Cancer-inspired Genomics Mapper Model for the Generation of Synthetic DNA Sequences with Desired Genomics Signatures

Genome data are crucial in modern medicine, offering significant potential for diagnosis and treatment. Thanks to technological advancements, many millions of healthy and diseased genomes have already been sequenced; however, obtaining the most suitable data for a specific study, and specifically for validation studies, remains challenging with respect to scale and access. Therefore, in silico genomics sequence generators have been proposed as a possible solution. However, the current generators produce inferior data using mostly shallow (stochastic) connections, detected with limited computational complexity in the training data. This means they do not take the appropriate biological relations and constraints, that originally caused the observed connections, into consideration. To address this issue, we propose cancer-inspired genomics mapper model (CGMM), that combines genetic algorithm (GA) and deep learning (DL) methods to tackle this challenge. CGMM mimics processes that generate genetic variations and mutations to transform readily available control genomes into genomes with the desired phenotypes. We demonstrate that CGMM can generate synthetic genomes of selected phenotypes such as ancestry and cancer that are indistinguishable from real genomes of such phenotypes, based on unsupervised clustering. Our results show that CGMM outperforms four current state-of-the-art genomics generators on two different tasks, suggesting that CGMM will be suitable for a wide range of purposes in genomic medicine, especially for much-needed validation studies.

Knowledge-integrated AutoEncoder Model

Data encoding is a common and central operation in most data analysis tasks. The performance of other models, downstream in the computational process, highly depends on the quality of data encoding. One of the most powerful ways to encode data is using the neural network AutoEncoder (AE) architecture. However, the developers of AE are not able to easily influence the produced embedding space, as it is usually treated as a \textit{black box} technique, which makes it uncontrollable and not necessarily has desired properties for downstream tasks. In this paper, we introduce a novel approach for developing AE models that can integrate external knowledge sources into the learning process, possibly leading to more accurate results. The proposed \methodNamefull{} (\methodName{}) model is able to leverage domain-specific information to make sure the desired distance and neighborhood properties between samples are preservative in the embedding space. The proposed model is evaluated on three large-scale datasets from three different scientific fields and is compared to nine existing encoding models. The results demonstrate that the \methodName{} model effectively captures the underlying structures and relationships between the input data and external knowledge, meaning it generates a more useful representation. This leads to outperforming the rest of the models in terms of reconstruction accuracy.