ActFound: AI Model Transforms Drug Development with Precision

Scientists from China and the United States have developed a cutting-edge artificial intelligence (AI) model, named ActFound, that promises to address significant challenges in drug development and discovery. According to a study published in Nature Machine Intelligence, ActFound outperforms existing models in bioactivity prediction, potentially offering a more cost-effective solution compared to traditional methods.

Overcoming Challenges in Bioactivity Prediction

Bioactivity refers to the various properties of compounds, including how they interact with biological targets, their effects on biological systems, and their therapeutic potential. Researchers from Peking University, the University of Washington, and AI tech firm INF Technology Shanghai collaborated on ActFound, which aims to bypass the typical hurdles faced by machine learning in this field. These challenges include limited data labeling and incompatibility between assays, which are tests used to measure the activity or potency of drugs.

Superior Performance and Cost-Effectiveness

ActFound not only outshines competing AI models but also rivals the accuracy of free-energy perturbation (FEP), a traditional computational method used in drug development. While FEP is known for its precision, it requires substantial computational resources, making it impractical for large-scale applications. Additionally, FEP relies on hard-to-obtain, three-dimensional protein structures, which necessitate expensive equipment and extensive laboratory procedures.

In contrast, ActFound delivers high accuracy with fewer data points, offering a more accessible and affordable alternative to FEP. “Our promising results indicate that ActFound could be an effective bioactivity foundation model for various types of activities,” said Wang Sheng, the corresponding author and assistant professor at the University of Washington.

The Role of AI in China's Pharmaceutical Industry

China's pharmaceutical industry is rapidly expanding, with significant government investment in research and development for innovative drugs. Many companies are now leveraging AI to identify potential drug targets, aiming to reduce development times. Some AI-developed products are already progressing through clinical trials, highlighting the growing impact of AI in the sector.

The Importance of Bioactivity Prediction in Drug Discovery

Bioactivity prediction is crucial in drug discovery and development, as it helps scientists identify promising compounds from a vast pool of candidates, reducing the need for time-consuming and costly experiments. Despite the potential of machine learning, its adoption has been limited due to several challenges, such as the poor generalizability of models in bioactivity prediction.

Foundation Models and Innovative Learning Techniques

To address these challenges, the team behind ActFound utilized foundation models, which are pre-trained on large datasets to improve predictions for unlabeled data. ActFound was trained using 35,644 assays from a well-known chemical database and 1.6 million experimentally measured bioactivities. The model employs two innovative machine learning techniques: meta-learning and pairwise learning.

Meta-learning allows the model to be optimized using a limited amount of labeled data, making it ideal for predicting properties of unmeasured compounds. This technique is particularly valuable in drug discovery, where obtaining bioactivity data can be prohibitively expensive.

Pairwise learning enhances the model's generalizability by calculating relative differences between compound pairs rather than predicting absolute values, which may vary across different assays. This approach mitigates issues with incompatible measurement metrics, enabling more accurate predictions.

ActFound's Success in Real-World Applications

ActFound was rigorously tested on six real-world bioactivity datasets and demonstrated superior performance compared to nine competing models in both in-domain and cross-domain predictions. In a case study focusing on cancer drugs, ActFound again outperformed other models, confirming its potential as a powerful tool for bioactivity prediction.

Paving the Way for AI-Driven Drug Development

The success of ActFound suggests that it could serve as a foundational model for bioactivity prediction, opening new avenues for machine-learning-based drug development and discovery. “Our promising results indicate that ActFound could be an effective bioactivity foundation model for compound bioactivity prediction, paving the way for machine-learning-based drug development and discovery,” the research team concluded.