Today’s guest post is from Dr. Alexander Kühn, Head of Bioinformatics at Alacris Theranostics, a Berlin-based spin-off company of the Max Planck Institute for Molecular Genetics, which uses next generation sequencing and other genomics data through its predictive modeling system ModCell™ for drug development and personalized medicine in oncology.
Cancer is a complex disease; differences in the genetic make-up of individuals as well as their tumors make every cancer patient unique. Yet, the majority of current medical practice fails to recognize this individuality and treats many patients identically, leading to wide variations in response to therapy. Typically, only 25% of patients benefit from the (often expensive) treatment they are given, with many suffering serious side effects. Even if some progress is visible, cancer drug therapy still seems to be largely based on a trial and error principle.
Computer models have enormous potential to overcome this mismatch for patients as well as for healthcare costs. We at Alacris use computer models based on millions of data points to carry out virtual clinical trials and virtual patient modeling. In order to better match patients to therapies and therapies to patients, Alacris developed the ModCell™ system.
ModCell™ generates a ‘Virtual Patient’ model for individualized prediction of therapy outcome. First, an individual patient and his or her tumor is analysed on the molecular level. This patient-specific genetic information is subsequently integrated into a cancer model. Thus, the ModCell™ system combines all available molecular information about a patient’s disease with the sum of the molecular and mechanistic knowledge about cancer as a whole within a patient-specific tumor model. Using this model, we can simulate the effects of hundreds of different drugs or drug combinations on the patient tumor right on a computer.
We used the ModCell™ system to simulate the effects of about 100 molecular targeted anti-cancer as well as non anti-cancer drugs/compounds on more than 700 different cancer cell lines, originating from 24 diverse tissue types. Molecular data of cancer cell lines were provided by the Cancer Cell Line Encyclopedia (CCLE). For each different cell line available, virtual models comprised of more than 6,000 parameters were generated. The models were employed to predict cell-line specific responses to a range of drugs by simulating a concentration of varying amounts for each compound (detailed description of the simulation approach can be found here). In total, we generated more than 5 million different models, each of which was expected to need up to one minute to be solved numerically and to produce about 500 kB of simulation data. This results in a simulation time of more than 3,000 days, if run on a single core, and more than 2 TB of data. Google Cloud Platform enabled us to handle this enormous simulation workload:
Here’s a breakdown of the steps we took to set up our workload on Google Compute Engine:
- Set up a single n1-highmen-8 instance as an SSH gateway
- Connected and exported a persistent 1 TB disk with Network File System version 4 (NFSv4)
- Created 125 VM instances, with a total of 1,000 cores, and connected these to the exported disk
- Used open source project TaskManager to control and schedule the generated cluster
- TaskManager enabled us to execute the job on a single core by distributing each of the 5 million simulation jobs over the whole cluster
- TaskManager enabled us to operate at full capacity and in parallel over several days. Results of each simulation were stored on the shared storage disk.
Simulation results are now being used to calculate the inhibition of cell growth, which will be compared to pharmacological profiles available from Cancer Cell Line Encyclopedia (CCLE). Potentially, these results will facilitate identification of cancer-related signal transduction pathways that are not yet covered in the model, and also further our understanding of the functional consequences of mutations at both the molecular pathway and cellular level. Moreover, we will gain insight into drug action at the molecular level and analyze cross-talk and potential redundancies between pathways. Using this information, the ModCell™ system can be expanded and modified, integrating the identified cancer-related signaling pathways and mutations, and additional drug information, such as drug uptake dynamics and drug metabolism.
Use of Google Cloud Platform has made it possible for us to rapidly refine and improve our modeling system, which has become about 10x faster than using the Alacris’ computer cluster (which contains only 100 cores). These improvements will help (a) to optimize personalized cancer therapy by removing some of the risks associated with classical empirically administered treatments and (b) to optimize the predictive power of virtual clinical trials to test drugs before they go into clinical trials. The goal is for this to lead to higher drug approval rates, saving time and a large fraction of the costs associated with clinical trials.
-Posted by Dr. Alexander Kühn, Head of Bioinformatics at Alacris Theranostics
Feed Source: Google Cloud Platform Blog
Article Source: Alacris helps better match cancer patients with drug therapies using Google Cloud Platform