Drug discovery AI stops predicting single steps: agents now plan, synthesize and test in a loop, graded on a six-level autonomy spectrum

AI in drug discovery is being repositioned: rather than passive programs that complete one prediction or generation task, agents are collaborative systems that break a high-level research goal into sub-tasks, invoke computational or experimental tools, and adjust later actions based on intermediate results. Earlier classical machine learning, deep learning and generative models each excelled at drug-target interaction prediction, molecular property estimation, or de novo design of small molecules and therapeutic proteins, but largely stayed within a single task or local step, leaving the multi-step, cross-disciplinary, feedback-heavy workflow to manual integration. The review follows the Design-Make-Test-Analyze cycle through target discovery, lead optimization, biomacromolecule design, clinical candidate evaluation, retrosynthesis planning and automated experimental platforms, labeling representative systems with the same autonomy grading plus their collaboration architecture and open-source accessibility.

To separate true agents from tool-augmented pipelines, the review proposes a Level 0 to Level 5 autonomy spectrum scored on scientific decision-making, system autonomy and the degree of human involvement required. Level 0 fixed tools and Level 1 assisted pipelines remain pre-agentic, with execution paths set by preset logic or manual processes; the algorithmic divide appears at Level 2, the task controller, where the large language model begins actively scheduling tools and recovering from interface-level errors. Level 3 closed-loop researchers keep adjusting strategy around a goal and close a higher-level reasoning loop, Level 4 self-evolving systems optimize how they work by retaining cross-project experience, and Level 5 virtual scientists would propose new research questions; the review states that fully independent Level 5 systems do not yet exist in drug discovery and that current advances concentrate at Levels 2 to 4.

In several concrete stages, the gain from agent architecture can be separated from base-model scale: performance improvements come mainly from tool scheduling, feedback loops and role division rather than from a stronger underlying model alone. Traditional non-agent workflows require researchers to hand-write scripts, debug reward functions and configure environments, or require computational biologists to handle data preprocessing and script development. In tuberculosis antibiotic discovery, with the same diffusion generative models and ADMET prediction tools, a sequential pipeline's molecules received a 58.0% expert approval rate, rising to 82.6% when the agent framework scheduled the tools and replanned on feedback; in a retrosynthesis safety-constraint task, LARC built on a 12-billion-parameter open-source model reached an 82.1% success rate on carcinogen avoidance, against 3.6% for GPT-4o without agent encapsulation.

The review addresses both computational and experimental researchers, covering target discovery, lead compound design, biomacromolecule design, clinical candidate evaluation, retrosynthesis planning and automated experimental platforms, organized around the Design-Make-Test-Analyze cycle. The Level 5 virtual scientist is explicitly framed as a theoretical ultimate vision, and the authors argue that for the foreseeable future the scientist-in-the-loop co-pilot paradigm remains the most practical and scalable deployment model.

The review states that fully independent Level 5 systems, able to define scientific questions and complete closed-loop research without human oversight, do not yet exist in drug discovery, and that current state-of-the-art advances concentrate at Levels 2 to 4. The authors also note that existing studies often fail to disentangle the contribution of agentic architectures from simply scaling foundation models or using better toolchains, and call for rigorous ablation and standardized quantitative metrics.