AI is moving beyond generating insights in drug discovery to actively coordinating entire research workflows, but experts warn that building safety guardrails and human oversight into these systems is just as important as the computational power behind them. As agentic AI systems take on more autonomous decision-making in molecular research, the field is grappling with how to ensure these tools remain accountable, transparent, and aligned with clinical safety standards.

What Are Agentic AI Systems in Drug Discovery?

Unlike traditional AI tools that simply respond to a single prompt and return an answer, agentic AI systems function more like a coordinated research team. They understand research objectives, plan sequences of tasks, retrieve relevant scientific context, coordinate multiple specialized tools, evaluate results, and recommend next steps. In molecular discovery, this means an agentic system can automatically route a drug candidate through protein structure prediction, binding affinity scoring, toxicity profiling, and lead optimization without researchers manually moving data between different platforms.

The challenge researchers face is that drug discovery involves fragmented workflows. A scientist investigating a potential drug target might need to consult multiple databases, retrieve scientific papers, run molecular analyses, compare outputs from different AI models, and move protein structure predictions from one platform into a docking tool on another. "What's missing is a system that coordinates them," according to research from NYB.AI, a Singapore-based company developing agentic AI infrastructure for molecular discovery.

How Are Researchers Building Safety Into AI Drug Discovery Systems?

Matthew Stewart, a postdoctoral researcher who transitioned from environmental science to AI safety, exemplifies how the field is approaching this challenge. After completing his Ph.D. at Harvard's School of Engineering and Applied Sciences, Stewart now leads AI research at Pelago Health, a digital substance use care provider, where he focuses on building AI agents with robust safety mechanisms.

• Human-in-the-Loop Monitoring: Stewart's team is developing systems where AI conversations with users are reviewed by healthcare professionals for warning signs of mental health crises, rather than allowing the AI to make autonomous clinical decisions.
• Clinical Oversight and Escalation: The system flags clinicians when concerning patterns emerge in user conversations, enabling human professionals to reach out via phone, send resources, or refer users to licensed therapists for direct care.
• Suicidality Assessment Research: Stewart's team has a peer-reviewed publication under review examining how to conduct suicidality monitoring using human-in-the-loop AI, demonstrating the field's commitment to rigorous validation before deployment.

"We're basically exploring this bridging factor of whether you can have an AI that allows people to get the clinical care they need, but with more safety guardrails and clinical oversight than a typical AI agent," Stewart explained.

Why Does Model Selection Matter in Molecular AI?

A critical insight emerging from AI research in drug discovery is that choosing the right model for the right task is as important as having access to powerful models. A peer-reviewed paper published in Briefings in Bioinformatics by NYB.AI researchers examined this "which model to use, and when" problem in graph-based drug-target interaction modeling. Different AI models operate at fundamentally different levels of analysis, each answering different scientific questions.

• Network-Level Models: These capture broad associations between drugs and targets across entire biological networks, useful for early-stage target discovery.
• Sequence-Level Models: These represent proteins as strings of amino acids, enabling researchers to predict how protein sequences affect drug binding without requiring full 3D structural data.
• Structural Models: These require full 3D structural data and model interactions at the resolution of individual atoms and binding pockets, providing precise mechanistic insights but demanding more computational resources.
• Mechanism-Focused Models: These specifically target interaction mechanisms, helping researchers understand not just whether a drug binds to a target, but how and why.

Using the wrong model wastes both time and computing resources. The research proposes a practical framework: before selecting any model, clarify what decision you actually need to make. Is your task association discovery, interaction classification, affinity estimation, candidate ranking, pocket identification, pose assessment, or mechanistic hypothesis generation? Only then should researchers examine model inputs, choose the right resolution, scrutinize how candidate models were evaluated, and plan experimental validation.

How Can Organizations Implement Agentic AI Responsibly?

For research-driven organizations, the strategic opportunity lies in orchestration rather than automation. Stewart's earlier work on algorithmic accountability provides a foundation for this approach. He published a paper titled "Beyond Explanation: Evidentiary Rights for Algorithmic Accountability," which analyzed 168 litigated cases involving algorithmic decisions to show that access to evidence was strongly associated with successful contestation. This research was one of three papers selected out of 325 entrants for the Best Paper Award at the Association for Computing Machinery Conference on Fairness, Accountability, and Transparency.

The core insight is that when an AI system makes a decision affecting people, those affected should have the right to understand and challenge that decision. In loan applications, for example, applicants currently have limited recourse if rejected. Under a counterfactual accountability framework, applicants could ask whether factors such as gender or age would change the outcome, and the institution would have to provide verifiable answers.

Organizations implementing agentic AI in drug discovery can apply similar principles by ensuring that:

• Decision Transparency: Researchers can trace which models were used, in what sequence, and why the system recommended a particular candidate for further testing.
• Iterative Refinement: Agentic systems continuously refine predictions based on real-world validation data, so each research cycle gets smarter and more aligned with actual experimental outcomes.
• Expert Integration: Rather than replacing researchers, agentic systems free scientists to focus on high-value decision-making, creativity, and asking novel questions that no automated system could anticipate.

What Are the Broader Implications for AI in Science?

Despite advances in AI models and computing power, much of the infrastructure required to apply AI effectively in scientific research remains out of reach for many organizations. Access to frontier AI models, large-scale computing resources, molecular simulation platforms, scientific retrieval systems, and workflow orchestration tools often requires substantial investment and technical expertise. This creates a growing divide between organizations that can afford to assemble sophisticated AI-driven discovery environments and those that cannot.

Agentic AI offers a potential path to democratize drug discovery by bridging the gap between generating an insight and acting on it. A lead that might have taken months to optimize with scattered tools and manual hand-offs can now be systematically enhanced in weeks. For executives across biotechnology, pharmaceuticals, healthcare, and scientific research, the organizations most likely to benefit from the next wave of AI adoption may not be those deploying the latest or best-performing models, but those who can most effectively integrate models, data, expertise, and workflows into a coherent research process.

The next phase of AI adoption in science is fundamentally about leveraging agentic systems to integrate fit-for-purpose AI models and the supporting technology ecosystem around them, making advanced research capabilities available to more skilled and passionate research teams worldwide. For research-driven industries, such increased access may prove to be the most important breakthrough of all.