Biological figure interpretation
A small, auditable pilot examining how multimodal AI interprets life-science evidence—and where automated checks need scientific review.
Relevant terminology. A missed biological error.
A motor-neuron figure asks the model to distinguish tract placement, commissural laterality, and dorsal nerve cord continuity. The recorded scientific review identifies a panel-level error that the configured lexical checks did not flag.
Recorded automated composite 56.7 · Provisional assessment 60.0 · Different scoring approaches
- 01Figure + context
- 02Model interpretation
- 03Lexical checks
- 04Scientific review
What the model said
Panel E: two arrowheads mark points along the DNC where the fluorescent process appears discontinuous or interrupted, with visible varicosities/puncta.
Exact excerpt from the saved response.
What the reference establishes
Panels D-E show continuous wild-type DD/VD dorsal nerve cord signal and a gap in the mutant dorsal nerve cord.
Recorded review: Misreads panels B and C, adds unsupported varicosities in E, and does not cleanly map wild-type versus mutant phenotypes.
Automated score is coincidentally similar but misses the key panel-level contradictions and morphology hallucinations. A lack of configured phrase matches does not establish that an interpretation is biologically correct.
Read the supplied figure context
The figure tests whether loss of the extracellular leucine-rich-repeat protein LRON-11 affects ventral nerve cord tract placement, commissural laterality, and dorsal nerve cord continuity in C. elegans motor neurons.
Panels A-C use DD/VD motor-neuron markers: wild type shows neurites in the right ventral nerve cord tract, whereas mutants can show neurites in the left tract or a commissure on the incorrect side. Panels D-E compare DD/VD dorsal nerve cord continuity and show a mutant gap. Panels F-G use a DA/DB motor-neuron marker and show mutant commissures extending on the incorrect side. Anterior is left; scale bars are 20 micrometres.
Original figure ↗Source paper ↗
Figure 4 from Kokan et al. (2022), PLOS ONE, CC BY. License: Creative Commons Attribution (CC BY).
Inspect the full response and scoring details ↓Inspect the evidence, case by case
Read the numerical summaries alongside the response, reference, and scientific notes. A low lexical match can reflect a correct paraphrase; a high lexical-check score can still miss a biological error.
| Case | Domain | Automated composite | Lexical coverage | Provisional review | Configured triggers | Requested model label |
|---|---|---|---|---|---|---|
| BFB-001 | Cell biology and quantitative microscopy | 55.0 | 30.8 | 92.0 | 0 | claude-sonnet-5 |
| BFB-003 | Single-cell genomics | 65.0 | 46.2 | 90.0 | 0 | claude-sonnet-5 |
| BFB-004 | Cancer biology and functional genomics | 62.9 | 42.9 | 80.0 | 0 | claude-sonnet-5 |
| BFB-005 | Preclinical therapeutics and drug delivery | 58.2 | 35.7 | 91.0 | 0 | claude-sonnet-5 |
| BFB-002 | Neurodevelopment and C. elegans genetics | 56.7 | 33.3 | 60.0 | 0 | claude-sonnet-5 |
Mean recorded response latency: 20.23 seconds. Legacy results record the requested model label; provider-returned model identity and original run environment were not retained.
What does label-free holotomography establish?
Automated composite 55.0 · Lexical coverage 30.8 · Provisional review 92.0 · Configured triggers 0
- What worked: Correctly interprets HTM as label-free structural imaging and explicitly requires orthogonal fluorescence for organelle identity.
- Scientific issue: Some morphology is described more specifically than necessary; sensitivity/specificity is not explicitly proposed.
- Scoring limitation: Large lexical false-negative: the response semantically covers most rubric concepts but uses different wording.
Read the structured model response
experiment
Panel A schematically compares two imaging modalities sharing the same 60x/0.8NA dry objective and sample plane: holotomography (HTM), which uses a rotating coherent object beam combined with a reference beam via a 50/50 beam combiner to reconstruct 3D refractive-index (RI) maps, versus epifluorescence, which uses excitation light reflected off a multiband dichroic mirror to collect fluorescence emission. Panel B applies HTM alone to unlabeled, unperturbed HeLa cells, showing a 2D RI map of multiple cells, a zoomed 2D RI inset of a subcellular region, and a corresponding 3D RI reconstruction of the same region, with an RI scale bar from 1.33 to 1.399.
observations
- Panel A shows HTM and epifluorescence share the same objective, sample stage, and focus plane but diverge in beam path: HTM uses a rotating object beam plus reference beam and a 50/50 dichroic combiner, while epifluorescence uses excitation light and a multiband dichroic mirror to select fluorescence emission.
- Panel B (top left) shows a 2D RI grayscale image of several flat HeLa cells with numerous small bright puncta scattered across the cytoplasm and a smooth, low-contrast nucleus-like region, with no fluorescent labeling.
- Panel B (bottom left) is a magnified 2D RI view of a boxed region showing distinct rounded high-RI (dark) structures of varying size intermixed with elongated, tubular/branching lower-contrast structures.
- Panel B (right) is a 3D RI reconstruction of a similar cellular region showing a dense population of ring-like and rounded high-RI structures concentrated in a perinuclear-like zone, with a large, low-RI circular void near the top.
- The RI scale bar (1.33-1.399) is provided for the 2D/3D HTM images, indicating that pixel/voxel intensity encodes refractive index rather than fluorescence intensity.
controls and statistics
- No fluorescence, immunostaining, or organelle-specific dye co-labeling is shown in Panel B to confirm the molecular identity of any RI structure (e.g., mitochondria, lipid droplets, vesicles).
- No quantitative RI value thresholds, size distributions, or statistical comparisons (e.g., number of puncta per cell, mean RI per structure type) are reported for the images shown.
- No sample size (number of cells, replicates, or independent experiments) is stated for the images in Panel B.
- No colocalization or correlative light-electron microscopy (CLEM) data are shown to validate that a given RI-bright or RI-dark structure corresponds to a specific organelle.
- No description of segmentation or classification algorithm criteria used to assign structures (round vs. tubular) to putative organelle categories is provided in this excerpt.
supported conclusion
HTM can generate label-free, high-resolution 2D and 3D refractive-index maps of unlabeled, unperturbed HeLa cells that resolve substructure with distinguishable morphologies (e.g., rounded versus elongated/tubular features) differing in refractive index, demonstrating feasibility of the imaging setup for capturing subcellular architecture without fluorescent probes.
unsupported claim
That the rounded high-RI puncta are lipid droplets or specific vesicles, or that the tubular structures are mitochondria or endoplasmic reticulum, cannot be concluded from RI morphology alone in this figure, since no orthogonal fluorescent or molecular marker is shown linking these RI patterns to a named organelle identity.
follow up
Perform simultaneous (or serial, co-registered) fluorescence labeling of candidate organelles (e.g., mitochondria, lipid droplets, endosomes) using the epifluorescence arm shown in Panel A on the same cells imaged by HTM, then quantitatively correlate fluorescent signal localization and intensity with specific RI value ranges and morphological classes to establish a validated RI-to-organelle identity mapping.
Inspect the scoring breakdown
Composite = 65% lexical coverage + 25% configured lexical-check score + 10% completeness. Lexical-check score: 100.0; completeness: 100.0. These are rubric components, not biological accuracy.
| Response field | Coverage /100 | Unmatched reference concepts |
|---|---|---|
| experiment | 50.0 | holotomographic microscopy coupled with epifluorescence |
| observations | 33.3 | multiple internal structures are visible; does not establish the molecular identity of each internal structure |
| controls and statistics | 50.0 | validated fluorescent organelle markers |
| supported conclusion | 0.0 | information-rich label-free images suitable for quantitative analysis; organelle identities require independent validation |
| unsupported claim | 0.0 | every high-refractive-index structure is a mitochondrion or lipid droplet; superior to all fluorescence microscopy methods |
| follow up | 50.0 | quantify classification sensitivity and specificity |
Configured claim/contradiction triggers: None.
Read the supplied context and reference
The study evaluates holo-tomographic microscopy (HTM), a label-free method that reconstructs cellular refractive-index maps, and compares it with fluorescence-assisted identification of subcellular structures.
Panel A diagrams the HTM setup coupled to epifluorescence. Panel B shows 2D and 3D HTM images of flat, unlabeled, unperturbed HeLa cells and motivates a quantitative strategy for interpreting the image content.
Reference conclusion: HTM can generate information-rich label-free images of living cells that are suitable for subsequent quantitative analysis, while organelle identities require independent validation.
Model task: Identify the workflow and describe what the figure directly demonstrates. Explain what control or orthogonal information would be required to assign a particular refractive-index structure to a named organelle. State the strongest supported conclusion, one overinterpretation, and one high-value follow-up.
Original figure ↗Source paper ↗
Figure 1 from Sandoz et al. (2019), PLOS Biology, CC BY. License: Creative Commons Attribution (CC BY).
Benchmarking a single-cell nascent-RNA sequencing workflow
Automated composite 65.0 · Lexical coverage 46.2 · Provisional review 90.0 · Configured triggers 0
- What worked: Accurately distinguishes strong aggregate agreement from weaker single-cell agreement and proposes an informative reproducibility/sensitivity study.
- Scientific issue: The statement that the slope indicates likely undercounting is plausible but not established by this figure alone.
- Scoring limitation: Lexical scorer misses semantically equivalent descriptions of barcoded nuclei, genomic composition, and assay limitations.
Read the structured model response
experiment
The figure introduces scGRO-seq, a single-nucleus method that labels nascent RNA with 3'-(O-propargyl)-NTPs during run-on, attaches single-cell barcodes and UMIs via CuAAC click chemistry, and sequences the resulting cDNA to capture strand-specific nascent transcription at single-cell resolution. Panels benchmark aggregate scGRO-seq signal at a gene (Hells) and an enhancer against orthogonal chromatin/transcription assays (PRO-seq, p300, ATAC-seq, histone marks, dREG) and against bulk inAGTuC and single-molecule intron seqFISH measurements, plus report reads-per-cell distributions and genomic feature distribution of reads.
observations
- Panel a diagrams the workflow: nuclei isolation, run-on with propargyl-NTPs, click-ligation of barcoded DNA adapters, single-nuclei sorting into barcoded/UMI wells, reverse transcription, PCR amplification with p5/p7 adapters.
- Panel b shows strand-specific scGRO-seq single-cell read tracks (sense/antisense) at a gene locus (Hells, Chr19) and an enhancer locus (Chr3), aligned with collapsed scGRO-seq, inAGTuC, PRO-seq, p300, ATAC-seq, H3K4me3, H3K27Ac, H3K4me1, and dREG tracks.
- Panel c shows a histogram of reads per cell across 10 experimental batches, with most cells having roughly 1,000–5,000 reads.
- Panel d left: scatter of scGRO-seq vs inAGTuC UMIs per million at gene bodies, fit y=0.94x, r^2=0.93.
- Panel d right: scatter of scGRO-seq vs inAGTuC UMIs per million at enhancers, fit y=0.5x, r^2=0.99.
- Panel e: scatter of single-cell scGRO-seq UMIs per cell vs intron seqFISH counts per cell, fit y=0.26x, r^2=0.58.
- Panel f: stacked bar chart of fraction of reads mapping to 3'UTR, 5'UTR, CDS, intronic, and intergenic regions for scGRO-seq versus scRNA-seq, showing scGRO-seq has a much higher intronic/intergenic fraction and lower CDS/UTR fraction than scRNA-seq.
controls and statistics
- Reported linear fit slopes and r^2 values (0.94x/r^2=0.93 for gene body; 0.5x/r^2=0.99 for enhancer; 0.26x/r^2=0.58 for single-cell seqFISH comparison) but no confidence intervals, p-values, or statistical test descriptions are given.
- Comparisons are made against an orthogonal bulk assay (inAGTuC) and single-molecule imaging (intron seqFISH), serving as external validation rather than internal replicate controls.
- Multiple experimental batches (10 listed) are shown for reads-per-cell distribution, suggesting some assessment of reproducibility across runs, but no explicit batch-effect statistics are reported.
- Color scale in scatter plots denotes local point density ('No. of neighbours'), used to visualize point overlap rather than as a statistical measure.
- No sample-size (n cells, n genes/enhancers) values, error bars, or reproducibility statistics (e.g., across biological replicates) are provided in the legend excerpt.
supported conclusion
Aggregate scGRO-seq signal correlates strongly with an independent bulk nascent-RNA method (inAGTuC) at both gene bodies and enhancers, and single-cell scGRO-seq signal shows a positive but weaker correlation with single-molecule intron seqFISH counts, supporting that scGRO-seq captures nascent transcriptional signal that is broadly consistent with orthogonal measurements at both aggregate and single-cell levels, with quantitative agreement being much stronger at the aggregate level than at the single-cell level.
unsupported claim
That scGRO-seq is more accurate, more sensitive, or superior to existing single-cell transcription assays such as intron seqFISH or scRNA-seq, or that it fully captures the same nascent transcription events as these methods, is not supported; the modest r^2 (0.58) and slope (0.26x) between scGRO-seq and seqFISH indicate substantial unexplained variance and likely undercounting, so the data show correlation/agreement in trend, not equivalence, identity, or superiority of detection sensitivity.
follow up
Perform paired scGRO-seq and intron seqFISH (or another single-molecule nascent-RNA imaging method) on the same individual cells across multiple independent biological replicates and sequencing depths, then assess per-cell reproducibility (e.g., technical replicate correlation, detection sensitivity at varying transcript abundance, and dropout rates) to determine whether the weaker single-cell correlation (r^2=0.58) reflects technical noise/capture efficiency limitations of scGRO-seq or true biological variability in nascent transcription between assays.
Inspect the scoring breakdown
Composite = 65% lexical coverage + 25% configured lexical-check score + 10% completeness. Lexical-check score: 100.0; completeness: 100.0. These are rubric components, not biological accuracy.
| Response field | Coverage /100 | Unmatched reference concepts |
|---|---|---|
| experiment | 50.0 | scGRO-seq profiles nascent transcription in individually barcoded nuclei |
| observations | 33.3 | strong aggregate correlation for gene bodies and enhancers; genomic-region composition differs between scGRO-seq and conventional scRNA-seq |
| controls and statistics | 50.0 | correlation does not establish unbiased measurement or equivalent sensitivity |
| supported conclusion | 100.0 | None |
| unsupported claim | 0.0 | error-free or quantitatively identical measurements; superiority to scRNA-seq for all biological questions |
| follow up | 50.0 | detection sensitivity and false-negative rates |
Configured claim/contradiction triggers: None.
Read the supplied context and reference
The figure introduces single-cell global run-on sequencing (scGRO-seq), which captures nascent transcription in isolated nuclei, and benchmarks aggregate and single-cell measurements against orthogonal assays.
The panels show the scGRO-seq workflow, representative gene and enhancer tracks, reads per cell, correlations with aggregate inAGTuC and intron seqFISH measurements, and the distribution of reads across genomic regions. Reported correlations include r-squared values of approximately 0.93 and 0.99 for aggregate gene-body and enhancer comparisons, and 0.58 for the single-cell seqFISH comparison.
Reference conclusion: The figure supports that scGRO-seq captures nascent transcription with strong aggregate agreement to orthogonal run-on data and moderate single-cell agreement to intron imaging measurements.
Model task: Explain what is being benchmarked and what the correlation panels support. Distinguish agreement from identity or superiority, note an important limitation of interpreting correlations from this figure, and propose a follow-up that tests reproducibility or sensitivity at the single-cell level.
Original figure ↗Source paper ↗
Figure 1 from Mahat et al. (2024), Nature, CC BY 4.0. License: Creative Commons Attribution 4.0 (CC BY 4.0).
What does an in vivo single-cell CRISPR screen validate?
Automated composite 62.9 · Lexical coverage 42.9 · Provisional review 80.0 · Configured triggers 0
- What worked: Accurately reconstructs the pooled in vivo screen, guide capture, cell annotation, and limits on gene-specific tumor claims.
- Scientific issue: Calls populations 'Cas9-edited' without direct per-guide validation and implies persistence of the same clones from P4 to P60.
- Scoring limitation: Scorer undercredits the answer but also fails to detect these subtler evidence-to-claim overextensions.
Read the structured model response
experiment
A pooled in vivo CRISPR screen was performed by in utero microinjection of a lentiviral sgRNA library (targeting 150 recurrently altered HNSCC/cSCC cancer genes plus non-targeting control guides) into Cas9-expressing, mCherry-reporter embryos at E9.5. Edited epithelial progenitor clones were tracked at postnatal day 4 (P4) and day 60 (P60) by mCherry fluorescence, FACS-enriched, and profiled by single-cell RNA-seq with simultaneous single-cell sgRNA (guide) capture, allowing each cell's transcriptomic state and perturbation identity to be linked. Cell types were annotated by UMAP clustering and canonical marker gene expression across the epidermal and hair follicle lineage.
observations
- Diagram (a) shows the sgRNA/mCherry lentiviral construct, in utero injection into Cas9 embryos, and downstream FACS enrichment plus scRNA-seq/sgRNA capture pipeline generating UMAPs at P4 and P60.
- Panel (b) shows a Venn diagram of gene overlap between HNSCC and cSCC mutation sets used to build the 150-gene library, plus a schematic of epidermal clones expanding differently for control vs. two test sgRNAs from E9.5 through P60.
- Panel (c) shows immunofluorescence images at P4 and P60 with mCherry, CD45, and DAPI staining, showing mCherry+ clonal patches within the tissue that are distinguishable from CD45+ immune cells.
- Panel (d) is a schematic of skin/hair follicle anatomy indicating locations of epidermal stem cells, bulge stem cells, sebaceous gland, dermal papilla, and various immune/pigment cell types.
- Panel (e) UMAP shows clustering of single cells into epidermal stem cell, suprabasal, lower bulge, bulge stem cell, sebaceous gland, UHF, macrophage/dendritic cell, and T cell populations, with a separate UMAP colored by cell-cycle phase (G1, S, G2/M).
- Panel (f) UMAP shows cells colored by whether an sgRNA was detected or not, with sgRNA-detected cells distributed across most major clusters including epidermal stem cells, suprabasal, bulge stem cells, sebaceous glands, and UHF, but sparse in T cell and macrophage/DC clusters.
- Panel (g) shows UMAP density plots for canonical marker genes (Krt14, Krt10, Mki67, Krt79, Gli1, Cd34, Cd3e, Cd74) confirming expected localized expression patterns corresponding to epidermal stem cells, suprabasal cells, cycling cells, UHF, lower bulge, bulge stem cells, T cells, and macrophages/DCs respectively.
controls and statistics
- Non-targeting control sgRNAs (50 guides) were included alongside the 150 gene-targeting guides to serve as a baseline for comparison of clone behavior (schematized in panel b as 'sgControl').
- Cell-type identity was validated using established marker genes (Krt14, Krt10, Mki67, Krt79, Gli1, Cd34, Cd3e, Cd74) rather than clustering alone, providing an internal biological control for UMAP cluster annotation.
- No quantitative statistics (e.g., clone size measurements, guide enrichment/depletion scores, p-values, or replicate numbers) are shown in this figure; panel (f) is presented as a qualitative UMAP overlay of sgRNA detection rather than a quantified capture-efficiency metric across cell types.
- No explicit sample size (number of embryos, mice, or cells per genotype) is stated in the visible legend excerpt.
supported conclusion
The figure demonstrates that the in utero CRISPR delivery and mCherry-based enrichment system successfully generates clonal, Cas9-edited epidermal populations that persist from P4 to P60, and that single-cell RNA-seq combined with guide capture can assign sgRNA identity to transcriptomically defined skin cell types (epidermal stem cells, suprabasal, bulge, sebaceous gland, and hair follicle populations), validating the technical feasibility of linking genetic perturbation to cell state in vivo.
unsupported claim
The figure does not show that any individual targeted gene causes altered clone growth, stem cell expansion, or tumor/neoplasia formation; it does not provide comparative clone-size or proliferation statistics between gene-targeting and control guides, nor does it demonstrate oncogenic transformation, malignant progression, or tumor formation for any specific gene from this dataset alone.
follow up
Perform quantitative comparison of clone size, proliferation (e.g., Ki67+ fraction or cell-cycle distribution), and relative sgRNA representation (enrichment/depletion) between gene-targeting and non-targeting control guides across matched P4 and P60 time points, ideally with statistical testing across multiple independent embryos/mice, to determine which specific gene perturbations significantly alter clonal expansion or progression toward a preneoplastic or tumorigenic state.
Inspect the scoring breakdown
Composite = 65% lexical coverage + 25% configured lexical-check score + 10% completeness. Lexical-check score: 100.0; completeness: 100.0. These are rubric components, not biological accuracy.
| Response field | Coverage /100 | Unmatched reference concepts |
|---|---|---|
| experiment | 66.7 | joint transcriptome and guide identity |
| observations | 33.3 | mCherry-positive epidermal clones at postnatal time points; guides detected across major annotated populations |
| controls and statistics | 66.7 | editing efficiency is not established for every guide |
| supported conclusion | 100.0 | None |
| unsupported claim | 0.0 | every targeted gene is edited equally; each enriched clone will progress to cancer through TNF |
| follow up | 0.0 | validate representative sgRNAs individually; independent guides and biological replicates |
Configured claim/contradiction triggers: None.
Read the supplied context and reference
The study combines in utero delivery of a pooled CRISPR library with mCherry enrichment, single-cell RNA sequencing, and guide capture to follow perturbed epidermal clones at postnatal time points.
The figure shows a library targeting 150 recurrently altered cancer genes plus control guides, in utero delivery to Cas9-positive embryos, P4 and P60 sampling, mCherry-positive clone imaging, cell-type annotation by UMAP and markers, and guide detection across major cell populations.
Reference conclusion: The figure validates the overall feasibility of coupling pooled in vivo CRISPR perturbation, epidermal clone recovery, guide capture, and single-cell cell-type annotation.
Model task: Describe the experimental design and the validation evidence in the figure. Explain what the control guides and cell-type marker panels contribute. State what cannot yet be concluded about individual genes or tumor formation from this figure alone, and propose the strongest next validation step.
Original figure ↗Source paper ↗
Figure 1 from Renz et al. (2024), Nature, CC BY 4.0. License: Creative Commons Attribution 4.0 (CC BY 4.0).
Calibrating efficacy and safety claims from a mouse dose-response figure
Automated composite 58.2 · Lexical coverage 35.7 · Provisional review 91.0 · Configured triggers 0
- What worked: Correctly interprets dose-dependent efficacy, delayed gross toxicity, the fitted TI, and the limits of clinical extrapolation.
- Scientific issue: Treats the error-bar definition as unknown despite the supplied legend and does not explicitly request PK and organ-exposure measurements.
- Scoring limitation: Severe lexical false-negative: nearly every expected observation is present in accurate paraphrase.
Read the structured model response
experiment
4T1-luc syngeneic mouse model of breast-cancer lung metastasis treated with escalating doses of a gemcitabine-paclitaxel drug-combination nanoparticle (GT DcNP), assessed by whole-body bioluminescence imaging, ex vivo lung nodule imaging, and body-weight tracking across a dose range from 0.125/0.0125 to 20/2 mg/kg (G/T).
observations
- Whole-body bioluminescence signal decreases progressively with increasing dose, becoming nearly undetectable at 10/1 and 20/2 mg/kg compared to saline.
- Ex vivo lung images show a parallel reduction in nodule number and signal intensity with increasing dose, with the fewest/faintest nodules at the highest doses.
- The fitted dose-response curves for %inhibition (nodule reduction and intensity reduction) show a sigmoidal increase, plateauing near 90-100% inhibition at doses ≥10/1 mg/kg.
- Body weight loss remains near zero or slightly negative (weight gain) at low-to-mid doses but rises sharply above ~15-16 mg/kg, reaching ~18-20% loss at 20/2 mg/kg.
- The efficacy curve (ED50 ≈2.306 mg/kg) is shifted well to the left of the toxicity curve (TD50 ≈36.48 mg/kg), giving a calculated separation (TI) of ≈15.8.
controls and statistics
- Saline-treated animals serve as the vehicle control for both imaging and weight-loss measurements.
- Group sizes are stated as 8-15 animals per dose, but no per-dose n, statistical test, or p-values are reported for the inhibition or weight-loss comparisons.
- ED50 and TD50 were derived from nonlinear regression (GraphPad Prism) fits to the inhibition and weight-loss curves, not from a predefined toxicological or survival endpoint.
- Error bars on the dose-response points appear to be SD or SEM but the specific measure is not defined in the excerpt.
- Body-weight loss is used as the sole surrogate for toxicity; no additional safety measures (organ histology, hematology, mortality) are described in this panel.
supported conclusion
In this syngeneic 4T1-luc lung metastasis model, the GT DcNP produces a clear dose-dependent reduction in both bioluminescent tumor signal and lung nodule burden, with substantial efficacy (~70-90% inhibition) achieved at doses well below those causing significant body-weight loss, indicating a wide preclinical margin between antimetastatic activity and this specific measure of gross toxicity in mice.
unsupported claim
That the calculated therapeutic index of ~15.8 demonstrates the drug combination will have an equivalently wide and safe therapeutic window in human patients.
follow up
Confirm toxicity with additional endpoints beyond body-weight loss (e.g., hematologic, hepatic, renal, and histopathological assessment) and test the formulation in an independent tumor model or in larger/more clinically relevant species before extrapolating the therapeutic index to human dosing.
Inspect the scoring breakdown
Composite = 65% lexical coverage + 25% configured lexical-check score + 10% completeness. Lexical-check score: 100.0; completeness: 100.0. These are rubric components, not biological accuracy.
| Response field | Coverage /100 | Unmatched reference concepts |
|---|---|---|
| experiment | 66.7 | tumor burden and body-weight loss dose responses |
| observations | 0.0 | increasing dose lowers bioluminescent tumor burden and lung nodules; body-weight loss occurs at higher doses; TD50 to ED50 ratio gives a therapeutic index near 15.8 |
| controls and statistics | 0.0 | saline-treated untreated disease control; 8 to 15 animals with mean plus or minus SEM; body-weight loss is a nonspecific gross toxicity proxy |
| supported conclusion | 100.0 | None |
| unsupported claim | 100.0 | None |
| follow up | 0.0 | repeat-dose pharmacokinetic and toxicology studies; clinical chemistry histopathology and organ exposure |
Configured claim/contradiction triggers: None.
Read the supplied context and reference
The figure evaluates a gemcitabine-paclitaxel drug-combination nanoparticle (GT DcNP) in a 4T1-luc syngeneic mouse model of breast-cancer lung metastasis and compares efficacy-related readouts with body-weight loss.
Representative bioluminescence and lung-nodule images are shown across doses. Dose-response curves quantify metastasis inhibition using integrated bioluminescence and nodule counts, while body-weight reduction is used as a gross toxicity measure. Group sizes are 8-15 animals; ED50 and TD50 estimates are fitted in GraphPad Prism, producing an estimated therapeutic index of 15.8.
Reference conclusion: In this 4T1-luc mouse model and dosing experiment, GT DcNP shows dose-responsive antimetastatic activity with body-weight loss emerging at higher doses, suggesting a preclinical efficacy-to-gross-toxicity separation.
Model task: Interpret the efficacy and toxicity trends and explain how the therapeutic index was estimated. Identify the control and the limitations of body-weight loss as a safety endpoint. State the strongest conclusion supported for this model, one clinical overclaim, and a follow-up needed before translational claims.
Original figure ↗Source paper ↗
Figure 2 from Mu et al. (2020), PLOS ONE, CC BY. License: Creative Commons Attribution (CC BY).
Interpreting motor-neuron axon-guidance phenotypes in lron-11 mutants
Automated composite 56.7 · Lexical coverage 33.3 · Provisional review 60.0 · Configured triggers 0
- What worked: Correctly identifies DD/VD versus DA/DB reporters, broad endpoints, mechanistic limits, and a class-specific rescue.
- Scientific issue: Misreads panels B and C, adds unsupported varicosities in E, and does not cleanly map wild-type versus mutant phenotypes.
- Scoring limitation: Automated score is coincidentally similar but misses the key panel-level contradictions and morphology hallucinations.
Read the structured model response
experiment
Fluorescent reporter lines labeling DD/VD GABAergic motor neurons (panels A-E) and DA/DB cholinergic motor neurons (panels F-G) were imaged in wild-type versus lron-11 mutant C. elegans to assess ventral nerve cord (VNC) tract placement, commissural laterality, and dorsal nerve cord (DNC) continuity.
observations
- Panel A: cell bodies lie along a single VNC tract with neurite processes labeled 'left' and 'right'; arrows indicate short lateral projections or turns near the midline.
- Panel B: a neurite (arrowhead) runs parallel to and slightly offset from the main VNC tract for an extended distance before rejoining it.
- Panel C: an arrow marks a short branch or misrouted process projecting away from the main VNC tract near a cell body.
- Panel D: a continuous, evenly spaced DNC process runs the full length of the imaged field with a small hook-like process crossing near the middle.
- Panel E: two arrowheads mark points along the DNC where the fluorescent process appears discontinuous or interrupted, with visible varicosities/puncta.
- Panel F: DA/DB neuron cell bodies (labeled DB3, DA2, DA3, DB4) sit along the VNC with commissures projecting dorsally at expected positions.
- Panel G: an arrow marks a commissure near DB3/DA2 and a second arrow marks a commissure near DB4, both appearing to project on a side or trajectory differing from panel F.
controls and statistics
- No explicit wild-type vs mutant genotype labels are visible on the panels themselves; the legend excerpt states that some panels represent wild type and others mutant, but the figure image alone does not identify which specific panel is which genotype.
- Scale bars (20 micrometres) are provided for each panel as a spatial reference, but no quantification, sample size (n), or statistical test is shown in the image.
- The legend states these are DD/VD and DA/DB motor-neuron markers, implying use of neuron-subtype-specific promoters, but the specific transgene/promoter identity is not given in the excerpt.
- No co-labeling control (e.g., a pan-neuronal or positional marker) is shown to independently confirm left/right identity of the VNC tract in each image.
supported conclusion
The images show variability in VNC tract occupancy (left vs right), occasional aberrant or extended neurite trajectories near the midline, and instances of apparent DNC discontinuity and commissures with atypical positioning in some samples, consistent with the legend's description of guidance/continuity defects being present in at least some animals of the imaged genotype(s).
unsupported claim
That these phenotypes are caused specifically by loss of LRON-11 protein function acting autonomously within these motor neurons to control axon guidance, rather than reflecting genetic background effects, marker-specific artifacts, or non-autonomous influences from surrounding tissue.
follow up
Perform neuron-class-specific rescue by expressing wild-type lron-11 cDNA under a DD/VD-specific or DA/DB-specific promoter in the lron-11 mutant background, then quantify the frequency of left-tract occupancy, incorrect-side commissures, and DNC gaps in rescued versus non-rescued mutant siblings across a defined sample size, to test whether restoring LRON-11 in the motor neurons themselves (cell-autonomous) is sufficient to correct the phenotypes.
Inspect the scoring breakdown
Composite = 65% lexical coverage + 25% configured lexical-check score + 10% completeness. Lexical-check score: 100.0; completeness: 100.0. These are rubric components, not biological accuracy.
| Response field | Coverage /100 | Unmatched reference concepts |
|---|---|---|
| experiment | 100.0 | None |
| observations | 20.0 | wild-type DD VD neurites in the right ventral tract; mutant DD VD neurites in the left ventral tract; DD VD commissure extends on the incorrect side; dorsal nerve cord gap in the mutant |
| controls and statistics | 0.0 | wild type with the same neuronal reporter is the morphological control; representative images do not establish penetrance or statistical significance |
| supported conclusion | 0.0 | loss of lron-11 is associated with motor-neuron axon-navigation defects; ventral tract placement commissural laterality or dorsal nerve cord continuity is disrupted |
| unsupported claim | 50.0 | directly senses a specific extracellular guidance cue |
| follow up | 50.0 | expression-matched controls and quantified rescue |
Configured claim/contradiction triggers: None.
Read the supplied context and reference
The figure tests whether loss of the extracellular leucine-rich-repeat protein LRON-11 affects ventral nerve cord tract placement, commissural laterality, and dorsal nerve cord continuity in C. elegans motor neurons.
Panels A-C use DD/VD motor-neuron markers: wild type shows neurites in the right ventral nerve cord tract, whereas mutants can show neurites in the left tract or a commissure on the incorrect side. Panels D-E compare DD/VD dorsal nerve cord continuity and show a mutant gap. Panels F-G use a DA/DB motor-neuron marker and show mutant commissures extending on the incorrect side. Anterior is left; scale bars are 20 micrometres.
Reference conclusion: Loss of lron-11 is associated with defects in motor-neuron ventral tract placement, commissural laterality, and dorsal nerve cord continuity.
Model task: Identify the reporter and phenotype shown in each panel group (A-C, D-E, and F-G). Compare wild-type and mutant morphology, identify the control, and distinguish direct observations from mechanistic claims. State the strongest supported conclusion, one overinterpretation, and a neuron-class-specific rescue experiment that would test cell autonomy.
Original figure ↗Source paper ↗
Figure 4 from Kokan et al. (2022), PLOS ONE, CC BY. License: Creative Commons Attribution (CC BY).
What these results support
Automated scoring
The composite weights lexical concept coverage at 65%, configured claim/contradiction checks at 25%, and response completeness at 10%. The historical claim_safety data field names the lexical-check component.
It can miss paraphrases, panel-level errors, and unsupported biological interpretations. It is not a biological-accuracy measure.
Provisional scientific review
The historical review was conducted by one reviewer without blinding. Published records contain totals and narrative notes; they do not include completed dimension-level scoring sheets.
A small pilot and one baseline do not support broad model rankings or general performance claims. Full methods and provenance ↗
Reproduce and inspect
The saved responses can be validated, re-scored, and rendered without an API key or new model call. Follow the offline quickstart ↗.
Expert notes (CSV)Public metrics (JSON)Frozen responses and scores