Biology subjects

Mahmoudabadi, G.

Publications and source records attributed to Mahmoudabadi, G..

6 recordsLinked to original sources

bioRxiv2026

Species-specific small models for cell type classification approach the performance of large single cell foundation models

Accurate cross-species cell type classification remains a key evaluation task in single-cell transcriptomics. Recent foundation models trained on millions of single-cell profiles demonstrate great in-distribution and out-of-distribution performance on this task, but their large parameter counts and substantial computational costs limit accessibility and interpretability. Here, we introduce CytoType, a simple and interpretable model for cell type classification that leverages pre-trained ESM-2 protein embeddings of protein-coding transcripts. By learning linear, cell-type-specific weights over transcript embeddings, without relying on gene count information, CytoType achieves F1 scores comparable to or exceeding those of large-scale transformer-based models. We further developed ESM-Cell Embedding (ESM-CE), an even simpler variant that only averages ESM-2 embeddings across expressed genes, which also performs competitively against foundation models. Both CytoType and ESM-CE are trained on species-specific data, maintaining high accuracy when classifying cell types with orders of magnitude fewer parameters compared to larger foundation models. For example, for human tissues, the average performance gap between CytoType and the best foundation model was 0.053 F1 points while CytoType uses 10,000x fewer trainable parameters. Additionally, we quantified the contribution of ESM-2 embeddings to cell type classification tasks and demonstrated a three fold reduction in the performance gap between CytoType and the best foundation model for nine species. Finally, we show that CytoTypes learned gene weights are biologically interpretable.

cell biology↗

bioRxiv2025

A Cross-Species Generative Cell Atlas Across 1.5 Billion Years ofEvolution: The TranscriptFormer Single-cell Model

Single-cell transcriptomics has revolutionized our understanding of cellular diversity, yet our understanding of the transcriptional programs across the tree of life remains limited. Here we present TranscriptFormer, a family of generative foundation models trained on up to 112 million cells spanning 1.53 billion years of evolution across 12 species. By jointly modeling gene identities and expression levels using a novel generative architecture, TranscriptFormer encodes multi-scale biological structure, functioning as a queryable virtual cell atlas. We demonstrate state-of-the-art performance on both in-distribution and out-of-distribution cell type classification, with robust performance even for species separated by over 685 million years of evolution. TranscriptFormer can also perform zero-shot disease state identification in human cells and accurately transfers cell state annotations across species boundaries. As a generative model, TranscriptFormer can be prompted to predict cell type-specific transcription factors and gene-gene interactions that align with independent experimental observations. Developmental trajectories, phylogenetic relationships and cellular hierarchies emerge naturally in TranscriptFormers representations without any explicit training on these annotations. This work establishes a powerful framework for quantitative single-cell analysis, and comparative cellular biology, thus demonstrating that universal principles of cellular organization can be learned and predicted across the tree of life.

systems biology↗

bioRxiv2024

Machine learning models of human tissue microbiomes for tissue-of-origin prediction

There is increasing interest in using microbial data diagnostically for tissue health monitoring such as in early cancer detection. To build such models, we need to understand whether normal tissue microbiomes can also be predictive of tissue of origin, and importantly ask how contaminants may contribute to model performance. In this study, using the Tabula Sapiens Microbiome dataset, we built machine learning models of human tissue microbiomes that can predict tissue of origin. This may in part explain how tumor types can be predicted based on the tumor microbiomes. We also demonstrate that machine learning models built using contaminants alone, though not as powerful as those built on true signal, can still predict tissue of origin. Reassuringly, the addition of contaminants to true signal does not increase the performance over models built on true signal. Overall, our findings raise the burden of proof for predictive models of the human tissue and tumor microbiomes. Without addressing the magnitude of contribution from contaminants to model performance, a models reproducibility and its clinical value becomes questionable. We also discuss the optimal microbial taxonomic resolution for building these models.

bioinformatics↗

bioRxiv2024

Machine learning models can identify individuals based on a resident oral bacteriophage family

Metagenomic studies have revolutionized the study of novel phages. However these studies trade depth of coverage for breadth. We show that the targeted sequencing of a small region of a phage terminase family can provide sufficient sequence diversity to serve as an individual-specific barcode or a "phageprint, defined as the relative abundance profile of the variants within a terminase family. By collecting [~]700 oral samples from [~]100 individuals living on multiple continents, we found a consistent trend wherein each individual harbors one or two dominant variants that coexist with numerous low-abundance variants. By tracking phageprints over the span of a month across ten individuals, we observed that phageprints were generally stable, and found instances of concordant temporal fluctuations of variants shared between partners. To quantify these patterns further, we built machine learning models that, with high precision and recall, distinguished individuals even when we eliminated the most abundant variants and further downsampled phageprints to 2% of the remaining variants. Except between partners, phageprints are dissimilar between individuals, and neither country-of-residence, genetics, diet nor cohabitation seem to play a role in the relatedness of phageprints across individuals. By sampling from six different oral sites, we were able to study the impact of millimeters to a few centimeters of separation on an individuals phageprint and found that such limited spatial separation results in site-specific phageprints.

bioinformatics↗

bioRxiv2019

Defining the Energetic Costs of Cellular Structures

All cellular structures are assembled from molecular building blocks, and molecular building blocks incur energetic costs to the cell. In an energy-limited environment, the energetic cost of a cellular structure imposes a fitness cost and impacts a cells evolutionary trajectory. While the importance of energetic considerations was realized for decades, the distinction between direct energetic costs expended by the cell and potential energy that the cell diverts into cellular biomass components, which we define as the opportunity cost, was not explicitly made, leading to large differences in values for energetic costs of molecular building blocks used in the literature. We describe a framework that defines and separates various components relevant for estimating the energetic costs of molecular building blocks and the resulting cellular structures. This distinction among energetic costs is an essential step towards discussing the conversion of an energetic cost to a corresponding fitness cost.

biophysics↗

bioRxiv2019

Human Phageprints: A high-resolution exploration of oral phages reveals globally-distributed phage families with individual-specific and temporally-stable community compositions

Metagenomic studies have revolutionized the study of novel phages. However these studies trade the depth of coverage for breadth. In this study we show that the targeted sequencing of a phage genomic region as small as 200-300 base pairs, can provide sufficient sequence diversity to serve as an individual-specific barcode or \"Phageprint\". The targeted approach reveals a high-resolution view of phage communities that is not available through metagenomic datasets. By creating instructional videos and collection kits, we enabled citizen scientists to gather [~]700 oral samples spanning [~]100 individuals residing in different parts of the world. In examining phage communities at 6 different oral sites, and by comparing phage communities of individuals living across the globe, we were able to study the effect of spatial separation, ranging from several millimeters to thousands of kilometers. We found that the spatial separation of just a few centimeters (the distance between two oral sites) can already result in highly distinct phage community compositions. For larger distances, spanning the phage communities of different individuals living in different parts of the world, we did not observe any correlation between spatial distance and phage community composition as individuals residing in the same city did not have any more similar phage communities than individuals living on different continents. Additionally, we found that neither genetics nor cohabitation seem to play a role in the relatedness of phage community compositions across individuals. Cohabitating siblings and even identical twins did not have phage community compositions that were any more similar than those of unrelated individuals. The primary factor contributing to phage community composition relatedness is direct contact between two habitats, as is demonstrated by the similarity between oral phage community compositions of partners. Furthermore, by exploring phage communities across the span of a month, and in some cases several years, we observed highly stable community compositions. These studies consistently point to the existence of remarkably diverse and personal phage families that are stable in time and apparently present in people around the world.

microbiology↗