Application of Autoencoders for Metagenomics Binning - Master Thesis in Computer Engineering

Aug 2, 2026

Upload by Oliver

Metagenomics is a discipline analyzing DNA sequences from entire microbial communities, offering strong impact in medicine and environmental monitoring. Large input datasets create major computational barriers, which deep learning methods increasingly address. This master thesis targets the contigs binning problem, clustering partially assembled DNA sequences, and introduces an autoencoder-based architecture named XAVAMB. A comprehensive evaluation against state of the art shows marked effectiveness, especially for strain identification at the most granular taxonomic level.

LanguageEnglish

CategoryResearch & Report

FormatPDF

Pages81

Reading Time3 h 24 min

1 Introduction

2 Basic concepts

2.1 Machine Learning

2.2 Deep Learning and Generative Models

2.3 Metagenomics and Metagenomics Binning

3 Related Works

4 XAVAMB architecture

5 Experiments Overview

5.1 Datasets and Evaluation Metrics

5.2 Hyperparameters Search Details

5.3 Intermediate Results

5.4 Bins Refinement

5.5 Distribution of contigs among clusters

5.6 Specificity of the architectures

5.7 Assistant Deep Clustering integration

6 Discussion

6.1 Estimating k has a good yet ambiguous effect

6.2 Contrastive Learning is not advisable for Adversarial Training

6.3 PC-mers usage is not effective yet contains unique information

6.4 XAVAMB’s main issue is instability

6.5 Assistant Deep Clustering has a stabilizing effect on XAVAMB

6.6 Performance and Quality Considerations

7 Conclusions and Future Works

8 Appendix

8.1 Computational resources for all datasets

8.2 Number of recovered species and genera

References

Acknowledgments

Master Thesis in Computer Engineering Application of autoencoders for metagenomics binning Master Candidate Lorenzo Spina Student ID 2054139 Supervisor Prof. Cinzia Pizzi University of Padova Academic Year 2023/2024 Graduation Date 24 April 2024 "Study, learn more, learn forever"Vladimir Il’ič Ul’janov Abstract Metagenomics, a scientific discipline focused on the analysis of DNA sequences derived from entire microbial communities, holds significant promise across various fields, such as in Medicine and Environmental Monitoring. However, it presents formidable computational hurdles due to the large sizes of input data. Machine Learning techniques, particularly Deep Learning, have emerged as powerful tools to tackle these challenges. In this thesis, we focus on the contigs binning problem, that is the clusterization of partially assembled DNA sequences, and propose an autoencoder-based architecture denoted as XAVAMB. Through a comprehensive evaluation against the State of theArt, our model demonstrated notable efficacy, particularly in discerning strains at the most granular taxonomic level. Sommario La Metagenomica, disciplina scientifica focalizzata sull’analisi di sequenze di DNA derivatedaintere comunità microbiche, offresignificative promessein vari settori, come la medicina e il monitoraggio ambientale. Tuttavia, essa presenta notevoli ostacoli computazionali a causa delle grandi dimensioni dei dati di input. Le tecniche di Machine Learning, in particolare il Deep Learning, sono emerse come potenti strumenti per affrontare tali sfide. In questa tesi, ci concentriamo sul problema del raggruppamento dei contigs, cioè la clusterizzazione di sequenze di DNA parzialmente assemblate, e proponiamo un’architettura basata su autoencoder denominata XAVAMB. Attraverso una valutazione completa rispetto allo stato dell’arte, il nostro modello ha dimostrato un’elevata efficacia, in particolare nel discernere ceppi al livello tassonomico più granulare. Contents List of Figures xi List of Tables xiii 1 Introduction 1 2 Basic concepts 3 2.1 Machine Learning ................................ 3 2.2 Deep Learning and Generative Models .................... 7 2.3 Metagenomics and Metagenomics Binning ................. 11 3 Related Works 15 4 XAVAMB architecture 19 5 Experiments Overview 23 5. 1 Datasets and Evaluation Metrics . . . . . . . . . . . . . . . . . . . . . . . 23 5.2 Hyperparameters Search Details . . . . . . . . . . . . . . . . . . . . . . . 25 5.2. 1 Estimation of Number of Clusters . . . . . . . . . . . . . . . . . . 25 5.2.2 Features Extraction and Contrastive Learning . . . . . . . . . . . 26 5.3 Intermediate Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26 5.4 Bins Refinement . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 32 5.5 Distribution of contigs among clusters .................... 36 5.6 Specificity of the architectures ......................... 37 5.7 Assistant Deep Clustering integration .................... 38 6 Discussion 47 6.1 Estimating k has a good yet ambiguous effect ................ 47 6.2 Contrastive Learning is not advisable for Adversarial Training ...... 48 6.3 PC-mers usage is not effective yet contains unique information ...... 49 6.4 XAVAMB’s main issue is instability ...................... 50 6.5 Assistant Deep Clustering has a stabilizing effect on XAVAMB ...... 50 6.6 Performance and Quality Considerations .................. 51 CONTENTS 7 Conclusions and Future Works 53 8 Appendix 55 8.1 Computational resources for all datasets ................... 55 8.2 Number of recovered species and genera .................. 59 References 61 Acknowledgments 67

FAQs

What problem does the thesis address in metagenomics?

It focuses on contigs binning, i.e., clustering partially assembled DNA sequences to group contigs into bins.

What model does the thesis propose?

The thesis proposes XAVAMB, an autoencoder-based architecture designed for contigs binning in metagenomics.

How is XAVAMB evaluated and what is its main benefit?

The model is comprehensively evaluated against state-of-the-art methods, showing notable effectiveness, especially for discerning strains at the most granular taxonomic level.

Aug 2, 2026

Upload by Oliver