Watch a 42-minute AutoML Seminar presentation exploring multi-objective Bayesian optimization techniques for biomedical and molecular data analysis workflows, particularly focusing on unsupervised bioinformatics problems. Learn how to tackle hyperparameter optimization challenges when dealing with undefined objectives by using multiple noisy heuristic metrics. Discover a novel method that infers useful heuristic objectives based on domain-specific criteria and adaptively updates scalarization functions through multi-output Gaussian process surrogate functions. Follow along as the speaker demonstrates practical applications in single-cell RNA sequencing and highly multiplexed imaging datasets, showing how this approach effectively handles clustering analyses where traditional optimization methods struggle. Gain insights into how this method successfully identifies biologically meaningful groupings of cells based on expression profiles, evaluates cluster separation, and validates results against expert annotations in Imaging Mass Cytometry data analysis.
Multi-objective Bayesian Optimization with Heuristic Objectives for Biomedical and Molecular Data Analysis Workflows
Overview
Syllabus
Multi-objective Bayesian Optimization with Heuristic Objectives for Biomedical and Molecular Data Analysis Workflows
Overview
Bayesian optimization in bioinformatics has been applied to protein design
Most bioinformatic analyses are unsupervised
A typical workflow involves many steps and parameters
Difficulty with defining objectives make AutoML challenging to apply in bioinformatics
AutoML approaches construct objectives for a given problem
Motivation for AutoGeneS constructed objectives
Our method automatically infers which objectives are useful to guide optimization
MOBO basics
We build on the random scalarizations approach that returns a subset of the Pareto front
We determine the region of the Pareto front using objective behaviours
Three examples of desirable behaviours
Toy data simulating useful and not useful objectives
Optimizing cofactor in the analysis of Imaging Mass Cytometry (IMC) data
We construct objectives for clustering workflow using pairs of co-expressed proteins
We construct two meta-objectives using expert annotations
Parameters selected by our method led to clusterings that agree with expert annotations
Quantitative evaluation of performance
Summary
Taught by
AutoML Seminars
