AI-Driven Biology Is Only as Good as the Data Beneath It
Opinion | The biotech and pharma industries are pouring billions of dollars into AI-driven drug discovery and development. But the genomic data used to train AI models are often incomplete, contain errors, or lack contextual information — a quality issue that threatens to slow the pace of biological research..
Artificial Intelligence (AI) is increasingly used in biology to analyze vast amounts of data, predict outcomes, and accelerate discoveries. AI systems rely on machine learning (ML), a type of AI where algorithms learn patterns from data without explicit programming. In biology, ML models are trained on datasets containing genetic sequences, protein structures, or clinical records to identify correlations or make predictions. For example, AI can help predict how a patient might respond to a drug based on their genetic makeup. However, the effectiveness of AI in biology hinges entirely on the quality and completeness of the underlying data. If the data is biased, incomplete, or unrepresentative, the AI’s predictions may be unreliable or even harmful. The article highlights that AI-driven biology is only as robust as the data it is built upon, emphasizing the critical role of data curation in scientific research.
A significant challenge in AI-driven biology is the existence of data gaps, which refer to missing or underrepresented information in datasets. These gaps can arise due to historical biases in data collection, such as focusing primarily on individuals of European descent in genetic studies, which accounts for over 80% of genetic data in public databases despite representing only about 16% of the global population. Such biases can lead to AI models that perform poorly for underrepresented groups, creating disparities in healthcare outcomes. The article notes that these gaps are not limited to genetics but also affect other areas like protein structure data, where certain protein families are overrepresented while others are scarcely documented. Addressing these gaps requires intentional efforts to collect diverse and comprehensive data, which is often resource-intensive and time-consuming.
When AI models are trained on biased or incomplete data, the consequences can be severe. For instance, AI tools used in drug discovery may fail to identify effective treatments for diseases prevalent in underrepresented populations because the training data lacks sufficient examples from those groups. The article cites a 2023 study showing that AI models trained on biased genetic data were less accurate in predicting disease risks for non-European populations by up to 30%. Similarly, in protein research, AI tools may struggle to predict the structures of proteins from understudied organisms, limiting their utility in fields like agriculture or environmental science. These inaccuracies can lead to wasted resources, misguided research directions, and even ethical concerns about fairness in healthcare and scientific applications.
Institutions such as universities, research organizations, and funding agencies play a crucial role in addressing data gaps in AI-driven biology. The article highlights efforts by institutions like the National Institutes of Health (NIH) in the United States, which has launched initiatives to fund research aimed at diversifying genetic datasets. For example, the NIH’s All of Us Research Program aims to collect health data from over 1 million participants, with at least 75% of the participants coming from groups historically underrepresented in biomedical research. Additionally, organizations like the Global Alliance for Genomics and Health (GA4GH) work to establish standards for data sharing and collaboration across borders. These initiatives are essential for creating more inclusive and representative datasets, but they require sustained funding and political support to succeed.
Technological advancements offer potential solutions to mitigate data gaps in AI-driven biology. One approach is the use of synthetic data, which involves generating artificial datasets that mimic real-world data to fill gaps where real data is lacking. For example, AI models can be trained on synthetic protein structures to improve predictions for understudied proteins. Another solution is the development of federated learning, a technique where AI models are trained across multiple decentralized datasets without sharing the raw data itself. This approach allows researchers to collaborate globally while preserving data privacy, which is particularly important for sensitive health data. The article notes that while these technologies show promise, they are not yet widely adopted and require further validation to ensure their reliability in real-world applications.
Beyond technical hurdles, ethical and policy challenges complicate efforts to improve data quality in AI-driven biology. The article discusses concerns about *data sovereignty*, where countries or communities may resist sharing their genetic or health data due to fears of exploitation or loss of control. For instance, Indigenous groups have raised concerns about the commercialization of their genetic data without their consent or benefit-sharing agreements. Additionally, the lack of standardized regulations for data sharing across countries creates legal uncertainties, particularly when data crosses international borders. Policymakers must balance the need for open data to advance science with the rights of individuals and communities to control their own biological information. The article emphasizes that addressing these challenges requires not only technological solutions but also robust ethical frameworks and international cooperation.
