BioMARS 2.0: a robotic biologist for formulating and adapting biological experiments
Authors: Yibo Qiu, Zan Huang, Yilin Qiao, Ziyang Ling, Yi Wang, Zhengyang Lu, Hangke Peng, Haoliang Ye, Ao Tian, Shu'ang Sun, Handi Liu, Yuqing Long, Mohan Yang, Kan Ding, Zhenzhen Zhou, Di Wu, Yu Xiong, Yaxuan Han, Jun'ao Song, Haoran Zhang, Ronald X Xu, Mingzhai Sun
Organizations: Suzhou Institute for Advanced Research, University of Science and Technology of China, Suzhou, Jiangsu, China. · School of Biomedical Engineering, Division of Life Sciences and Medicine, University of Science and Technology of China, Hefei, Anhui, China.
Biological experimentation is not reducible to a single optimization problem, but requires decisions at multiple levels of abstraction: translating scientific knowledge into experimental procedures, adapting culture conditions to measured phenotypes and selecting interventions that test biological hypotheses. Automating experimental execution does not, by itself, determine which experiments should be performed next or how they should change in response to biological feedback. Here we introduce BioMARS, a human-in-the-loop multi-agent robotic system that connects knowledge-guided experimental decision-making with robotic execution and biological feedback. At the procedural level, BioMARS generated literature-grounded protocols and traceable action plans, improving expert-reviewed protocol quality scores across four model backends. At the parametric level, BioMARS used image-derived phenotypes from robotic retinal organoid culture to select subsequent experimental conditions, with higher reported scores in later rounds of the three-round experiment. At the intervention level, BioMARS nominated drug-repurposing candidates from biomedical evidence and advanced six candidates to wet-lab testing by researchers. Together, these studies demonstrate feedback-driven robotic biological experimentation within a framework that supports decisions about experimental procedures, culture conditions and biological interventions.
Figures & tables
Figure 1
Figure 2: Procedural-level formulation and action mapping. Full legend follows.
Figure 3Figure 4
Appendix figures & tables18 assets
Supplementary material from the paper’s appendix.
Appendix
Figure 1: Reagent dispensing in the automated laboratory. a , Automated cell-culture laboratory. b–d , Reservoir and reagent-bottle retrieval, followed by cap opening. e–h , Single-channel tip pickup, aspiration, transfer to the reservoir and tip ejection. i–m , Multichannel tip pickup, aspiration, 96-well-plate lid opening, dispensing and tip ejection.
Figure 2: Laboratory stations and infrastructure. a–k , Cell counter, reagent-reservoir platform, reagent-reservoir work area, CO 2 incubator, isolation chamber, optical table, constant-temperature incubator, plate lid-opening station, waste bin, centrifuge-tube cap-opening station and pipette-tip rack holder, respectively.
Function
Main arguments
Operation scope
Get container get_new_container_from_storage
Storage location; target workstation; container type
Acquire a new receiving vessel. Acquisition does not fill, coat or load it.
Discard container discard_container
Source workstation; container type
Dispose of a vessel designated as waste; distinct from returning a vessel to storage.
Retrieve an existing reagent or sample in its container.
Return reagent return_reagent
Reagent identity; container type; workstation
Return the same reagent or sample container to storage after use.
Transfer liquid transfer_liquid_between_workstations
Source and target workstation/types; volume; addressed wells; optional mixing
Transfer liquid or suspension to the specified receiver. Plate addressing follows the supported column or well convention.
Resuspend cells resuspend_cells_in_tube
Tube workstation and type; medium and volume fields
Mix cells already in a supported tube. This call does not itself add medium or adjust concentration.
Appendix
Table 1: Action catalogue used for protocol mapping. Thirteen functions covering material handling, liquid operations, culture and measurement. Entries summarize function arguments and operation semantics.
Code
Maximum
Assessment criterion
C
30
Task-scope alignment: coverage of required functions, with deductions for explicitly unrelated experimental functions.
A
30
Content accuracy: fidelity of described operations and key conditions to the source.
L
20
Experimental logic: consistency of samples, branches, ordering, repetitions and input–output relationships.
S
10
Protocol structure: hierarchy, step localization, reading order, information accessibility and concision.
D
10
Procedural detail: retention of applicable parameters in the operations described.
Appendix
Table 1: Protocol-assessment rubric and source-data codes. The five criteria sum to 100 points. Each task uses a common source-based checklist across all models and pipelines. Domain experts reviewed system-assisted assessments and confirmed final scores.
Model
FlashRAG
BioMARS
Gain, points (95% CI)
Gain, % (95% CI)
Counts
GPT-4o
73.80
86.71
12.91 (9.23, 16.53)
17.5 (12.2, 23.1)
40/1/6
Qwen 3.6 Flash
78.35
94.69
16.34 (13.44, 19.33)
20.9 (16.6, 25.6)
44/0/3
DeepSeek V4 Flash
81.98
93.89
11.92 (8.72, 15.28)
14.5 (10.3, 19.2)
40/1/6
Gemini 3 Flash Preview
76.51
90.11
13.60 (10.90, 16.40)
17.8 (13.8, 22.1)
42/1/4
Appendix
Table 2: Paired protocol-assessment gains. BioMARS and FlashRAG were evaluated on the same 47 tasks for each model. Scores range from 0 to 100 under the five-dimensional rubric, with system-assisted assessments reviewed and confirmed by domain experts. Absolute gains are paired mean differences; relative gains express the difference as a percentage of the FlashRAG mean. Parentheses show percentile 95% confidence intervals from 20,000 paired-task bootstrap resamples. Counts are improved/tied/declined tasks.
Figure 1: Changes in protocol-assessment dimensions. Cells show mean paired BioMARS-minus-FlashRAG score differences, expressed as percentages of dimension maxima ( n=47 tasks per model). Blue and red indicate positive and negative differences, respectively. This normalization differs from the relative mean gains in Fig. 2 c.
Dimension
Weight
Qwen n
Qwen score
DeepSeek n
DeepSeek score
Operation matching
30
119
94.29
119
93.20
Known arguments
25
66
83.33
67
71.64
Objects and handling
25
111
76.58
112
80.36
Required order
10
30
96.67
32
93.75
Unknown information
10
47
97.87
47
95.74
Total
100
119
85.56
119
83.68
Appendix
Table 3: Step-level Action assessment. Dimension scores are unweighted means on a 0–100 scale across applicable steps; n is the number of scored steps for that dimension. The total is the mean normalized step score across all 119 steps; dimension means use their respective applicable subsets.
Round
Algorithm
Selected pair
Mean
Median
Maximum score
Top-3 mean
Fraction ≥0.8
Fraction ≥0.9
Failures
High-noise
Round 1
BO
BO–Round1–Plate1 BO–Round1–Plate2
0.429
0.418
0.798
0.725
0.000
0.000
1
1
MASBO
MASBO–Round1–Plate2 MASBO–Round1–Plate3
0.435
0.418
0.974
0.757
0.100
0.100
1
2
Round 2
BO
BO–Round2–Plate1 BO–Round2–Plate3
0.654
0.578
0.915
0.885
0.400
0.100
0
0
MASBO
MASBO–Round2–Plate1 MASBO–Round2–Plate2
0.752
0.826
0.897
0.881
0.700
0.000
0
0
Round 3
BO
BO–Round3–Plate2 BO–Round3–Plate3
0.764
0.855
0.920
0.906
0.700
0.200
1
–
MASBO
MASBO–Round3–Plate2 MASBO–Round3–Plate3
0.923
0.941
0.949
0.949
1.000
0.800
0
–
Appendix
Table 3: Round-wise optimization outcomes. All statistics are calculated from the final score obtained using the highest-Pearson-r plate pair.
Figure 1: Plate-level concordance and variation. a , Scores from the most highly correlated plate pair for each method and round ( n=10 conditions); the diagonal indicates equality. b , Pearson correlation, concordance correlation coefficient (CCC) and mean absolute error (MAE) for these pairs. c , Three-plate score ranges, with points indicating selected-pair means. Grey, BO; blue, MASBO.
Figure 2: Plate-level phenotype scores in optimization round 1. a–c , BO plates; d–f , MASBO plates. Numbers show original well scores on a shared 0–1 colour scale; white cells indicate masked or missing values. Condition-level failure-zero assignment is described in Methods.
Figure 3: Plate-level phenotype scores in optimization round 2. a–c , BO plates; d–f , MASBO plates. Numbers show original well scores on a shared 0–1 colour scale; white cells indicate masked or missing values.
Figure 4: Plate-level phenotype scores in optimization round 3. a–c , BO plates; d–f , MASBO plates. Numbers show original well scores on a shared 0–1 colour scale; white cells indicate masked or missing values. BO column 11 retains its original well scores here but is assigned zero as an experimental failure in condition-level analyses.
Region or unit
Role in analysis
Columns 1 and 12
Control columns (W01/W12); excluded from candidate scores.
Rows A and H
Edge rows; excluded from candidate scores.
Rows B–G, columns 2–11
Candidate region. Each column corresponds to one condition, with up to six technical-replicate wells per plate.
Blank or invalid measurements
Excluded from the within-plate mean; not converted to zero.
Plate1, Plate2 and Plate3
Independent biological batches, each represented by one plate. Full identifiers specify algorithm, optimization round and plate (for example, BO–Round1–Plate1).
Final condition score
Average valid wells within each plate, then equally average the condition means from the highest-Pearson-correlation plate pair.
Appendix
Table 4: Plate layout and score aggregation for organoid optimization. W02–W11 identify conditions, whereas a coordinate such as B2 identifies an individual well.
Figure 5: Observed culture conditions within the six-factor search space. a–f , Scores for the same 60 conditions plotted against the six indicated factors (10 conditions per method and round). Scores are aggregated as in Fig. 4 f. Grey and blue denote BO and MASBO; circles, squares and triangles denote rounds 1–3, respectively.
Candidate/ origin
System-generated hypothesis
Retrieved evidence from the same lineage
Ifenprodil One-hop link
Inhibition of GluN2B-containing NMDARs would reduce intracellular calcium overload and improve viability during NMDA exposure in retinal organoids.
Receptor pharmacology and developing rat cortical neurons support GluN2B-dependent antagonism and subtype-dependent ifenprodil sensitivity (rounds 2–5) [ 38 , 59 ] . These findings ground the receptor component of the hypothesis; organoid protection is an extrapolation.
Milnacipran HCl One-hop link
Non-competitive NMDAR antagonism would reduce GRIN1-associated calcium influx and subsequent inflammation. The system proposed IL-6 and TNF- α measurements to test this response.
Retrieved literature reports non-competitive NMDAR antagonism by milnacipran itself [ 40 ] and NMDA-channel blockade and neuroprotection by conformationally restricted analogues (rounds 1–6) [ 60 , 41 ] . These findings support an NMDAR-related pharmacological rationale, but do not establish the proposed anti-inflammatory mechanism in organoids.
Agmatine sulfate One-hop link GRIN1
Competitive antagonism at GRIN2B and activation of Nrf2 were proposed to reduce calcium influx and oxidative stress during NMDA exposure.
Polyamine-site binding and spinal NMDA-evoked calcium transients support receptor modulation (rounds 1–5) [ 39 , 61 ] . Rotenone-lesioned rats provide Nrf2-related evidence (rounds 1–5) [ 62 ] ; a retinal-neuroinflammation study reports protection in mice (round 5) [ 63 ] . Polyamine-site evidence does not establish competitive GRIN2B antagonism.
Stiripentol Predicted link
GRIN2A/2B modulation was refined into a hypothesis of enhanced GRIN2B-mediated signalling, improved mitochondrial dynamics and reduced oxidative stress in retinal organoids.
A drug-specific study reports neuronal protection and reduced sodium/calcium entry, but no detectable binding affinity for the tested glutamate receptors (round 4) [ 64 ] . Retinal-organoid evidence links mitochondrial oxidative stress to RGC death (round 5) [ 65 ] . Neither establishes the proposed enhancement of GRIN2B signalling.
Phenytoin Predicted link
GRIN2B/2D modulation would reduce reactive oxygen species and improve mitochondrial respiration during NMDA exposure. Proposed readouts were oxygen consumption and malondialdehyde levels.
Phenytoin produced modest electrocorticographic correction in a GRIN2D-variant mouse model (rounds 1–5) [ 66 ] . A neuronal-injury study links GluN2B signalling to mitophagy with THSG, a different compound (round 2) [ 67 ] . This supports a pathway rationale, not direct receptor engagement by phenytoin.
Memantine One-hop link GRIN1
Selective inhibition of extrasynaptic NMDARs and GRIN1-dependent desensitization were proposed to reduce calcium overload and NMDA-induced organoid injury.
Retrieved studies report protection against NMDA-induced retinal degeneration (round 1) [ 68 ] , calcium-dependent NMDAR desensitization (rounds 1–4) [ 69 ] , and autophagy in ARPE-19 cells (rounds 3–4) [ 70 ] . Organoid-specific receptor selectivity remains part of the hypothesis.
Appendix
Table 1: System-generated hypotheses and retrieved evidence for six BioMARS candidates. Hypotheses summarize archived outputs from one selected parent-linked lineage per candidate. Evidence is selected across all rounds of that lineage, prioritizing drug-specific findings, relevant mechanisms and retinal models while retaining findings that constrain interpretation. Proposed mechanisms are distinguished from literature findings.
Candidate/ source / rounds
Ref.
Study model and principal finding
Relevance to this study
Ifenprodil D1 R2–5
[ 38 ]
Receptor pharmacology: selective antagonism of NR2B/GluN2B-containing NMDARs.
Direct pharmacological support; retinal organoids were not examined.
Ifenprodil D2 R2–5
[ 59 ]
Developing rat cortical neurons: NMDAR subtypes differ in ifenprodil sensitivity.
Supports receptor-subtype dependence; extrapolation from developing cortical neurons is required.
Ifenprodil D3 R2–5
[ 71 ]
A mouse PTSD model: Cheonwangbosimdan reduced PTSD-like behaviours through GluN2B-containing NMDAR antagonism.
Supports the target mechanism using a different compound and disease model.
Milnacipran HCl D1 R1–6
[ 60 ]
Conformationally restricted milnacipran analogues show NMDA-channel-blocking activity.
Analogue-based support, rather than direct evidence for the parent drug.
Milnacipran HCl D2 R1–6
[ 40 ]
Milnacipran itself has non-competitive NMDAR-antagonist activity; its conformationally restricted analogue PPDC is approximately 30-fold more potent.
Supports NMDAR pharmacology of the parent drug and analogue development; no retinal model was tested.
Milnacipran HCl D3 R1–6
[ 41 ]
Conformationally restricted analogues block open NMDA channels and protect against NMDA-induced neurotoxicity.
Analogue potency and dose cannot be transferred directly to milnacipran.
Appendix
Table 2: Literature evidence retrieved across candidate-refinement rounds. Twenty-one sources are grouped by candidate and drawn from the same lineages as Table 1 . D identifies a source within a candidate; R identifies refinement rounds containing that source. Study findings are reported separately from model inferences.
Figure 1: Additional drug experiments during NMDA exposure. a–c , Memantine experiment 3, agmatine experiment 2 and phenytoin experiment 3; d,e , MK-801 experiments 1 and 2. Numbers identify datasets. Panels b,e share Ctrl and NMDA controls; panel d shares its control groups with Mem, Agm and Phe in Fig. 5 e. Dots denote organoids; bars show means with SD error bars for area-normalized ATP signals. Group sizes are listed in Table 4 . Ctrl, normal medium; NMDA, NMDA alone; drug groups received simultaneous NMDA exposure. Bracketed comparisons use two-sided Welch t-tests with unadjusted P values: ∗P<0.05 , ∗∗P<0.01 , ∗∗∗P<0.001 . Other experiments appear in Fig. 5 d,e.
Compound/experiment
Group
n
Mean
SD
Change (%)
Fig. 4 panel
Ifenprodil 1
Ctrl
6
0.455
0.109
–
d
NMDA
5
0.363
0.083
–
1 μ M
8
0.297
0.107
-18.1
10 μ M
8
0.318
0.104
-12.3
100 μ M
8
0.297
0.101
-18.2
Stiripentol 1
Ctrl
6
0.838
0.130
–
d
Appendix
Table 4: ATP viability measurements across the reported drug experiments. Means and SDs summarize ATP luminescence divided by projected organoid area; n denotes the number of organoids assayed. Drug-group changes are calculated relative to the displayed NMDA-group mean, without pooling datasets. Ifenprodil, stiripentol and milnacipran hydrochloride each have one preliminary screen. The three experiments per selected candidate include one preliminary screen and two subsequent experiments. Mem 2, Agm 3 and Phe 2 are preliminary-screen datasets; numbers do not indicate chronological order. Drug-treated and control groups in Agm 3, and drug-treated and NMDA reference groups in Phe 2, came from different batches; these comparisons are descriptive only. Ctrl denotes normal medium. MK-801 is a separate positive control, not one of the six nominated candidates. This table summarizes descriptive statistics; tests are specified in Methods and the corresponding figure legends.
Biological wet-lab protocols are written for trained researchers and often leave routine operations, state-dependent conditions, and contextual parameters implicit, making them difficult to translate into robot-executable actions. We present ProtoAct, a structured protocol-grounding framework that converts free-form biological procedures into state-aware, embodiment-ready action sequences. ProtoAct uses ProtoRAG to retrieve manually annotated examples for context-sensitive parsing, employs RefineChecker to detect and revise missing or inconsistent steps, and applies ActSchema to map the refined procedure into constrained JSON function sequences. We further introduce BioP2E, for which we manually annotate 22 cell-culture protocols into 258 monitoring conditions, 910 executable subtasks, and 962 grounded action calls. Evaluation across seven large language models demonstrates that ProtoAct can be effectively instantiated with different backbones. Ablations confirm that retrieval, posterior checking, and schema constraints make complementary contributions. The parsed subtasks further support demonstration collection and VLA model training, enabling successful execution in both simulation and real-robot settings. ProtoAct thus provides a practical interface between biological protocol understanding and embodied robotic execution.
Zhe Liu, Jiaming Gu, Zhaohui Du +7
Key Laboratory of Smart Manufacturing in Energy Chemical Process, Ministry of Education, East China University of Science and Technology, Shanghai, China · Department of Computer Science and Engineering, East China University of Science and Technology, Shanghai, China · Department of Laboratory Medicine, Ruijin Hospital, Shanghai Jiao Tong University School of Medicine, Shanghai, China +1
Autonomous wet-lab experimentation requires more than plausible protocol text: biological intent, quantitative procedures, device constraints and experimental feedback must remain aligned from protocol and SOP design to code and physical execution. We developed ProtoPilot, a self-evolving multi-agent system, together with an expert-grounded benchmark and evaluation framework for testing this conversion as an experimental automation problem. The framework spans 294 synthetic-biology and molecular-biology tasks derived from 98 gold-standard protocols, wet-lab expert rubrics, device-level validity gates and real experimental tests. ProtoPilot incorporates layer-wise verifiability, multi-agent orchestration and a runtime-updated skill library to generate protocols, expand SOPs, synthesize SDK-compliant code and revise workflows from wet-lab feedback. It achieved a Top@3 expert-preference rate of 90.2%, an overall protocol-to-code gate pass rate of 89.5% and an Opentrons pass rate of 88.24%, compared with 32.35% for OpenTrons-AI. Wet-lab validation produced interpretable readouts, Sanger-confirmed products and feedback-corrected PCA-assembled DNA targets, establishing a verifiable route to autonomous experimentation. Together, these results show that the evaluation framework captures execution-relevant requirements for autonomous wet-lab automation, and that ProtoPilot can meet them by converting protocol and code generation into validated execution and feedback-guided revision.
Yankai Jiang, Weiting Tang, Haoran Sun +11
1Shanghai Artificial Intelligence Laboratory, Shanghai, China · 2Genoria AI, Shenzhen, China
In this paper, a bio-robot is an engineered living or biohybrid system in which living cells perform one or more core functions, such as sensing, information processing, actuation or output. We focus on systems whose cell-based functions are programmed by genetic circuits; physical movement is optional. Designing such a system requires translating application requirements into sensing, logic or memory, output, assembly, host and containment modules, while grounding each choice in traceable parts and evidence. We present micro_biorobot_agent, an offline multi-agent system built on Qwen3.5-27B. The system combines requirement analysis, module-specific retrieval, candidate assembly, conflict checking, local repair, independent review and validation over an integrated library of 23,762 records covering biological parts, measured combinations, literature-supported relationships and actuation evidence. Deterministic output checks align the final report with the retrieved part set and correct false gaps, unsupported part mentions and source-tracking errors. On two author-developed evaluation sets of 50 queries each, the system obtains mean overall scores of 7.35 and 8.04, the highest among the seven evaluated systems; on Scenario Design it exceeds the runner-up by 2.23 points. A 50-query paired ablation shows that the source-tracking check reduces false-gap incidents from 15 to 3, an 80% reduction, and increases source accuracy by 0.75 points. This paper reports the Qwen3.5-based v1 system and evaluates high-level design reports rather than experimentally validated circuits.