Reasoning and decision-making at the molecular level could lead to incredible advances in biotechnology (as well as other sciences). Imagine a future with "intelligent drugs"-drugs that have been developed to travel through the bloodstream with the capability to detect and assess diseases as well as treat them. What if these drugs were capable of being specifically designed for the unique chemistries taking place within your body? This may sound far-fetched, but the theoretical foundations for such materials are taking shape.
The detection and assessment method for intelligent drugs would require developing chemistries with artificial intelligence (AI) capabilities. Some mixture of chemical components would have to be capable of assessing their environmental situation through reactions (or lack thereof) and determining the correct components to modify or release, if any. The framework for this process requires algorithms similar to machine-learning templates, but with biochemical inputs and outputs.
Harvard computer scientists have addressed the algorithm issue by demonstrating that certain useful AI algorithms can be carried out via chemical reactions. By realizing that chemical reaction dynamics are analogous to computational methods used in certain AI algorithms, the team from Harvard's Wyss Institute and SEAS (School of Engineering and Applied Sciences) developed a tool that can translate these algorithms into molecular representations.
The algorithms for this study use a combination of probability and graph theory components known as "message-passing inference on factor graphs". These algorithms are already in use in many everyday devices as predictive and filtering mechanisms, for tasks such as verification processes and the extraction of information from difficult-to-interpret, noisy, or partially corrupted data.
The Harvard tool takes unknowns that are represented in probabilistic terms (otherwise known as probabilistic graphical models), and produces a group of simulated chemical reactions with predictive properties. Entities that can't be directly measured and observed are estimated. This ability to estimate is one of the first steps in developing so-called smart drugs and targeted chemical systems.
This work may have implications in other intersections between the fields of biochemistry and machine learning. Biological systems have a similar task to robots; both must sometimes estimate and evaluate their surroundings with incomplete information, and take actions based on their conclusions. AI and machine learning principles are well suited to these situations. The ability to design an artificial environment that models cellular-level biological systems could give researchers a powerful tool to solve complex problems on a molecular level in real-world biological systems.
Don't run down to your local pharmacy searching for smart drugs anytime soon-this remains a theoretical framework, and has yet to involve direct chemical reactions. Any bench chemist will tell you that while simulations are very useful, real-world chemical reactions can often provide surprises. Combine this with the complex environment of living biological systems (like us) and the challenges are still significant.
Even so, this work provides a notable advancement in modeling capabilities by addressing the issue of bio-systems complexity with proven AI principles. It's reasonable to expect this work will contribute to many biotechnology advances.