The complexity of biological systems makes medication development difficult. Diseases include complex gene, protein, and environmental interactions, making treatment targets difficult to identify.
The Board on Health Sciences Policy, Institute of Medicine, Washington, held a workshop on “Improving and Accelerating Therapeutic Development for Nervous System Disorders” and found that this complexity makes it difficult to predict how a drug will behave in the body, resulting in high clinical trial failure rates.
Thus, new medicine market success is minimal. At least 7.9% of clinical trials fail due to insufficient efficacy and unexpected side effects, according to a Springer Nature Journal and National Library of Medicine review. This high attrition rate emphasizes the need to understand disease biology to optimize medication development.
Two key takeaways from their discourse are examined here:
Artificial intelligence in biology relies on organizing complex, inconsistent data into a structured, high-veracity dataset to provide accurate insights and novel discoveries.
Balance crucial decisions and routine jobs: AI should improve early-stage judgments (hypothesis generation, risk assessment) and routine tasks (experimental design) to improve accuracy and processes.
Listen to the complete episode:
Cleaning Data Is Key To AI Success
Liran begins by highlighting two major AI science opportunities:
1. Improving Efficiency: AI can improve productivity by 20-50% by doing ideation, due diligence, and planning, allowing scientists to work quicker and better.
2. Enhancing Scientific Discovery: AI can improve science beyond processes. Deeper AI integrations beyond deterministic, manual operations may improve scientific research and outcomes step-by-step.
He outlines three major AI difficulties in biological research:
1. Understanding Entities and linkages: AI must classify genes, proteins, and diseases and their semantic linkages. Classification is key to understanding scientific data.
2. Interpreting Scientific Meaning: After AI finds entities and correlations, scientists must interpret them into usable insights.
3. Handling Inconsistent Scientific Language: Biology has a “dysfunctional dictionary,” with 20-30 labels for the same topic, according to Liran. AI struggles to interpret and link information due to this inconsistency.
“We aim to understand all scientific discoveries and establish a foundational truth. Data has exploded in recent decades, but it needs structure and veracity to be useful.
A clean, structured, and high-veracity dataset allows AI, such as generative AI, to generate new ideas and connections. However, today’s data landscape is too cluttered and complex for algorithms to produce meaningful results.” – BenchSci co-founder and CEO Liran Belenzon
Liran discusses the two ways to use an organized, accurate knowledge graph in drug discovery:
1. Targeted Disease Insight using Bioinformatics:
Several tech-bio or AI-driven enterprises develop knowledge graphs for one or two diseases. They analyze graphs to find treatments for certain illnesses. A specific biological problem is solved by this specialized technique.
The Productized AI Assistant Approach (Broad Scientific Support)
In contrast, Liran’s team builds a comprehensive disease biology map based on evidence, explainability, and scientific accuracy. Instead of treating ailments, they want to construct a scientist’s “brain behind the brain” AI assistant. This aide improves scientific workflows and decision-making for every drug discovery program.
For the latter approach, BenchSci studies and improves scientists’ key procedures, notably in major pharmaceutical organizations with thousands of researchers:
Key biological decision points must be discovered and optimized in a major pharmaceutical firm with thousands of scientists. Understanding their workflows and boosting productivity by 30–50% is crucial. BenchSci has two main disease biology research areas.
The first is ideation and due diligence—determining mechanisms of action, drug due diligence, risk assessment, and biomarkers. The second is improving experiment planning and execution to speed up science and decrease costs.
AI improves ideation, due diligence, and experiment execution, helping scientists validate or dispute discoveries. – BenchSci co-founder and CEO Liran Belenzon
Balance Important And Routine Tasks
Liran then discusses a crucial use case frequency-impact trade-off in adding AI to scientific workflows that pharmaceutical CEOs must consider in adoption strategy:
1. Low-Frequency, High-Impact Use Cases: Only 20-30% of scientists do ideation, hypothesis development, and due diligence early in a project. These tasks are rare, yet their accuracy determines the project’s path.
2. High-Frequency, Lower-Impact Use Cases: Experimental design and validation are done more often across the company but have less consequences, such as saving days or months in experiment optimization. These duties impact more scientists.
The ideal AI system, according to Liran, should provide deep insights for early judgments and broad support for normal experimental activity. A dual strategy embeds AI as a continual assistant for any scientist rather of confining it to certain projects or specialized users, increasing acceptance.
He highlights that organizational strategy influences how AI is implemented, whether as a specialized tool for certain tasks or a universal AI copilot for every scientist.
Liran’s trade-off emphasizes BenchSci’s focus on scientific intelligence by employing multimodal AI to extract, arrange, and harmonize data from public research and proprietary pharmaceutical company information. Their system decodes and organizes decades of unstructured internal data like electronic lab notebooks and SharePoint files, surpassing standard extraction approaches.
They construct a customized disease biology map using internal and public knowledge, giving pharmaceutical companies a competitive edge. By improving disease mechanism comprehension, their integrated method improves medication discovery and decision-making.
Liran admits that the ultimate goal is to create a “Jarvis of science”—an Iron Man-style AI aide. In business words, he means a powerful AI assistant that provides fresh scientific insights and streamlines research operations, helping scientists produce breakthrough ideas and improve R&D efficiency.