Vivan Therapeutics conducts personalized drug screens for cancer patients.
During their patented drug screening process (called Personal Discovery Process), a team of scientists models a patient’s tumor mutations in fly avatars, and then uses them to test a library of FDA-approved drugs—to identify drug combinations best suited to treat the patient’s cancer. Additionally, AI-based software scans the patient’s cancer profile against a vast amount of data and research to find the best treatment recommendations.
The Vivan Therapeutics team decided to automate parts of their process by training a machine learning model to count and classify flies, intending to reduce time and staffing requirements and eliminate human error. To achieve that, the scientists needed a training data platform that could host large amounts of data, support the entire model training pipeline from data curation through annotation and model training, and provide AI expertise and dedicated labelers.
Vivan Therapeutics uses V7 to train classification and segmentation models that streamline distinguishing and quantifying flies used for analyzing responses to potential cancer therapies. Thanks to V7, the Vivan Therapeutics team has set up training data workflows that include model-assisted data curation and labeling, as well as semi-automated QA processes.
With the current iterations of their models reaching 96% accuracy on average, the team has saved up to 4.5 days of work per patient and significantly improved the throughput of their lab. The team continues to work with V7 to automate, accelerate, and refine their patented Personal Discovery Process to help more cancer patients receive effective treatments and reduce fatality rates.
"We were looking for third-party software that would cover the whole training data pipeline, from labeling to model training, since our company didn’t possess a full-time machine learning team with expertise in computer vision at the time. V7 provided us with all of that, along with the APIs, which are crucial for further productionisation of the system."
Vivan Therapeutics is a cancer therapeutics platform that provides personalized drug screens to cancer patients.
Personal Discovery Process, a patented drug screening platform for a single patient, analyzes an individual's genetic sequencing data to capture their tumor network. The mutations unearthed by the process are engineered into an army of 500,000 fruit flies. Then, Vivan Therapeutics’ scientists use the army to screen up to 2000 FDA/NICE-approved drug cocktails and discover the most efficient combinations.
To conduct the trial, the scientists put fly embryos— the patient’s avatars alongside control embryos—into screening tubes, where they are fed combinations of drugs mixed with food. The staff monitors which drugs increase the survival rate of the flies under different conditions. All drug 'hits' that increase the survival rate of avatar pupae are then put through the secondary screen.
In addition to personalized cancer therapeutics, the company conducts oncology drug discovery research for biopharma partners, which requires testing the entire library of FDA-approved drugs.
The Vivan Therapeutics team has automated parts of the Personal Discovery Process. They use specialist machines to capture 360-degree images of the screening tubes. The images are then used to build machine-learning models that count and distinguish the flies.
Vivan Therapeutics conducts a separate drug screen for every patient. Each drug screen comprises several thousands of fly pictures—and each fly has to be counted and classified.
The scientists count all the pupae and insert their findings into Excel sheets. However, this fully manual method causes a lot of bottlenecks in the project—due to time, staffing, and storage requirements, as well as the inevitable human error.
The Vivan Therapeutics team has decided to work towards automating parts of the process to overcome these obstacles and shorten the time to delivery. Their imaging system helps capture high-quality images of pupae inside the vials, facilitating the creation of datasets for training computer vision models that count and classify the flies.
Currently, the Vivan Therapeutics team uses the V7 platform and labeling workforces for their entire data annotation and model training pipeline. With the help of V7 experts, the team has trained mini-models, which support the data curation and annotation stages in their workflows.
The data preparation workflow starts with Vivan Therapeutics staff uploading images to V7. Then, the images are run through an image quality classification model, which helps eliminate low-quality images. Afterward, a segmentation model labels the pupae on the remaining images—based on whether or not they have tumors and whether they survived. The model-assisted annotation is followed by an automated logic stage—if the annotations are of insufficient quality, a Vivan Therapeutics team member reviews and corrects them. Otherwise, the pre-annotated images go directly to V7-provided labelers. Finally, the images undergo a final quality assurance by the Vivan Therapeutics scientists.
One key step towards automating the Personal Discovery Process is training a machine learning model to help count and classify flies.
With no dedicated machine learning team on board at the time, the Vivan Therapeutics team searched for a comprehensive training data platform that would let them easily set up a machine learning pipeline and provide them with AI experts and a labeling workforce. After testing several vendors, Vivan Therapeutics decided to go with V7.
The team initially picked V7 because of its convenient pricing model—considering the necessary data volumes, quantity-based prices were out of the question. However, what convinced them to continue using V7 was its powerful model training capabilities. The team discovered they could develop a fully working mini-model with as few as 125 manually annotated images, and then use it to annotate more data and train more accurate models on larger datasets. This meant they could start working on automating their process within just a few days of subscribing to V7.
V7’s machine learning and customer success specialists, as well as the on-demand labelers, turned out to be perfect partners for Vivan Therapeutics. Within the budget and resources dedicated to working with V7, the company couldn’t have assembled the same level of AI expertise and technology in-house.
The Vivan Therapeutics team also benefits from V7’s intuitive interface. With the easy-to-use workflows, the team can employ model and logic stages and organize an airtight QA process. Aided by auto-annotation features, the Vivan Therapeutics team has accelerated their annotation process and boosted the overall quality of their work.
V7’s API and integrations are helping to integrate and automate the Personal Discovery Process, reducing workforce requirements, human error, and the time needed to handle one patient.
"The time it takes to count the flies, storage requirements, and human error are the project’s biggest bottlenecks. If we get rid of them by automating parts of the research currently performed manually, we’ll be able to process more patients and give them more accurate reports much faster. We assume that with the help of V7-trained models, we can save up to 4.5 days of work per patient—and optimize the screenings even more in the future."
Vivan Therapeutics' current manual system allows the team to process around three patients a month. Once the V7-trained models are deployed in the lab, the team will be able to save up to 4.5 person days of work per patient. And the team’s not stopping here.
The segmentation model currently used in the workflows reaches 96% accuracy. It will be iterated on to develop a full-scale machine-learning model that will fully automate the counting and classification of flies. The team also plans to add additional models to the workflow.
With the help of V7, Vivan Therapeutics is working towards automating as many parts of the Personal Discovery Process as possible to improve the speed and accuracy of the screenings. Fewer human errors and less statistically varied results help boost the lab's credibility, opening up new commercial opportunities for the team. And the faster and better the analysis, the more cancer patients are saved.
Discover how other AI-first companies solved knowledge tasks at scale with V7