Gabija Kavaliauskaite

EXPERIENCE

About

Hello there,

I’m Gabi, a PhD Fellow in Computational Biology by day and a professional patissier by evening. I was born and raised in the small Baltic country of Lithuania, in a town known for its convergence of three rivers and endless forests. My love for nature was imprinted in me from a young age through countless camping trips and car journeys exploring the country's most remote places.

This deep appreciation for nature led me to pursue a career in functional genomics, dedicating my years to understanding the mechanisms behind the genome landscape. I’ve followed a unique path in learning the laws of nature, guided by the belief that, as Roger Bacon wrote in Opus Majus (1267), mathematics is both the door and the key to the sciences.

Driven by my curiosity about mathematical models, I found my way to Denmark, where I first delved into designing experiments for pattern recognition and data-driven insights into cis-regulatory elements. This experience led me to pursue a PhD at the University of Southern Denmark under the supervision of Assoc. Prof. Jesper Grud Skat Madsen. Here, I design and develop models to understand cellular differentiation pathways and link these complex processes to the unique genetic variations between individuals.

My work on differentiation processes has also led me to the Novo Nordisk Foundation Center for Genomic Mechanisms of Disease at the Broad Institute, where I serve as a Co-Lead in the Molecular QTL Track. In this role, I develop statistical and deep learning models to link genetic variation to cellular fate.

Develop a passion for learning. If you do, you will never cease to grow. — Anthony J. D’Angelo

Dream Challenge

Genes, the fundamental units of inheritance, drive various phenotypes through their context-specific expression. Variations in these genes can alter their functions and disrupt biological processes, potentially leading to diseases. Understanding the relationship between genetic variations and complex traits requires exploring the mechanisms that regulate gene expression. In eukaryotes, these regulatory processes are primarily controlled by cis-regulatory elements, such as transcription factors (TFs) that bind to specific regulatory sequences, resulting in unique gene expression patterns [1, 2]. The involvement of multiple TFs adds to the complexity of cis-regulation, emphasizing the intricate nature of gene regulation.

Multiple neural network-based architectures have been tested for genomic applications, but the challenge remains in how these models generalize when introduced to new, non-ad hoc data. To address this issue, the Random Promoter DREAM Challenge was organized by the de Boer Lab (School of Biomedical Engineering at UBC) and the DREAM Challenge committee [3]. In this competition, participants were asked to train machine learning models to predict gene expression from promoters with random DNA sequences.

Our team, MadLab, consisting of postdoc Andreas Møller, PhD Fellow Gabija Kavaliauskaite, and Associate Professor Jesper Madsen, developed a model with three key components: a convolutional network, a transformer, and a recurrent network. The model processes one-hot encoded sequences, which are fed into four separate 1D convolutions with increasing kernel sizes (15-30). The purpose of using convolutions with different kernel sizes is to learn sequence features at varying scales. The outputs of these convolutions are then combined and passed into an Enformer-style transformer with relative positional embedding [4]. Finally, the transformer's output is fed into a bidirectional LSTM layer followed by a linear activation. We included a bidirectional LSTM layer to capture both forward and backward regions within the sequence, enhancing the model's predictive power. For a more detailed description of the model architecture, please refer to the link provided.

To test our model, we split the data into training (N = 4,703,150) and validation (N = 95,982) sets, using a 98-2 percent split. We trained our model using mean squared error, and the model with the lowest validation Pearson correlation (0.5907) was selected for further evaluation.

The committee evaluated the models based on several criteria: a) the ability of the model to predict gene expression given a sequence; and b) the model's capability to predict changes in gene expression from closely related sequences. This included predictions based on single nucleotide variants (SNVs), perturbations of specific TF binding sites, and tiling of TF binding sites across background sequences. Each model's performance was evaluated using Pearson r^2 and Spearman ρ. The challenge ran for 12 weeks with 292 registered participants, and the MadLab model ranked 21st on the final leaderboard. All models from the challenge can be found here.

References:

1. Worsley-Hunt, R., Bernard, V. & Wasserman, W.W. Identification of cis-regulatory sequence variations in individual genome sequences. Genome Med 3, 65 (2011). https://doi.org/10.1186/gm281

2. Pividori, M., Lu, S., Li, B. et al. Projecting genetic associations through gene expression patterns highlights disease etiology and drug mechanisms. Nat Commun 14, 5562 (2023). https://doi.org/10.1038/s41467-023-41057-4

3. Rafi, A. M., Nogina, D., Penzar, D., Lee, D., Lee, D., Kim, N., Kim, S., Kim, D., Shin, Y., Kwak, I. Y., Meshcheryakov, G., Lando, A., Zinkevich, A., Kim, B. C., Lee, J., Kang, T., Vaishnav, E. D., Yadollahpour, P., Random Promoter DREAM Challenge Consortium, Kim, S., … de Boer, C. (2024). Evaluation and optimization of sequence-based gene regulatory deep learning models. bioRxiv : the preprint server for biology, 2023.04.26.538471. https://doi.org/10.1101/2023.04.26.538471

4. Avsec, Ž., Agarwal, V., Visentin, D. et al. Effective gene expression prediction from sequence by integrating long-range interactions. Nat Methods 18, 1196–1203 (2021). https://doi.org/10.1038/s41592-021-01252-x

Contact me

Email Email me at: kavaliauskaite.g.e@gmail.com

• Twitter
• GitHub
• LinkedIn

Elements

i = 0;

while (!deck.isInOrder()) {
    print 'Iteration ' + i;
    deck.shuffle();
    i++;
}

print 'It took ' + i + ' iterations to sort the deck.';