Big Mechanism was a research program funded by the U.S. Defense Advanced Research Projects Agency (DARPA) with a budget of $45 million. Launched in 2014, the program aimed to develop software that could read cancer research literature, integrate the extracted knowledge into a comprehensive cancer model, and generate new hypotheses by the end of 2017. The initiative sought to automate the collection of big data and bridge disciplines such as knowledge-based natural language processing, curation and ontology, and systems and mathematical biology. By reading research abstracts and papers, the software was expected to extract pieces of causal mechanisms, thereby accelerating the pace of scientific discovery in oncology.
The program focused on mutations in the Ras gene family, which are implicated in approximately one-third of human cancers. At the time, researchers had a rough road map of interaction sequences among proteins affecting cell replication and death, but the causal relations were poorly understood. Big Mechanism aimed to clarify these relationships by building detailed, computable models from the scientific literature.
Program Structure
The program was organized into three stages. The first stage involved reading the literature and converting it into formal representations. The second stage focused on integrating the knowledge into computational models. The third stage was to produce experimentally testable explanations and predictions. Research teams developed four separate systems that collectively targeted all three tasks.
In February 2015, an evaluation meeting reviewed progress on the first stage. Multiple tasks were considered, including extraction of experimental procedure details and evaluation of statements such as "we demonstrate" and "we suggest." Another task involved mapping sentence meaning and relationships. The best machine-reading system extracted 40% of relevant information from a small corpus and correctly determined how each passage related to the model.
The second stage was scheduled to become active in summer 2015, when team members attempted to produce a single reference model. The third stage was considered the most challenging, because the artificial intelligence community had limited success in developing hypothesis generators. However, molecular biology was seen as more amenable, because most domain knowledge is technical and available in written form.
Technical Approach
Big Mechanism relied on advances in machine learning and natural language processing (NLP) to parse scientific texts. The program used knowledge-based NLP techniques to extract causal relationships, which were then formalized using ontologies and curated databases. The extracted mechanisms were integrated into mathematical models of cellular signaling pathways, particularly those involving the Ras family of proteins.
The use of deep learning and neural networks was not explicitly mentioned in the program's early documentation, but the underlying principles of automated knowledge extraction and model building align with broader trends in generative AI and large language models that emerged later. The program's emphasis on reading and understanding scientific text anticipated later developments in AI-driven scientific discovery.
Challenges and Outcomes
One of the primary challenges was the complexity of biological systems. The causal mechanisms underlying cancer are intricate, involving multiple proteins and feedback loops. The program's goal of generating novel hypotheses was particularly ambitious, as it required the AI to not only synthesize existing knowledge but also propose new experiments.
The evaluation in February 2015 highlighted the difficulty of extracting accurate information from scientific papers. The best system achieved only 40% accuracy, indicating that significant improvements were needed. The program's timeline was tight, with the final goal set for 2017.
While the program concluded in 2017, its legacy persists in the form of tools and methodologies for automated literature analysis and model building. The concepts pioneered by Big Mechanism have influenced subsequent efforts in AI-driven biomedical research, including the use of transformers and multi-head attention in reading and summarizing scientific texts.
Impact and Legacy
Big Mechanism was one of the first large-scale initiatives to apply AI to the entire scientific discovery pipeline, from reading papers to generating hypotheses. It highlighted the potential of AI to accelerate research in complex domains like cancer biology. The program also fostered collaboration between computer scientists, biologists, and mathematicians, leading to cross-disciplinary insights.
Although the program's specific goals were not fully achieved, it contributed to the development of techniques for knowledge graph construction and sequence-to-sequence modeling that are now common in AI research. The focus on causal reasoning and hypothesis generation remains an active area of study, with modern large language models beginning to tackle similar challenges.
See Also
- Artificial intelligence
- Machine learning
- Deep learning
- Neural network
- Large language model
- Transformer (architecture)
- Generative AI
- Knowledge graph
- Sequence-to-Sequence (Seq2Seq)
- Multi-Head Attention
References
- DARPA Big Mechanism program documentation (2014-2017)
- Evaluation meeting reports (February 2015)
External Links
- DARPA official website (archived)