Cancer has been known as a deadly and complex disease to tackle. By applying machine learning algorithms we hope to improve personalized treatment for cancer patients. These machine learning algorithms are trying to learn a (latent) representation of the input. The problem is that this representation is hard to interpret and to observe the actual connections between the input and the output. That is why these algorithms are considered to behave like a black-box. In this research, a benchmark is conducted to measure how sensitive these algorithms are to changes in their hyper-parameters. The focus of this experiment are different types of variational auto encoders. We will measure how sensitive they are to changes in their: latent space dimension, learning rate and type of optimizer. The models will be trained on a dataset that contains the RNA gene-expression of different types of cancer tissues. To conclude that the optimizer may play the most important role performance wise for VAE models. Using the optimizer Adam and RMSprop results overall in lower reconstruction loss and overall in a more consistent performance.

Using RNA sequence data for predicting patient properties is fairly common by now. In this paper, Variational Auto-Encoders (VAEs) are used to assist in this process. VAEs are a type of neural network seeking to encode data into a smaller dimension called latent space. These latent features are then used to do downstream task analysis such as cancer types, survival time and cancer stages, with the help of a MLP classifier. Furthermore, the training process itself is also analyzed with the usage of UMaps. The purpose of this paper is to compare different VAE models on their effectiveness in providing training data used for the predictions. The predictions mostly consist of guessing when using any of the latent spaces, constructed by the VAE models, as input data for the MLP classifier. The NoVAE model is the only model with slightly better performance when it comes to mean accuracy and standard deviation. The guessing issue is further analyzed with the help of UMaps. The VAEs are able to classify the input data during the training process, but when faced with new data, this end up not being the case. Both the learning rate and β term yield interesting results regarding the modification of the input data and variational property respectively. A lower learning rate leads to better classification, but this is due it deviation less from the original input data. When using a small β term with the β-VAE, the output is similar to that of the VanillaVAE. Meaning the VanillaVAE does not perform better than a regular autoencoder.

Variational Auto-Encoders are a class of machine learning models that have been used in varying context, such as cancer research. Earlier research has shown that initialization plays a crucial part in training these models, since it can increase performance. Therefore, this paper studies the effect initialization methods on VAEs. This research shows that if using only one hidden layer, Uniform methods and Xavier methods perform best depending on the VAE model, where the standard VAE shows the most sensitivity to these methods. But, if using more hidden layers, the uniform method performs significantly worse than a method that uses the number of inputs of the layer such as the default implementation of PyTorch, Xavier Normal or XavierUniform. However, after enough epochs in all other models these initialization methods converge.

This study presents a comparison of different VariationalAutoencoder(VAE) models to see which VAE models arebetter at finding disentangled representations. Specificallytheir ability to encode biological processes into distinct la-tent dimensions. The biological processes that will be lookedat are the cell cycle and differentiation state. The cell cycleis expressed as a S- and G2M-Score and the differentiationstate is expressed as a number that quantifies the develop-ment time of the cells. First the models will be trained, afterthat the models will be evaluated. The evaluation is doneby checking the latent dimensions for a correlation with thetwo aforementioned biological processes. From this it be-came quite clear that VAE and DIP-VAE performed theworst out of the four models tested. On the other handβ-VAE andβ-TCVAE performed by far the best.

Personalized treatment methods for a complex disease such as cancer benefit from using multiple data modalities from a patient's cancer cells. Multiple modalities allow for analysis of dependencies between complex biological processes and downstream tasks, such as drug response and/or expected survival rate. To this end, it is important to gain an understanding of the relationships between modalities in tumor cells. Multimodal Variational Auto-Encoders (MVAEs) are a combination of generative models trained on different sets of data modalities. In this research, the ability of MVAEs to capture common information between different data views from the same tumor cells is assessed. MVAE models discussed here are a Mixture-of-Experts (MoE) and a Product-of-Experts (PoE) approach to combining the generative model posterior distributions into a single common latent space. The performance assessment is done by: i) comparing the loss of information when reconstructing the training data to MOFA+, a linear method for combining multimodal data, and ii) measuring if one modality of a tumor cell can generate another modality, based on characteristics of the latent space learned by the MVAE. Biological data modalities considered are RNA-seq, gene-level copy number and DNA methylation (DNAme), gathered by The Cancer Genome Atlas. It is found that PoE reconstructs data from all data types with a higher accuracy compared to MoE and MOFA+. The mean squared error of PoE's average reconstruction loss is about a quarter of MOFA+'s, and less than a seventh of the MoE's average reconstruction loss. In terms of predicting modalities from other modalities, the PoE again outperforms MoE on all cross-modal predictions. Additionally, it can be concluded that both models have higher losses in their prediction of DNAme from other modalities, indicating a lesser correlation between this data type and the others.