LLM-Based Agents for Competitive Landscape Mapping in Drug Asset Due Diligence
This paper introduces and benchmarks an LLM-based agent system for drug asset due diligence that transforms unstructured VC memos into a structured evaluation corpus, achieving 83% recall in competitor discovery and reducing analyst turnaround time from 2.5 days to approximately 3 hours.
Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer
Imagine you are a detective trying to solve a massive case: "Who are all the other players in this specific drug market?"
In the world of biotech investing, before a fund decides to back a new medicine, they need to know the entire "competitive landscape." They need to find every single other drug that fights the same disease. This is called Due Diligence.
Currently, doing this is like trying to find a needle in a haystack, where the haystack is made of thousands of different languages, broken websites, blurry images, and secret reports. Human experts have to spend days (about 2.5 days per drug) manually digging through these messes to find the answers.
This paper introduces a new AI Detective Team (called BIOPTIC AGENT) that does this job much faster and better. Here is how it works, broken down into simple concepts:
1. The Problem: The "Lost and Found" of Drug Names
The authors explain that finding competitors is incredibly hard because:
- Names are messy: One drug might be called "Drug X," "The Blue Pill," or by its chemical code in different countries.
- Information is scattered: Clues are hidden in scientific papers, government trial lists, press releases, and PDFs that look like spreadsheets.
- The rules change: New drugs appear every day, and old ones disappear.
Existing AI tools (like standard search engines or simple chatbots) often miss the "long-tail" competitors—the obscure ones that are hard to find. They also tend to "hallucinate" (make things up) when they can't find a real answer.
2. The Solution: A Team of AI Agents
Instead of asking one AI to "go find everything," the authors built a hierarchical team of specialized AI agents. Think of it like a well-organized newsroom or a detective squad:
- The Parser (The Archivist): This agent reads the messy, unstructured reports (PDFs with charts and text) and turns them into neat, organized lists. It extracts the drug names and the diseases they treat.
- The Explorer (The Detective): This is the main agent. It doesn't just guess; it goes out onto the "web" (the internet) to hunt for competitors. It uses a strategy called REACT (Reasoning, Acting, Observing).
- Analogy: Imagine a detective who doesn't just ask one question. They ask a question, read the answer, realize they need more info, ask a new question based on that, and repeat this loop many times. They also send out multiple "scouts" to search different websites at the same time.
- The Validator (The Judge): This is the most critical safety step. The Explorer might find 50 drugs, but 10 might be fake or irrelevant. The Validator acts like a strict judge. It checks every single drug the Explorer found against the rules: "Is there real proof this drug fights this specific disease?" If the answer is no, it throws the drug out. This ensures the final list is highly accurate.
3. The Training Data: Learning from the Pros
To teach this AI team, the authors didn't just use public data. They used 5 years of private, real-world investment memos from a biotech venture capital fund.
- They took these messy, handwritten-style reports and used their own AI to clean them up into a "Gold Standard" dataset.
- This dataset became the exam to test if their new AI team was better than the competition.
4. The Results: Speed and Accuracy
The authors tested their BIOPTIC AGENT against other famous AI systems (like OpenAI's Deep Research and Perplexity).
- The Score: The BIOPTIC AGENT found 83% of the known competitors (Recall).
- OpenAI Deep Research found 65%.
- Perplexity found 60%.
- The Speed: In a real-world test with a venture capital fund, the time it took an analyst to finish a competitive analysis dropped from 2.5 days to about 3 hours. That is a 20x speedup.
- The Quality: Even though it was faster, it didn't sacrifice accuracy. The "Judge" agent successfully filtered out false alarms, ensuring the final list was trustworthy.
5. Why It Works: The "Bitter Lesson" vs. The "Scaffold"
The paper discusses a debate in AI: Should we just use a super-smart brain (a big model) and hope it figures it out? Or should we build a structure (a "scaffold") that forces the AI to work step-by-step?
The authors found that for this specific, messy job, the scaffold wins.
- A single-pass AI (one that tries to answer in one go) fails when the information is hard to find.
- The BIOPTIC AGENT succeeds because it is built to keep searching, check its own work, and verify evidence before giving an answer. It's not just "thinking"; it's "hunting."
Summary
This paper presents a deployed system that acts like a super-efficient research team. It takes the chaotic job of mapping drug competitors, uses a multi-step AI process to hunt down every possible clue, and then uses a strict "Judge" AI to verify the findings. The result is a tool that is 20 times faster than human experts and finds more competitors than other top AI systems, all while keeping the error rate low.
Drowning in papers in your field?
Get daily digests of the most novel papers matching your research keywords — with technical summaries, in your language.