BGPT: Author Review: Jerry M. Parks

Explore by Goal

Quick Explanation Copied

Jerry M. Parks — scientific strength (evidence-based critique)

Across the provided paper set, the author’s work spans enzyme mechanisms/structure, advanced biophysics (neutron crystallography), and computational biology/ML for biomolecular modeling (e.g., RNA-ligand docking pose reranking) .

Evidence quality is moderate-to-strong where papers include explicit datasets, evaluation design, and mechanistic/structural measurement —but the prompt provides insufficient details to independently verify all other listed works’ methods/results.

Long Explanation

BGPT Author Review (Critical, Skeptical, Evidence-Based): Jerry M. Parks

Scope note: This review is constrained to the information explicitly provided in your prompt (paper titles list + one detailed raw-data-like study summary for IRIS + OpenAlex author-metrics snapshot). Where details are missing, I mark uncertainty instead of filling gaps.

1) Citation/impact signals (from the metrics you supplied)

Prompt-provided (local): total citations 44; h-index 2; paper count 6.
OpenAlex snapshot (from your prompt): cited_by_count 5923; h_index 42; works_count 181 for the top match (OpenAlex ID A5037417242).

Skeptical interpretation: bibliometrics are signals, not mechanistic proof. High citation counts can reflect broad relevance, but also field size, network effects, and selective visibility; low values can reflect misattribution, name disambiguation, or short time windows.

2) Evidence-rich deep-dive: IRIS (ML reranking for RNA/DNA-ligand docking)

The most detailed, audit-able component in your prompt is a study summary describing IRIS trained on 734 nucleic acid–ligand complexes (608 train / 126 test) and evaluated via Top-1/Top-5 pose ranking, including interface-level RMSD analysis and SHAP-style feature attribution .

Figure A — Normalized pose-ranking gains (RNA vs DNA)

What this supports: the reported evaluation indicates stronger gains for RNA than DNA in the provided summary .

Key uncertainty: this is not a stand-alone pose refinement method; it reranks poses produced by rDock. Thus, any baseline limitations or pose diversity issues in the upstream docking step can cap performance .

Figure B — Interface-level accuracy (I-RMSD reported)

The summary reports that interface-level accuracy improvements (I-RMSD) are particularly strong for RNA .

Figure C — SHAP feature attributions (magnitude, from summary)

In the provided summary, the largest SHAP attributions are reported for rDock-derived interaction/restriction terms (e.g., SCORE.INTER.norm, SCORE.RESTR) .

3) Scientific strength signals across the provided work set (title-level evidence + one full study)

Mechanistic/structural biochemistry & biophysics: Titles indicate enzyme characterization and neutron crystallography for hydrogen/protonation insights (high evidentiary value when experimentally executed). Example provided DOIs include neutron work on hydrogen bonding/protonation in glycoside hydrolases and active-site hydrogen observations in tryptophan synthase .
Constraint: the prompt did not provide the full bibliographic metadata for these neutron papers (DOIs beyond what’s in your list are not fully validated in this chat), so I cannot fully audit experimental specifics.
Protein engineering / biosensing: A provided title indicates split-intein design for a small-molecule biosensor in plants—an engineering-heavy program where success depends on signal dynamic range, specificity, stability, and expression context (details not provided here).
RNA/DNA targeting via ML + structural data: The IRIS summary suggests fairly rigorous dataset curation and explicit evaluation framing (Top-1/Top-5, interface RMSD, cross-validation, SHAP). That combination is a positive sign for scientific rigor in computational work .

4) Rigor & limitations: what could mislead?

IRIS-specific blind spots (from the provided summary)

Upstream dependence: IRIS reranks rDock-generated poses, so performance is bounded by rDock’s ability to generate near-native conformations. This is a structural limitation, not a purely ML limitation .
Data-limited generalization: The summary notes chemical diversity/size limitations and a persistent train/validation gap, which can indicate a data-limited regime (i.e., overfitting risk or unmet coverage) .
Partition/redundancy choices: Redundancy filtering by receptor identity and ligand SMILES equivalence reduces exact overlap but may not remove all near-duplicate geometry/chemotype patterns; model performance may reflect these specific dataset priors .

5) Practical assessment (fit-for-purpose)

Best-supported claim from this prompt: Parks’ work (at least in IRIS context) demonstrates an ML pipeline trained and tested on experimentally anchored nucleic-acid–ligand complex sets, with explicit evaluation and attribution, improving docking pose ranking for RNA more than DNA .
Lower-confidence region: the prompt lists other Parks works (nylon hydrolases; split-intein plant biosensor; neutron diffraction papers; etc.), but provides insufficient methodological detail here to independently score rigor for those papers.

6) Buttons for deeper BGPT exploration

Feedback:

Updated: March 25, 2026