Cancer Genomics / Computational Biology
In silico generation of synthetic cancer genomes using generative AI
Researchers developed OncoGAN, an AI pipeline that generates unlimited, privacy-preserving synthetic cancer genomes indistinguishable from real patient data, offering a powerful new resource for benchmarking and improving cancer genome analysis tools.
Imagine you have a big puzzle, but you can't see all the pieces because they're hidden for privacy reasons. This makes it hard to solve the puzzle. Scientists have found a way to create new puzzle pieces that look just like the hidden ones, so they can share them with others to help solve the puzzle faster. This means they can understand cancer better and find new ways to treat it.
- 01
OncoGAN is a multimodel ensemble pipeline combining GANs, TVAEs, and random sampling that generates realistic synthetic cancer genomes for eight distinct tumor types, accurately reproducing somatic mutations, copy number alterations, and structural variants.
- 02
Synthetic genomes produced by OncoGAN faithfully replicate tumor-specific mutational signatures, genomic positional mutation patterns, and driver mutation frequencies and intercorrelations observed in real PCAWG data.
- 03
On average, only 0.021% of simulated mutations exactly match those in the training set, demonstrating effective donor privacy protection, making the synthetic genomes fully open access.
- 04
DeepTumour achieved nearly 100% tumor-type prediction accuracy on OncoGAN-generated synthetic donors, validating the biological fidelity of the simulated genomes.
- 05
Augmenting DeepTumour training data with 100 OncoGAN-simulated donors per tumor type improved overall classification accuracy from 89.26% to 90.16%, with the largest gains for underrepresented tumor types such as Lymph-MCLL (F1 score: 75% to 84%).
Understanding how genomic alterations drive cancer is key to advancing precision oncology. To detect these alterations, accurate algorithms are used; however, due to privacy concerns, few deeply sequenced cancer genomes can be shared, limiting benchmarking and representing a major obstacle to the improvement of analytic tools. To address this, we developed OncoGAN, a generative AI model combining adversarial networks and variational autoencoders to create realistic synthetic cancer genomes. Trained on large-scale genomic datasets, OncoGAN accurately reproduces somatic mutations, copy number alterations, and structural variants across cancer types while preserving donors' privacy. The synthetic genomes reflect tumor-specific mutational signatures and positional mutation patterns. Using DeepTumour, we validated the synthetic data's fidelity, showing high concordance between generated and predicted tumors. Moreover, augmenting the training data with synthetic genomes improved DeepTumour's accuracy, underscoring OncoGAN's potential to generate shareable datasets with known ground truths for benchmarking and enhancement of cancer genome analysis tools.