Stavroula Ntoufa- Director of Scientific Affairs, Causaly
Early-stage life sciences research is a complex and resource-intensive process. Scientists weigh an overwhelming amount of information from scientific literature, including journals, Google Scholar, and PubMed. Researchers must search through vast data sources to accurately identify and contextualize all relevant biomedical information.
This process takes significant time and money while introducing human bias, a known factor affecting research quality and accuracy. Most R&D leaders acknowledge bias influences pipeline prioritization, often overlooking contradictory safety signals.
Despite using external experts for mitigation, understanding which efforts are effective remains unclear. It’s reasonable to think this is a major contributing factor to 90% of drug development projects ultimately failing to make it all the way through approvals.
Human biases – from sampling errors, favoring desirable results, or emphasis on positive outcomes – create research blind spots. Though typically unintentional, these biases impact the validity of scientific conclusions and delay or prevent research progress. To minimize bias, research must:
- Comprehensively include all available data
- Enhance accuracy beyond current standards
- Improve time and cost efficiencies
How AI enhances research efficiency
Researchers typically rely on a variety of external and internal data sources. Teams spend months conducting keyword searches, scanning titles, authors, and abstracts from thousands of documents. This manual, arduous process requires patience and multitasking, making it highly susceptible to bias and potentially yielding incomplete results.
Ideally, researchers want to minimize time discovering and evaluating information, freeing themselves to focus on analyzing and interpreting materials relevant to their research. AI can now effectively process vast amounts of biomedical information, reducing research time, costs, and inaccuracies.
Machine learning models process data from multiple sources, contextualize information, and deliver results to scientists in plain speaking, conversational interfaces instead of a simple search bar. Large Language Models (LLMs) assess study context, identify relevant connections between seemingly unrelated research, and generate accurate, comprehensive, and properly cited results.
Purpose-built AI systems, fine-tuned for scientific research workflows, perform this entire process continuously. Connected to all key data sources, they gather ongoing updates and new data to accurately represent the latest findings.
Beyond enhancing efficiency, AI exposes the challenges of human bias in research. Addressing bias using AI requires a deeper understanding of the issues inherent in scientific exploration.
Human Bias in Research
Three major bias categories – sampling, familiarity, and positivity – create problematic research outcomes. While scientists try to minimize biases, AI can mitigate even more potential errors.
Countering Sampling Bias: Incorporating All Data
Scientists strive to conduct research representing the full spectrum of scientific data available. Researchers are mindful to avoid drawing conclusions based on limited information that mislead or invalidate studies.
Given the volume of available information and sources, researchers unintentionally overlook relevant studies because of human limitations – a challenge scientists understand.
Research teams combat sampling bias by cross-referencing findings and reviewing multiple sources with similar keywords. While these methods improve information completeness, they remain time-consuming and imperfect. The fragmentation and high cost of data access further complicate this process, requiring large teams to effectively extract insights.
AI addresses sampling bias by effectively managing large-scale datasets. Instead of reviewing 10-15 papers a day, AI can distill and contextualize thousands of papers in seconds. By adopting AI into their process, researchers gain confidence that underrepresented findings and hidden connections will be found. Information is integrated and weighted across the board, ensuring research conclusions are based on more comprehensive information and answer their research questions with greater accuracy and transparency.
Mitigating Familiarity Bias: Shining a Light on Blind Spots
Researchers spend years reading, testing, and forming beliefs in their fields. While this expertise is critical, it can lead to familiarity bias, closely related to confirmation bias.
Scientific studies undergo constant reexamination and updates. When researchers unintentionally overlook valid but contradictory evidence, research outcomes could become questionable, potentially creating a domino effect that influences other researchers to favor similar results.
Researchers minimize familiarity bias through peer review, collaboration with multiple external experts, and adopting policies that recognize diverse perspectives. While effective, these approaches require more time and effort.
AI accelerates this process and reduces familiarity bias by objectively analyzing scientific literature, internal documents, and data. It examines the full spectrum of data, incorporating recent updates, contradictory findings, and seemingly unrelated information that scientists might overlook or even consider.
Including all available data balances data synthesis. By including novel or dissenting perspectives, AI expands researchers’ perspectives beyond their conscious or unconscious biases.
Countering Positivity Bias: Uncovering Negative Results
Positivity bias, also known as publication bias, occurs when studies with positive or statistically significant results are favored for publication over studies with negative or null findings. This distorts scientific findings, and while researchers recognize this issue, it persists.
Researchers understand that negative outcomes provide equally important insights as positive results. However, publications inadvertently make it difficult for scientists to capture and analyze relevant findings from negative results, as these studies receive fewer citations. These create research blind spots.
Researchers go to great lengths to counter this, spending significant time reviewing preprint studies, unpublished data, and clinical trial registries. This additional effort not only requires more effort to gather accurate, comprehensive data, but also risks wasting time on potential non-viable targets.
AI analyzes and references information from all sources and quickly surfaces hard-to-find results. Using cutting-edge LLMs, AI can quantify bias in published studies by analyzing citation patterns, identifying contradictory results, and highlighting otherwise overlooked negative results.
This presents a complete picture of available information from multiple sources and provides a balanced representation of all previous study outcomes. Most importantly, research teams can de-risk clinical trials through better decision-making on the targets and drugs before clinical trials begin, resulting in significant cost savings and ultimately better drugs for the patients.
AI’s Potential in Mitigating Bias
Research bias isn’t theoretical or philosophical. These are issues with real-world consequences that can shape the future of life sciences.
While researchers apply rigorous methodologies to minimize bias, these efforts demand additional time and investment.
AI can play a pivotal role in reducing research bias by:
- Analyzing a vast collection of information across multiple sources to create a broader representation of findings
- Identifying underrepresented findings
- Highlighting unexpected correlations and surface contradictions
- Structuring results to balance sources and deter dominant narratives
As AI evolves, the opportunity extends beyond making research faster and more cost-efficient to improving the quality of results by including a more complete spectrum of information.
Though researchers take steps to minimize the impact of bias, traditional approaches remain time-intensive and imperfect. Systematic approaches to bias mitigation beyond relying on external experts and teams are needed.
While AI is increasingly transforming the pharma industry, scientists still face limitations. AI providers recognize these challenges and are actively addressing them by designing systems that enhance transparency, verifiability, and reproducibility - critical elements for addressing various trust issues:
- The black box problem: Many powerful models provide predictions without clear explanations, a challenge for scientists who need mechanistic hypotheses or justifications for experimental follow-up.
- Data quality limitations: Scientific datasets often come from diverse sources that may be incomplete, biased, or noisy. Furthermore, much of the highest-quality data remains confidential within organizations or protected by intellectual property rights.
- Complex scientific interpretation: Biological systems operate dynamically, nonlinearly, and through multi-scale processes, which are challenging to model with current AI capabilities. Though AI excels at identifying correlations, inferring causality - a cornerstone of scientific discovery - demands expert human interpretation.
By complementing human expertise, AI can reshape early-stage research in life sciences, reducing bias, improving accuracy, and unlocking discoveries that will shape life sciences for years to come.
Subscribe to our e-Newsletters.
Stay up to date with the latest news, articles, and events. Plus, get special
offers from American Pharmaceutical Review delivered to your inbox!
Sign up now!