BioDiscoveryAgent · Agentic Research Directory

BioDiscoveryAgent

BioDiscoveryAgent iteratively selects genetic perturbations using earlier screening outcomes and optional literature, database and critic tools. The authors evaluate it by replaying existing genetic screens, with dataset-specific batch sizes. This supports a documented experiment-selection workflow; it does not demonstrate an autonomous wet laboratory. Code is public, but this review did not run it.

L3Multi-step tool-using agentic workflow Source Preprint (not peer-reviewed) Outcome Author-reported

Evidence & verification

Source level

Preprint (not peer-reviewed)
Grounded in a preprint that had not completed peer review at the time the source was checked.

Outcome verification

Author-reported
Outcomes are reported by the team that built the system. The entry is source-checked against the team's own publications; not independently reproduced.

Sources checked

2026-10-19 (UTC)

Checked source version: 2405.17631v3. Claims refer to this manuscript snapshot; final journal/proceedings version parity was not verified. Author-reported; independent replication not established in this review.

Source-checked does not mean independently reproduced. See the Method page for how these labels are assigned.

Sources (2)

Research applicability

Documented biology research workflow, with source-version and evaluation limits retained.

Getting started

Read the primary manuscript and its limitations first. Check implementation access and terms separately; this review did not run an installation, API call to a research system, or experiment.

Maintenance

Source snapshot checked 2026-10-19. Current availability and experimental version parity are separate claims.

License

Research code/data/manuscript terms require separate verification; repository metadata alone is not a reuse clearance.

Workflow & human oversight

L3 applies only to the documented tool-assisted retrospective experiment-selection variant. Researchers define the phenotype, historic screen and fixed round/batch budgets, and assess recovered hits against the measured screen. Optional literature/database/critic tools and previous measured responses inform later selections. The retrospective data lookup is not a newly executed wet-lab experiment. Evaluated replay uses five rounds and dataset-specific batch sizes. Scope does not imply all baselines/ablations used identical optional tools or claim prospective wet-lab autonomy.

Workflow level is not an evidence score; higher is not better evidence.

Approach

Selects perturbations, reads previously measured outcomes, then revises the next batch; literature/database tools and an AI critic support selection.

Outcomes (as reported by the originating team)

Six single-gene datasets; five rounds; normally 128 genes/round, but 32 for the 1,061-perturbation dataset; Figure 2 averages ten runs. Author-reported 21% more hits, and 46% on non-essential genes, relative to Bayesian-optimization baselines. These are relative improvements, not percentage points.

Limitations

Retrospective replay of existing screens, including one unpublished dataset; not an autonomous wet-lab demonstration or independent replication.

Reproducibility

Author-reported; independent replication not established in this review. Documentary review only; no candidate code, model, simulation or experiment was executed. A public repository or result artifact is not proof of replication.

Adoption requirements

Untested suggestion: inspect the pinned sources and validate data, evaluator assumptions, implementation availability and researcher controls before planning a bounded trial. No setup or trial was executed.

Comments

No comments yet.

Add a comment

Comments are held for moderation before appearing. Critical or corrective comments are welcome — especially evidence that an entry overstates what has been verified.