Helping Fred Hutch cut medical research time by 90% while maintaining reliable, customizable, and human-guided research

Helping Fred Hutch cut medical research time by 90% while maintaining reliable, customizable, and human-guided research

Customer: Fred Hutchinson Cancer Center

Dr. Soheil Meshinchi, a pioneer in pediatric oncology, is dedicated to advancing therapies for high-risk pediatric AML leukemia. As a result of decades of collecting patient samples and modern sequence analysis of those samples, the Meshinchi lab at Fred Hutchinson Cancer Center has identified several protein targets of interest that are differentially expressed in the cancerous cells of pediatric AML patients.

A very promising line of therapy is to identify existing drugs that may be effective against those protein targets, and repurpose those drugs to treat pediatric AML patients. The best opportunity to find such drugs is from the thousands of ongoing medical trials aimed at other more common forms of cancer.

The key initial step in the research process is to analyze ongoing trials to identify the protein targets of those trials and the drugs used, and explore related publications to collect information about those drugs. This is a highly demanding task requiring the research team to meticulously read hundreds of clinical trial documents, and then extract and verify information from them.

The Meshinchi lab at Fred Hutchinson Cancer Center collaborated with the Thunk.AI platform to fully automate this crucial initial stage of the research pipeline, placing the clinical trial analysis in the hands of AI agents.

In the span of just 6 months and with no additional staff, the Meshinchi lab automated its clinical trials analysis workflow using AI agents implemented on the Thunk.AI platform.

“Thunk.AI transformed a painstaking time-consuming task into an efficient and simple workflow.”

– Rhonda Ries,

Genomics Research Manager

– Rhonda Ries,

Genomics Research Manager

Now that the workflow has been automated on Thunk.AI, work that might have taken months to perform is completed in a matter of hours, verified in a matter of days, and the research team is freed up to work immediately on the downstream stages of research.

<$1

<$1

cost per trial analyzed

Cost per trial analyzed

90%+

90%+

reduction in processing time for clinical trials analysis

Reduction in processing time for clinical trials analysis

1:1

1:1

Accuracy of results comparable to human workflow

Accuracy of results comparable to human workflow

The original human-powered workflow

The original human-powered workflow

The original human-powered workflow

There were various challenges with the “human-powered” approach to this research process.

Demanding skilled time

Demanding skilled time

Analyzing each clinical trial required highly skilled professionals to dedicate substantial time and effort. On average, it took about 20 minutes per trial to thoroughly read and understand the documents, extract relevant entities like gene targets, interventions, study details, and sponsors, and verify all information. The process was labor-intensive and repetitive, with tasks such as performing web searches for sponsors and checking if interventions were FDA-regulated drugs. For antibody drug conjugates (ADCs), additional effort was required to identify the payload linked to the antibody by researching relevant online publications. These detailed steps made scaling the process difficult, especially as trial volumes continued to grow.

Scaling human effort

Scaling human effort

With hundreds of ADC and CAR-T-related clinical trials and thousands more expected, the volume of work rapidly outpaced the capacity of the existing team. Manual workflows, while thorough, could not keep up with the increasing demands. This bottleneck risked delays in extracting critical insights needed for downstream decision-making, impacting the ability to stay ahead in research and innovation.

Maintaining accuracy

Maintaining accuracy

Despite the high workloads, there was no room for error in clinical trial analysis. Precision was critical for ensuring the integrity of research findings and compliance with regulatory standards. Exhaustive data collection and validation were essential, yet the risk of human fatigue and errors increased with the growing workload. Ensuring high accuracy and consistency under these challenging circumstances was a constant struggle for the team.

After attempting to execute this workflow solely through manual effort, it became clear that a smarter, faster solution was essential to maintain the pace of innovation.

The automated AI-powered workflow

The automated AI-powered workflow

The Meshinchi lab partnered with Thunk.AI to describe the workflow as an intelligent application called a ‘thunk’.

•  They created a thunk that can download hundreds of clinical trials from
clinicaltrials.gov and process each of them with a simple workflow.
• They used English to define each step of the workflow and the instruction logic
to be used at every step.
• They tested the results and tuned the instruction logic appropriately to meet a
high quality bar.

All of this was done rapidly, taking advantage of the self-service capabilities and no-code design of the Thunk.AI platform.

The Meshinchi lab partnered with Thunk.AI to describe the workflow as an intelligent application called a ‘thunk’.


• They created a thunk that can

download hundreds of clinical trials

from clinicaltrials.gov and process each

of them with a simple workflow.

• They used English to define each step

of the workflow and the instruction

logic to be used at every step.

• They tested the results and tuned the

instruction logic appropriately to meet a

high quality bar.


All of this was done rapidly, taking advantage of the self-service capabilities and no-code design of the Thunk.AI platform.

For example, the instructions for the last workflow step (“Research drug and manufacturer”) were specified using straightforward English instructions along with an example of the desired result format:

Search the web or sponsor website to find out if this is an FDA regulated drug. Record urls from where you captured this information.


Example of how I want you to format output:


Lirilumab is not FDA regulated.

Evidence: https://www.sciencedirect.com/topics/immunology-and-microbiology/lirilumab,https://www.sciencedirect.com/topics/medicine-an

Search the web or sponsor website to find out if this is an FDA regulated drug. Record urls from where you captured this information.


Example of how I want you to format output:


Lirilumab is not FDA regulated.

Evidence: https://www.sciencedirect.com/topics/immunology-and-microbiology/lirilumab,https://www.sciencedirect.com/topics/medicine-an

That’s it! Notice how the instructions are exactly how they would be specified to a human teammate who would be expected to exercise judgment and subjectivity in making decisions. That is sufficient for the AI agent to do its work.

Efficient Research & Information Extraction:

Efficient Research & Information Extraction:

The thunk AI agent automated the extraction of all key information from clinical trial APIs and specialized websites like ADCReview, processing single trials or batches of trials as needed. The relevant data was organized into structured columns, ready for review.

The thunk AI agent automated the extraction of all key information from clinical trial APIs and specialized websites like ADCReview, processing single trials or batches of trials as needed. The relevant data was organized into structured columns, ready for review.

Human-guided AI Collaboration:

Human-guided AI Collaboration:

Fred Hutch staff tailored their thunk by providing clear, natural language instructions on what information to extract and how the AI should retrieve it. These instructions, along with examples, allowed the AI to adapt to specific requirements. After the AI completed its tasks, human reviewers checked the outputs, fine-tuning the AI's behavior with additional instructions when necessary. The platform also enabled compliance checks to ensure results met all specified requirements.

Fred Hutch staff tailored their thunk by providing clear, natural language instructions on what information to extract and how the AI should retrieve it. These instructions, along with examples, allowed the AI to adapt to specific requirements. After the AI completed its tasks, human reviewers checked the outputs, fine-tuning the AI's behavior with additional instructions when necessary. The platform also enabled compliance checks to ensure results met all specified requirements.

Auditable Work:

Auditable Work:

Every action taken by the AI or the human within the thunk was logged for complete transparency and auditability. This traceability allowed users to see detailed records of the AI's activities, such as which web searches were performed and the sources of the extracted answers. This ensured a high level of accountability and trust in the workflow.

Outcomes and Impact of AI agent automation

Outcomes and Impact of AI agent automation

In the span of just a month in production, the Meshinchi lab has analyzed more than 1000 clinical trials using AI agent automation – an undertaking that would otherwise have taken many months of focused effort. AI automation has provided tangible and as well as intangible benefits for the AML research process at the Meshinchi lab:

90% Time Reduction:

90% Time Reduction:

By automating the repetitive and time-consuming aspects of clinical trial analysis, AI reduced the absolute time required to process each trial from 20 minutes to less than 2 minutes. This dramatic 90% time savings allowed the research team to analyze more trials in less time, significantly improving operational efficiency and accelerating research timelines.

Fred Hutch staff tailored their thunk by providing clear, natural language instructions on what information to extract and how the AI should retrieve it. These instructions, along with examples, allowed the AI to adapt to specific requirements. After the AI completed its tasks, human reviewers checked the outputs, fine-tuning the AI's behavior with additional instructions when necessary. The platform also enabled compliance checks to ensure results met all specified requirements.

Cost Efficiency:

Cost Efficiency:

The automated workflow also brought substantial cost savings. Each trial was processed for under $1, a fraction of the cost of traditional manual methods. These savings not only reduced overall operational expenses but also enabled the research team to allocate resources more effectively to other high-priority research areas.

Fred Hutch staff tailored their thunk by providing clear, natural language instructions on what information to extract and how the AI should retrieve it. These instructions, along with examples, allowed the AI to adapt to specific requirements. After the AI completed its tasks, human reviewers checked the outputs, fine-tuning the AI's behavior with additional instructions when necessary. The platform also enabled compliance checks to ensure results met all specified requirements.

Empowered Teams:

Empowered Teams:

The AI automation freed researchers from routine, manual tasks, allowing them to focus on complex, high-value analyses that require their expertise. This shift not only improved job satisfaction but also empowered the team to contribute more meaningfully to advancing cancer research and discovering critical insights.

Fred Hutch staff tailored their thunk by providing clear, natural language instructions on what information to extract and how the AI should retrieve it. These instructions, along with examples, allowed the AI to adapt to specific requirements. After the AI completed its tasks, human reviewers checked the outputs, fine-tuning the AI's behavior with additional instructions when necessary. The platform also enabled compliance checks to ensure results met all specified requirements.

Enhanced Accuracy:

Enhanced Accuracy:

The AI automation started out with an impressive 90% accuracy in its first implementation. Over time, this accuracy improved further as researchers provided feedback and fine-tuned instructions, ensuring that the AI consistently met Fred Hutch's exacting standards. The combination of AI precision and human oversight resulted in error-free, reliable outputs, essential for such a critical domain.

Fred Hutch staff tailored their thunk by providing clear, natural language instructions on what information to extract and how the AI should retrieve it. These instructions, along with examples, allowed the AI to adapt to specific requirements. After the AI completed its tasks, human reviewers checked the outputs, fine-tuning the AI's behavior with additional instructions when necessary. The platform also enabled compliance checks to ensure results met all specified requirements.

“What would have been an error prone, manual process, became rapid, reliable, and reproducible.”

– Rhonda Ries,

Genomics Research Manager

– Rhonda Ries,

Genomics Research Manager

Book a demo

Talk with our CEO and founders about how AI can transform your business.

Book a demo

Talk with our CEO and founders about how AI can transform your business.

Book a demo

Talk with our CEO and founders about how AI can transform your business.