Empowering Biomedical Discovery with AI Agents

In plain words

Scientists already use software to help with lab work, but most of it does one narrow job and a person steers every step. This paper argues for "AI agents": programs that plan a research workflow, call other tools such as models, databases and lab machines, check their own results and learn from them, with people kept in the loop.

To make the idea concrete, the authors sort such agents into four levels by how much they do on their own. At the lowest level, software only helps a scientist form a hypothesis. At the top, an agent would propose hypotheses that no existing paper states and design new ways to test them. Today's agents reach the first level in narrow fields. The two higher levels have not been built yet.

The paper also lists what has to be solved first: unreliable answers, weak ways to test agents, missing biomedical data, and unclear rules for who is responsible. Its purpose is to give the field a shared vocabulary and a plan. It does not report experiments.

TL;DR

The paper is a perspective; it offers a view rather than a new system or a measured result §1. It defines four levels of autonomy, Level 0 (no agent) to Level 3 (agent as scientist), and scores an agent on three areas: hypothesis, experiment and reasoning. An agent gets the lowest level it reaches in any area §4 tab. 1. It describes four modules a compound agent needs: perception, interaction, memory and learning, and reasoning §5 fig. 4. The authors state that Level 1 agents exist in narrow fields, while Levels 2 and 3 "have yet to be realized" §4. The paper reports no experiments and no numbers.

Problem

Biomedical models have so far been built for one task each. Databases hold curated facts but cannot filter or refine them, and search engines retrieve but do not reason about what they found §2. The authors argue that conventional foundation models trained on existing data are not enough for "AI scientists", because a genuinely new hypothesis cannot already be contained in what the model learned from §1. They also name a data gap: few large experimental datasets cover biomedical areas beyond a handful, such as structural biology and single-cell science §1.

Method

The paper has no algorithm, model or training procedure. Its contribution is a conceptual framework made of three parts: the autonomy levels, a taxonomy of multi-agent collaboration, and a module decomposition for building an agent. Each part is summarised below.

Levels of autonomy §4 tab. 1

Agents are judged in three areas of discovery: hypothesis, experiment and reasoning. A level needs its capabilities in all three areas, so the lowest area sets the level §4. Table 1 lists what each level does in each area and how the scientist's role changes tab. 1.

The rule for combining areas is stated in the text: an agent with Level 3 in the experiment area and Level 2 in the other two is a Level 2 agent §4.

# Autonomy level of an agent  [§4] [tab. 1]
# Each area is rated with the level it reaches there, 0 to 3
# The agent is classed at the lowest of the three ratings
def autonomy_level(hypothesis_level, experiment_level, reasoning_level):
    return min(hypothesis_level, experiment_level, reasoning_level)

Table 2 gives one example per level for genetics, cell biology and chemical biology tab. 2. The text expands three of these examples §4:

• Genetics §4.1: a Level 1 agent writes code for quality control, imputation and association tests on genotype data from a genome-wide association study. A Level 2 agent looks for patient subgroups in which a drug may work and proposes in vitro checks. A Level 3 agent would start new study groups and develop validation methods.
• Cell biology §4.2: a Level 1 agent predicts cellular responses with integrated models and runs predefined pipelines. A Level 2 agent manages the experimental loop and updates its own models. A Level 3 agent would form hybrid virtual-cell models that combine digital and high-throughput physical agents.
• Chemical biology §4.3: current approaches are at Level 0. Level 1 agents run ML tools and design simple experiments. Level 3 is described as studying all types of molecular interactions in a cell.

Multi-agent collaboration schemes §3.2 fig. 3

The paper's view is that a single LLM can play several roles, but that heterogeneous teams of agents, which may include ML tools, specialised tools and human experts, fit biomedical work better §3. It names five collaboration schemes:

• Brainstorming: agents pool many ideas, with volume valued above quality at first §3.2.
• Expert consultation: one agent or human gathers evidence and gives feedback that others use to refine their work §3.2.
• Research debate: one team gathers evidence for a position and the other team critiques it §3.2.
• Round-table discussion: agents exchange and revise proposals over several rounds. Reconcile is cited as an example §3.2.
• Self-driving laboratory: the end-to-end discovery workflow runs in a loop, under the scientist's broad direction, without step-by-step oversight §3.2.

For assigning roles to agents, the paper names three approaches: fine-tuning on biological tasks, in-context role prompts, and automatic generation of roles §3.1.

Modules of a compound agent §5 fig. 4 fig. 5

The paper describes four modules and gives an example that uses all of them §5 fig. 5e. It does not specify how they connect.

• Perception §5.1: input from scientists, the environment and other agents, in modalities such as text, microscopy images, video, sensor readouts, proteomics and assay devices. Two alignment strategies are named. Textual translation turns readings into text. Representation alignment maps the output of a modality-specific encoder into the LLM's representation space through an adaptor and instruction tuning §5.1.
• Interaction §5.2: human-agent interaction (fine-tuning on dialogues, RLHF, or direct preference optimisation), multi-agent interaction, and tool use. Tools include application APIs, search engines, ML models, databases and robots. Agents call tools either by in-context prompting or by fine-tuning on tool-use demonstrations §5.2.
• Memory and learning §5.3: long-term memory holds knowledge either in the weights or in an external store, such as a GWAS database. Short-term memory is held in the context window or as latent embeddings. Long-term memory can be updated by parameter-efficient fine-tuning, interactive learning, model editing, or by editing the external store §5.3 fig. 5b fig. 5a.
• Reasoning §5.4: direct reasoning includes chain-of-thought, least-to-most, self-consistency, tree-of-thoughts and graph-of-thoughts. Reasoning with feedback adjusts the plan from experiment results or human feedback, as in ReAct, Voyager and Reflexion-style self-assessment §5.4 fig. 5c fig. 5d.

Hyperparameters

The paper reports no training runs or model settings. The table lists what it does state.

Results

Not applicable. The paper reports no experiments, benchmarks or measured outcomes. Its claims are qualitative. In particular, the claim that Levels 2 and 3 "have yet to be realized" is stated without evidence in the text §4. Systems it cites as examples, such as ChemCrow, AutoBa, Coscientist and Reconcile, are described by what they do, without numbers from this paper §3.1 §3.2 §4 §5.2.

Paper vs code

No official code.

The paper names no code repository, project page or GitHub link in its text, and the arXiv metadata has no comment with a link. A GitHub repository search on the title words returned nothing. A search inside the mims-harvard organisation returned two repositories for other papers, KGARevion and ATHENA, not for this perspective. The paper has no code to compare against.

Pitfalls

• Nothing is measured. Every level, module and collaboration scheme is a description. The paper does not show that any agent of Level 2 or 3 works, or that the scheme helps.
• The level boundaries are judgement calls. The paper gives no test for "creative", "extrapolated" or "linear continuation of literature". Level 2 is described as hypotheses that continue trends, and Level 3 as hypotheses that do not §4 tab. 1. Two people could rate the same agent differently.
• Section and table numbers are counted, not printed. The LaTeX source uses unnumbered section headings. This card numbers them in order: §1 Introduction, §2 data-driven models, §3 types of agents, §4 levels, §5 roadmap, §6 challenges, §7 outlook. Subsections are counted within these. The reference to "Figure 5e" in §5 matches the panel labels in the figure caption.
• Figure and table numbers follow the source order. Figure 1 is the overview, Figure 2 the timeline, Figure 3 the agent roles, Figure 4 the four modules, Figure 5 the worked example, Figure 6 the challenges. Table 1 is the levels table and Table 2 the examples. The figures themselves are not reproduced here.
• Version. The source used is v2, updated 2024-07-24 [raw metadata]. The published date on this card is that of v1.
• Ethics and governance are opinion. The paper argues for human oversight, staged autonomy and international standards §6 §7. It gives no concrete rules, thresholds or regulations.
• Costs are not quantified. No compute, data-collection or lab costs are given. The paper says agents can make discovery workflows more resource-efficient §1 and that an autonomous experiment without careful planning may cause harm §6, with no figures for either.
• Multi-agent risks are asserted, not shown. The paper says errors compound across interacting agents and that hazardous-substance misuse is a concern §6. No failure rate or example study is given.

Lineage

• Builds on: the LLM agent work it cites as building blocks, including ReAct, Toolformer, Reflexion, Voyager and HuggingGPT §5 §5.2 §5.4. It also builds on chain-of-thought prompting §5.4.
• Compared against: an existing autonomy taxonomy for general AI systems, which the authors say does not consider biomedical discovery §4.
• Superseded by: not stated.