This interview analysis is sponsored by BenchSci and was written, edited, and published in alignment with our Emerj sponsored content guidelines. Learn more about our thought leadership and content creation services on our Emerj Media Services page.
The intricate nature of biological systems significantly complicates drug development for major pharmaceutical enterprises. Diseases often involve complex interactions among genes, proteins, and environmental factors, making it challenging to pinpoint precise therapeutic targets.
Published findings from a workshop on “Improving and Accelerating Therapeutic Development for Nervous System Disorders”, conducted by the Board on Health Sciences Policy, Institute of Medicine, Washington, highlights that the complexity of biological systems leads to difficulties in predicting how a drug will behave in the human body, resulting in high rates of clinical trial failures.
Consequently, the success rate for new drugs reaching the market is low. A review of studies from the Springer Nature Journal and cataloged in the National Library of Medicine indicates that only about 7.9% of clinical trials succeed, with many failures attributed to inadequate efficacy and unforeseen adverse effects. High attrition rates underscore the critical need for a deeper understanding of disease biology to improve drug development outcomes.
Emerj Editorial Director Matthew DeMello recently spoke with Liran Belenzon, Co-founder and CEO of BenchSci, to talk more about how the challenges in drug discovery due to complex disease biology and the low success rate of clinical trials, emphasizing the need for AI to enhance scientific workflows and improve decision-making.
This article examines two critical insights from their conversation:
- Clean data is essential for AI success: AI’s full potential in biology depends on organizing messy, inconsistent data into a structured, high-veracity dataset to drive accurate insights and new discoveries.
- Balance critical decisions and routine tasks: AI should enhance both early-stage decisions (hypothesis generation, risk assessment) and routine tasks (experimental design) to boost accuracy and accelerate workflows.
Listen to the full episode below:
Guest: Liran Belenzon, Co-founder and CEO, BenchSci
Expertise: Entrepreneurship, Artificial Intelligence, Business Strategy
Brief Recognition: Liran served as an army commander in the Israel Defense Forces for three years before founding an e-commerce startup, which became Israel’s leading B2B marketplace by 2015. After moving to Toronto, Canada, he earned an MBA from the University of Toronto’s Rotman School of Management. His entrepreneurial ventures also include Biz Coupon and Standy.
Clean Data is Essential for AI Success
Liran opens the conversation by identifying two significant opportunities with AI in science:
1. Improving Efficiency: AI can enhance productivity by performing tasks like ideation, due diligence, and planning, increasing efficiency by 20-50% and allowing scientists to do the same work faster and better.
2. Advancing Scientific Discovery: Beyond improving current processes, AI can also help create better science itself. Deeper integrations of AI beyond deterministic, manual workflows can potentially drive significant, step-wise improvements in scientific research and outcomes.
He further highlights three key challenges in using AI for biological research:
1. Understanding Entities and Relationships: AI must first classify biological entities (like genes, proteins, and diseases) and identify their semantic relationships. Classification is foundational for extracting meaning from scientific data.
2. Interpreting Scientific Meaning: Once AI identifies entities and relationships, the next challenge is translating those findings into meaningful, actionable insights for scientists.
3. Handling Inconsistent Scientific Language: Biology, Liran says, has an extremely “dysfunctional dictionary,” where the same concept may have 20-30 different names. This inconsistency makes it difficult for AI to interpret and connect information accurately.
“BenchSci’s approach is to comprehensively understand everything ever discovered in a scientific area and establish a foundational truth. The past few decades have seen an explosion of data, but without structure and veracity, that data cannot be effectively leveraged.
Once a clean, structured, and high-veracity dataset is in place, AI’s full potential — such as generative A I— can be harnessed to form novel ideas and new connections. However, today’s data landscape is too messy and complex to feed into an algorithm and expect meaningful outcomes.”
— Liran Belenzon, Co-founder and CEO at BenchSci
Liran explains to the Emerj audience the two different approaches to using a structured, accurate knowledge graph in drug discovery:
1. The Bioinformatics Approach (Targeted Insight for Specific Diseases):
Some companies, often called tech-bio or AI-driven companies, build knowledge graphs focused on one or two diseases. They dive deep into these graphs to extract insights aimed at curing specific conditions. This approach is narrow, focusing on solving a particular biological problem.
2. The Productized AI Assistant Approach (Broad Scientific Support):
Liran’s team follows a different approach — building a comprehensive map of disease biology that emphasizes evidence, explainability, and scientific accuracy. Instead of targeting specific diseases, they aim to create an AI assistant that is the “brain behind the brain” for scientists. This assistant supports every drug discovery program by enhancing decision-making and scientific workflows.
To succeed in the latter approach, the BenchSci team focuses on understanding and improving the core workflows of scientists, especially in large pharmaceutical companies with thousands of researchers:
“In a large pharmaceutical company with thousands of scientists, key decision points in biology must be identified and optimized. Understanding their workflows and increasing productivity by 30 to 50% is critical. Our focus at BenchSci falls into two primary areas centered on unraveling disease biology.
The first is ideation and due diligence — determining mechanisms of action, conducting drug due diligence, assessing risk, and identifying biomarkers. The second is enabling more efficient experiment planning and execution to accelerate scientific progress and reduce costs.
By enhancing these workflows, AI helps scientists streamline ideation, due diligence, and experiment execution, ensuring they can validate or challenge insights more effectively.”
— Liran Belenzon, Co-founder and CEO at BenchSci
Balance Critical Decisions and Routine Tasks
Liran then explains an essential trade-off between use case frequency and impact in applying AI to scientific workflows that is instrumental for pharmaceutical leaders to understand in adoption strategy:
1. Low-Frequency, High-Impact Use Cases: These involve ideation, hypothesis generation, and due diligence, which are typically performed by only 20-30% of scientists at the early stages of a scientific project. While these tasks happen less often, their impact is significant because getting them right or wrong shapes the project’s direction.
2. High-Frequency, Lower-Impact Use Cases: Tasks like experimental design and validation occur much more frequently across the organization but have more minor, incremental impacts — such as saving a few days or months in optimizing experiments. These tasks affect a broader portion of the scientific workforce.
In drawing his distinction between the two, Liran suggests that the ideal AI system should address both categories: providing deep insights for critical early decisions and broad support for routine experimental work. A dual approach increases AI adoption by embedding it as a constant assistant for every scientist rather than limiting it to specific projects or specialized users.
He also emphasizes that organizational strategy determines how AI is integrated — whether as a niche tool for specific tasks or a universal AI copilot supporting every scientist.
Ultimately, Liran’s emphasis on the trade-off he describes underscores that BenchSci’s approach focuses on scientific intelligence by using multimodal AI to extract, structure, and harmonize data from public research and proprietary information within pharmaceutical companies. Their deployment of AI capabilities decodes and organizes decades of unstructured, internal data — such as electronic lab notebooks and SharePoint files — overcoming the limitations of traditional extraction methods.
By combining internal and public knowledge, they create a proprietary map of disease biology, offering pharmaceutical companies a unique competitive edge. Their integrated approach enhances decision-making and drug discovery by providing a more comprehensive understanding of disease mechanisms.
Liran is unabashed that the ultimate goal here is to build a “Jarvis of science” — a pop culture reference to the AI assistant character from the Iron Man movies. In business terms, he means an advanced AI assistant that delivers new scientific insights and streamlines research workflows, enabling scientists to generate breakthrough ideas while increasing efficiency across the R&D process.
