Why BGPT?
logo

Check Claims Before You Publish

See how your reported results read as versioned evidence records — linked to experiments and limitations.Know what the science actually supports before you trust the answer.

Press Enter ↵ to review claims



    Explore by Goal












     Quick Explanation


    BGPT note: The only specific “Shao Li” work product you provided is a set of paper excerpts plus some author-level citation metadata. From those excerpts alone, this appears to be a computational/biomedical modeling–heavy author profile spanning: (i) multi-modal ML for antimicrobial peptides, (ii) equivariant GNN potentials for molecular physics, (iii) immune cell-type annotation for BALF scRNA-seq, (iv) mechanistic immunology/fibrosis work in mice, and (v) mechanobiology/cell signaling studies—plus several review/synthesis pieces.



     Long Explanation


    Author Review (Science Strength): “Shao Li”

    Evidence basis used: Only the paper-excerpt research items you included (each with DOI and extracted quantitative claims). No external bibliometrics were verified beyond what you provided.

    Visual evidence map (where the excerpts are strongest)

    Core citations used for this visualization

    Quantitative performance snapshot (from provided extracted numbers)

    Cited bases for the plots
    Skeptical check: the ViSNet plot uses only the excerpt-provided subset (not the full benchmark table), to avoid inventing missing entries.

    Scientific strength assessment (based strictly on your excerpts)

    1) Quantitative modeling + explicit evaluation loops

    The strongest “signal” in these excerpts is explicit reported metrics under multiple regimes (e.g., in silico “test vs low-similarity subset” for AMP; multiple geometric/MD benchmarks for ViSNet; LOSO and external macro-F1 for BAL annotation). For example, MultiAMP reports AUROC/AP/MCC/F1/Accuracy on both a test split and a low-similarity subset, and shows ablations that reduce performance when key modules are removed ().

    2) Cross-domain breadth—good for transfer, risky for depth

    The excerpt set spans multiple biology and ML subfields: immune-fibrosis in mice (), synaptic mechanobiology with FRET biosensors (), and BAL scRNA-seq annotation standardization (). This breadth can indicate strong interdisciplinary computational competence, but it also increases the chance that some results are context-specific (e.g., model assumptions, cohort/geography effects, or preclinical species limitations).

    3) Methodological transparency signals (data/code availability in excerpts)

    Multiple excerpted works include code/data pointers (e.g., MultiAMP code on GitHub and ocean data sourcing description ; ViSNet code/data pointers via GitHub and benchmark dataset URLs ). Such transparency typically improves reproducibility prospects—though absence of full accession lists in some excerpts remains a concern.

    Critical skepticism: where the excerpted science could mislead

    A) In silico success ≠ wet-lab generality

    MultiAMP’s design claims in the excerpt are computational and appear to rely on in silico structure prediction and predicted MIC evaluation rather than comprehensive experimental MIC testing for all designed candidates (). A skeptic should demand (i) independent external AMP datasets and (ii) at least some wet-lab activity confirmation, especially for “ocean mining” candidates.

    B) Dataset bias and negative sampling risk

    In MultiAMP, negatives are sourced from UniProt with antimicrobial annotations excluded, and ocean candidates are filtered by identity thresholds (). Such procedures can systematically change the “easy vs hard” classification boundary; without rigorous external datasets and activity testing, generalization claims remain probabilistic.

    C) Preclinical causality and translational uncertainty

    TRIM13 CAR-M in mice reports multiple mechanistic readouts and fibrosis reversal in CCl4 and HFMCD models (). But the excerpt explicitly notes preclinical species limits and longer-term safety/off-target and head-to-head comparison gaps ().

    What would most strengthen the scientific case (falsification targets)

    • For MultiAMP: show that independently collected AMP datasets (not derived from the same filtering/identity-thresholding logic) preserve the superiority over strong baselines, and that motif/structure attributions correlate with experimentally measured activity (not just predicted MIC). This is consistent with the excerpt’s own limitations around dataset bias and lack of extensive external experimental validation ().
    • For ViSNet: quantify performance drift beyond benchmark regimes and demonstrate robust force/energy behavior for conformational edge cases; the excerpt acknowledges generalization/robustness boundaries such as benchmark dependence and invariance/representation tradeoffs ().
    • For BAL-EA: continuously add genuinely new BAL patient/disease contexts and re-test generalization, because atlas-driven annotation can lag behind novel rare states; the excerpt flags external accuracy variance for certain cell types and the dependence on available BAL-dominant biology ().

    Overall judgment from provided excerpts

    Based only on your excerpts, Shao Li’s scientific contributions look strongest where (i) architectures are explicitly evaluated with multiple regimes/ablations and (ii) computational pipelines are tied to quantitative metrics (MultiAMP, ViSNet, BAL-EA). The weaker scientific “weight” (for causal certainty) is where results are primarily in silico or preclinical without full experimental externalization, which the excerpts themselves acknowledge (MultiAMP, TRIM13, and also TuFEst’s cohort-generalizability caveats).


    Feedback:    

    Updated: March 23, 2026

     Top Data Sources ExportMCP


     Analysis Wizard



    Summarize excerpted quantitative metrics into comparative plots (AUROC/MCC/F1/MAE/force MAE/R2), then compute rank stability across subsets to expose sensitivity to evaluation regime shifts.



     Hypothesis Graveyard


    “All performance gains come purely from better biology” is unlikely: the excerpts explicitly note biases, dataset construction choices, and reliance on predicted structures/predicted MIC and atlas coverage limitations, which can explain gains without representing underlying biology perfectly.


    “Preclinical mechanistic chains guarantee human efficacy” is too strong: TRIM13’s excerpt emphasizes mouse-only evidence and incompletely assessed long-term safety/off-targets, so translational generality cannot be assumed.

     Science Art


    Author Review: Shao Li Science Art

     Science Movie


    Make a narrated HD Science movie for this answer ($32 per minute)




     Discussion


    Stay current without chasing every paper.

    Know what changed, what holds up, and what remains uncertain. Every Friday. No ads.


    My BGPT






     Trending