Ovarian Cancer Case Study
Daimensions™ took two minutes to do what would normally have taken many months
Finding the genes that predict cancer takes too long and is too expensive
The data set we worked with contained 21,000 gene protein expressions (that’s 21,000 columns) derived from cells drawn from 584 patients (584 rows). Each cell sample was labeled by experts as belonging to one of three categories: the patient has ovarian cancer, or two other categories that indicated “no cancer”.
Using existing best practices, searching through all of the possible combinations of factors (columns) that might interact to predict the cancer/no cancer outcome would take many months. Finding the answer through computing alone would not be achievable in any reasonable amount of time. (This is especially true if we include the possibility of multiple gene expressions contributing to cancer – which is important!)
Until now, many human experts from many disciplines needed to collaborate in order to apply their knowledge and experience to cut the computational problem down to a manageable size. The existing approach – heavily dependent on expert time – is slow and error-prone.
Brainome’s measurement-driven approach finds the answer
The goal is to quickly find, using computing alone, a solid, dependable solution with very high accuracy. In Brainome’s vocabulary, we wanted to do the cancer prediction with “near perfect accuracy” and “extreme generalization”.
This means that we wanted to find a solution that could be clearly demonstrated to be both correct and to apply broadly to all women with potential ovarian cancer – not just to the 584 patients included in the study.
We applied Brainome’s Daimensions™ system in measurement-driven mode. What happened was remarkable.
A generalized model with 100% accuracy in under 2 minutes
In less than 2 minutes we found a single gene, VWA7, out of 21,000, that predicts ovarian cancer with 100% accuracy.
This result was created by Brainome Daimensions™ on a single-GPU computer with 1 TB of SSD and 32 GB of RAM.
The resulting predictor model was tiny – just two neurons.
Our results, calculated in isolation from any knowledge of the domain and with no help from experts, are fully consistent with the genomics research community’s current best understanding of the causes of ovarian cancer.
Build an Audio detector using 10 neurons in less than 1 day
Ovarian Cancer Study
Find the key genes expression out of 21,000 for ovarian cancer diagnosis