications. We present the first algorithm for the identification of modified nucleotides without the need for prior training data along with the open source software implementation, nanoraw. Nanoraw accurately assigns contiguous raw nanopore signal to genomic positions, enabling novel data visualization, and increasing power and accuracy for the discovery of covalently modified bases in native DNA. Ground truth case studies utilizing synthetically methylated DNA show the capacity to identify three distinct methylation marks, 4mC, 5mC, and 6mA, in seven distinct sequence contexts without any changes to the algorithm. We demonstrate quantitative reproducibility simultaneously identifying 5mC and 6mA in native E. coli across biological replicates processed in different labs. Finally we propose a pipeline for the comprehensive discovery of DNA modifications in any genome without a priori knowledge of their chemical identities.

bioinformatics↗