oFVSD: optimized Forward Variable Selection Decoder

Introduction  
The cutting-edge oFVSD toolbox, a revolutionary decoding package tailored for neuroimaging data offers unparalleled precision and efficiency. My methodology transforms the challenge of high-dimensional data into an opportunity for accurate and meaningful analysis, employing a sophisticated forward variable selection (FVS) algorithm combined with hyper-parameter optimization.

The oFVSD toolbox stands out by implementing 18 versatile ML models by default, each rigorously evaluated through k-fold cross-validation within the FVS framework. This process not only determines the optimal subset of features for each model but also refines the hyperparameters dynamically, ensuring each model is finely tuned for the highest performance. Moreover, the oFVSD toolbox is designed for practicality, supporting execution in a parallel computing environment.

Funding: Moonshot R&D Program (2-years project: ¥3,000,000 including staff salary + + academic conference + paper fee) 

Paper: Frontiers in Neuroinformatics 

Open-source code: https://github.com/tungtokyo1108/FVS_decoder
Language: Python

SVVS: Stochastic Variational Variable Selection

Introduction  
Introducing the groundbreaking Stochastic Variational Variable Selection (SVVS) method—a transformative advance in the field of microbial community analysis. This innovative approach reshapes our ability to manage and interpret the vast, complex data typical of microbiome research, setting a new standard for accuracy and efficiency.

SVVS introduces three significant improvements that directly address the challenges faced by traditional Dirichlet multinomial mixture (DMM) models:

1. Enhanced Precision in Feature Selection: SVVS integrates a novel indicator variable that identifies key operational taxonomic units essential for cluster differentiation. This innovation allows for a refined focus on the most impactful microbial species, significantly enhancing the model's accuracy and the relevance of its findings.

2. Advanced Computational Efficiency: the computation of SVVS is the application of stochastic variational inference, a sophisticated technique that utilizes a controllable variational distribution. This approach, combined with stochastic optimization algorithms, facilitates incredibly fast computations. By drastically reducing the time required to process complex datasets, SVVS enables the analysis of larger datasets than ever before, achieving results in a fraction of the time taken by traditional methods.

3. Dynamic Cluster Estimation: SVVS incorporates Dirichlet process mixtures to transition into an infinite model framework. This adaptation allows for automatic estimation of the number of clusters based on the data itself, introducing a level of adaptability and scale previously unattainable. 

Funding: Research Fellowships for Young Scientists and Grants-in-Aid for Scientific Research-KAKEN (2-years project: ¥6,200,000 including staff salary + compuational machine cost + academic conference + paper fee) 

Paper: Microbiome (2022 Impact Factor: 16.837) 

Open-source code: https://github.com/tungtokyo1108/SVVS 

Language: Python

SVBPhylo: Stochastic Variational Inference for Bayesian Phylogenetics

Introduction  
A significant advancement in phylogenetic analysis with our new variational Bayesian procedure, was specifically designed to enhance the PhyloBayes MPI program. This innovative approach is set to revolutionize genomic studies by significantly reducing the computational demand of Bayesian analysis, particularly in the context of large and complex genomic datasets.

Traditional methods, while robust, rely on computationally intensive Markov chain Monte Carlo (MCMC) sampling techniques, which become increasingly impractical as data sets grow in size and complexity. Our new method cleverly circumvents this issue by introducing a variational Bayesian approach that approximates the (unknown) posterior distribution with a controllable variational distribution. What sets this method apart is its optimization strategy: rather than sampling, it estimates the parameters of the variational distribution by minimizing Kullback-Leibler divergence. This innovative technique not only speeds up the computation dramatically but also maintains the accuracy and integrity of the Bayesian inference.

Funding: Grant-in-Aid for Scientific Research (B) (3-years project: ¥1,000,000 including compuational machine cost + academic conference + paper fee) 

Paper: Molecular Biology and Evolution (2019 Impact Factor: 14.797) 

Open-source code: https://github.com/tungtokyo1108/My-Project--Machine-Learning-Algorithm-in-Parallel-Environment-for-Biological-Computation 

Language: C++