AI Lags in Global Drug Asset Discovery by 27%

Scientists are tackling the increasing challenge of identifying promising new drug candidates emerging from a rapidly globalising biopharmaceutical landscape. Alisa Vinogradova and Vlad Vinogradov of Bioptic.io, in collaboration with Luba Greenwood from Harvard Business School and colleagues, present a novel approach to drug asset scouting, addressing the fact that over 85% of patent filings now originate outside the United States, with a significant proportion from China. This research is significant because current AI agents struggle to comprehensively and accurately identify these ‘under-the-radar’ assets, creating substantial financial risk for investors and hindering business development. The team, including Ilya Yasny, Dmitry Kobyzev, Shoman Kasbekar, Kong Nguyen, Dmitrii Radkevich, Roman Doronin, and Andrey Doronichev from Bioptic.io, developed and benchmarked a ‘Bioptic Agent’, a tree-based self-research AI, demonstrating a substantial performance improvement over leading models like Claude Opus 4.6, Gemini 3 Pro + Research, GPT-5.2 Pro, Perplexity Research, and Exa Websets in identifying relevant assets from diverse, multilingual sources.

Scientists are developing artificial intelligence to locate promising new drugs originating outside the United States. A growing proportion of pharmaceutical innovation now happens globally, with much of the information published in non-English sources. This technology aims to give investors and developers a crucial edge in a rapidly changing landscape.

Scientists have developed a new artificial intelligence agent capable of significantly improving the identification of promising drug candidates from across the globe. This innovation addresses a critical need in biopharmaceutical research, where a growing proportion of novel assets originate outside the United States and are often reported through regional, non-English language channels.

The research team constructed a challenging benchmark to assess the performance of these AI agents, focusing on completeness and accuracy in identifying relevant drug assets, a task where even small omission rates can translate to multi-billion-dollar losses for investors and pharmaceutical companies. The newly developed Bioptic Agent, a tree-based self-learning system, demonstrably outperforms existing deep research AI tools in this complex scouting process.

By leveraging a multilingual, multi-agent pipeline, the researchers created a rigorous testing ground that prioritizes identifying assets largely overlooked by U.S.-centric searches. This approach reflects the reality of modern drug development, where China alone accounts for approximately 30% of global activity and over 1,200 novel candidates are currently in development.

The team collected complex search queries directly from experienced investors, business development professionals, and venture capitalists to ensure the benchmark accurately reflects real-world scouting scenarios. Furthermore, the agent’s performance continued to improve with increased computational resources, suggesting that further investment in processing power could yield even more substantial gains in drug asset scouting capabilities.

The work presents a new benchmarking methodology specifically designed for the drug asset scouting problem, and the Bioptic Agent represents a significant step towards automating and enhancing this critical process within the biopharmaceutical industry. This technology promises to accelerate the identification of promising new therapies and reduce the risks associated with overlooking valuable assets hidden within the increasingly global landscape of pharmaceutical innovation.

Bioptic Agent significantly outperforms rival models in complex drug asset identification

Bioptic Agent attained a 79.7% F1 score in a drug asset scouting benchmark, demonstrating a substantial lead over competing AI agents. This performance represents the proportion of correctly identified assets relative to all retrieved assets, and signifies a marked improvement in recall and precision. Evaluations revealed a strong correlation between computational resources and performance, suggesting that increased compute power directly translates to more effective asset discovery.

Light deduplication, employed as the default setting, achieved deduplication quality comparable to a more intensive “heavy” mode, while maintaining efficiency. The heavy deduplication mode, involving exhaustive web searches for aliases and cross-lingual variants, was available for scenarios demanding the highest confidence in accuracy, but was not necessary to achieve state-of-the-art results.

The core of Bioptic Agent’s success lies in its tree-based, self-Bioptic approach, guided by a “Coach Agent”. This agent dynamically proposes refinements to search directives, focusing on under-explored angles and addressing errors identified by the “Investigator Agent”. Node rewards, calculated as the product of local precision and the change in validated assets, prioritize directives that reliably add new, valid assets to the global results, rather than those generating high volumes of low-quality candidates. Backpropagation of these rewards through the directive tree ensures that promising search paths receive increased attention in subsequent epochs.

Regional News Mining and Discoverability Assessment for Drug Asset Benchmarking

A multilingual, multi-agent pipeline underpins the construction of a challenging completeness benchmark designed to evaluate drug asset scouting capabilities. Initially, a Regional News Miner Agent systematically surfaces potential drug assets from non-English language sources, addressing the increasing trend of biopharmaceutical innovation originating outside the United States and being initially disclosed through regional channels.

Following asset identification, an Attributes Enrichment Agent validates and structures the extracted information, ensuring data consistency and facilitating subsequent analysis. To prioritise assets with limited English-language visibility, a Google Search Agent assesses discoverability, specifically comparing the presence of information in both English and the asset’s origin language.

This pipeline extends beyond simple asset identification to encompass realistic query generation. Real-world screening queries were collected directly from experienced investors, Business Development professionals, and Venture Capitalists, providing a foundation of practical search criteria. These queries underwent a process of clustering by intent and distillation into reusable templates by a Template Generator Agent.

Conditioned on these templates, a Query Generation Agent produced a diverse set of benchmark queries, each paired with the corresponding ground-truth asset identified earlier. Rigorous validation is central to the methodology. A Query Validator Agent, alongside expert human reviewers, ensured the quality and investor-realism of each query-asset pair, mitigating the risk of spurious matches or irrelevant results. This multi-layered approach, combining automated agents with human oversight, aims to create a robust and representative benchmark that accurately reflects the complexities of real-world drug asset scouting, where completeness and the avoidance of hallucinations are paramount.

Evaluating large language models for global pharmaceutical innovation scouting

The relentless global expansion of biopharmaceutical research demands a new approach to identifying promising drug candidates. For years, the assumption that innovation largely flowed from US and European laboratories has obscured a critical reality: the majority of new assets now emerge from elsewhere, often disclosed in regional languages and channels easily missed by conventional scouting methods.

This isn’t merely a matter of geographical awareness; it’s a fundamental challenge to the entire process of drug discovery, with billions of dollars at stake for those who can swiftly and comprehensively map the landscape. This work addresses that challenge head-on, not by offering another search engine, but by benchmarking the performance of existing large language models against a deliberately complex task.

The impressive gains achieved by the ‘Bioptic Agent’ , significantly outperforming leading commercial models, suggest that a focused, tree-based approach, coupled with sufficient computational power, can meaningfully improve the recall of relevant assets. Crucially, the methodology prioritises completeness and avoids the ‘hallucinations’ that plague many AI systems, a vital consideration when dealing with sensitive investment decisions.

However, the limitations are clear. The benchmark, while sophisticated, remains focused on a specific area, chronic hepatitis B, and relies on expert-generated queries. Scaling this approach to encompass the entirety of pharmaceutical research will require substantial effort and ongoing refinement of the evaluation metrics. Furthermore, the reliance on compute power raises questions about accessibility for smaller players.

Looking ahead, the real potential lies not just in optimising AI agents, but in integrating them seamlessly into the workflows of human experts. The future of asset scouting will likely involve a hybrid model, where AI handles the initial, exhaustive search, and human analysts provide the critical judgment and contextual understanding needed to translate data into actionable insights. This is a step towards a more inclusive and efficient global innovation ecosystem, but it demands continued investment in both technology and human expertise.