AI in Service of Scientific Discovery: Two New Approaches

The landscape of scientific research is being enriched by two new artificial intelligence tools, described in recent publications in Nature. These systems, developed by Google and the non-profit organization FutureHouse respectively, have been conceived to assist scientists in the process of formulating and verifying hypotheses, with an initial focus on drug retargeting. The primary objective is not to replace human ingenuity, but rather to address the growing complexity and massive volume of information that characterize modern science.

Google has introduced its Co-Scientist, a system that adopts an approach defined as "scientist in the loop." This means that researchers maintain a central role, constantly applying their judgment to guide and direct the system's operations. Although Google indicates its applicability to physics as well, initial demonstrations have focused exclusively on biological data, particularly on straightforward hypotheses such as the effectiveness of a drug for a specific condition.

"Agentic" Architectures and Data Management

FutureHouse's system, on the other hand, goes a step further, training an AI capable of autonomously evaluating biological data derived from specific classes of experiments. Despite the differences between the two, both share a fundamental characteristic: they are "agentic." This implies that they operate in the background, calling upon and orchestrating external tools to perform their functions. A similar approach has also been adopted by Microsoft for its science assistant, while OpenAI stands out for having simply performed fine-tuning of an LLM for the biology domain.

The raison d'être of these systems lies in AI's ability to process quantities of information that far exceed human capabilities. In sensitive sectors such as biology and pharmacology, where scientific literature and experimental datasets are growing exponentially, tools like Co-Scientist and FutureHouse's system become crucial for identifying patterns, correlations, and new hypotheses that would otherwise remain hidden.

Implications for Infrastructure and Data Sovereignty

The introduction of AI assistants capable of "chewing through" massive volumes of data raises significant questions for technological infrastructure. Organizations intending to adopt such solutions must carefully consider computing and storage requirements. Processing complex biological datasets and running advanced AI models demand considerable computational resources, often translating into a high need for VRAM and processing power for inference.

For sensitive sectors like pharmaceutical research, data sovereignty and regulatory compliance (e.g., GDPR) are critical aspects. The choice between a cloud deployment and an on-premise infrastructure becomes strategic. A self-hosted or air-gapped approach can offer greater control over sensitive data and ensure compliance, but it entails a higher TCO and the need for internal management of hardware and software. AI-RADAR offers analytical frameworks on /llm-onpremise to evaluate the trade-offs between these options, considering factors such as latency, throughput, and operational costs.

Future Prospects and the Role of AI in Science

These developments mark an important step in the integration of artificial intelligence into the scientific process. Far from seeking to replace scientists, these assistants are configured as powerful co-pilots, capable of amplifying research capabilities and accelerating discovery. Their effectiveness in drug retargeting, a high-impact area with traditionally long development times, demonstrates AI's transformative potential.

However, it is essential to maintain a critical perspective. AI excels at data processing and correlation identification, but the formulation of truly innovative hypotheses and a deep understanding of scientific phenomena remain the prerogative of human intellect. The future of research will likely see an increasingly close symbiosis between human intuition and AI's computational power, with systems like Co-Scientist and FutureHouse's acting as catalysts for new discoveries.