The key initial step in the research process is to analyze ongoing clinical trials to identify protein targets, associated drugs, and relevant publications. This requires research teams to meticulously read hundreds of clinical trial documents, extract key details, and verify information—a highly demanding and time-consuming task.

With Thunk.AI, this process can be automated while keeping humans in control. This sample thunk demonstrates how AI can find, extract, verify, and supplement clinical trial data, freeing up researchers to focus on higher-value analysis and decision-making.

This sample thunk researches clinical trials of targeted cancer therapies: antibody-drug conjugates (ADCs), CAR-T cell therapies and T-cell engagers. Each work item is one trial, identified by its NCT number. You can add trials by NCT number, or ask the thunk to find them for you, such as "add ten glofitamab trials in France".

For each trial, the thunk works through five steps:

1️⃣ Fetches the trial record – reads the trial's registry details from ClinicalTrials.gov: its title, phase, status, sponsor, conditions, countries, interventions and age range. A trial that isn't in the registry is marked as not found, and nothing else is researched.
2️⃣ Identifies the therapy and its target – picks out the investigational therapy among the trial's interventions (even when the registry lists it only by a code name), records its modality, and finds the gene or protein it targets, with a link to the evidence.
3️⃣ Finds the ADC payload – for an antibody-drug conjugate only, finds the drug the antibody carries, with a link to the evidence.
4️⃣ Checks FDA approval – looks the therapy up in Drugs@FDA, the FDA's own records. Cell therapies aren't in Drugs@FDA, so for those it looks for the approval on fda.gov. A news story alone never counts as an approval.
5️⃣ Checks development activity – finds the newest publication about the therapy in Europe PMC, and records whether the therapy is still in active development, from that publication's date and the trial's status.

Steps 3 to 5 run only when they apply. A trial that tests no targeted therapy (for example, an imaging study) stops after step 2.

In Thunk.AI, you can build tools that extend an agent's capabilities. This thunk has four API tools that read public, read-only databases: one fetches a trial from ClinicalTrials.gov, one searches it for trials, one checks Drugs@FDA through openFDA, and one searches Europe PMC for publications. Each step has only the tools it needs, and web search is on only for the research steps.

Every finding is recorded in its own column with the type of value it holds: the modality and the FDA status are fixed lists of choices, the target is a list, the ages are numbers, the latest publication date is a date, and each piece of evidence is its own link. The thunk's saved reports use these columns to list, for example, the antibody-drug conjugates, the approved therapies, and the therapies still in development but not approved.

The sample comes with seven test work items, each a real trial whose answer is known. They include a therapy listed only by its code name, an investigational ADC, a CAR-T therapy approved through fda.gov, an MRI study where "ADC" means something else, and an NCT number that doesn't exist.

Here's a short video demonstration of this thunk:

Fred Hutch Cancer Research Center researchers used a thunk much like this one to reliably extract and synthesize information over 1,000s of trials, with AI handling data extraction, organization, and validation.

Instead of spending hours manually processing each trial, they were able to:
✅ Automate repetitive research across diverse trial formats
✅ Maintain accuracy and compliance with structured AI outputs
✅ Scale their research efforts significantly

More specifically, they achieved:

With AI handling the research-heavy work, skilled researchers could now focus on reviewing insights and applying them to decision-making. Explore the attached white paper to deep dive on this case study.

Copy & Customize this thunk: Head over to Samples and copy the thunk to adapt it to your own research. Change the kinds of therapy it looks for, add the details your team needs as new columns, or connect your own databases in place of the public ones, all without coding.

Build your own thunk: You can also design your own AI-driven research workflow tailored to your requirements and processes. Follow our step-by-step tutorial to create your own AI-powered Thunk.

Need help? Contact us at [email protected] if you need any assistance in using the platform.