Review papers with raw data transparency

1 Key claims and evidence

• Claim GDV linear model predicts atom-level RMSF directly from coordinates with high correlation to MD-derived RMSF for globular proteins — supported by cross-validation on ATLAS (mean Spearman ~0.792–0.794 across three replicates; best-select mean 0.828). ATLAS cross validation stats
• Claim Generalization to experimental benchmarks: NMR median/mean correlations ~0.729/0.776 (140 proteins) and cryo-EM mean correlation ~0.704 (321 proteins after QC); performance is slightly below specialized neural models on cryo-EM that use experimental density. NMR and cryo-EM benchmarks
• Claim The approach is extremely fast and memory-efficient with a by-parts scheme enabling predictions on very large proteins on laptops (<1 GB RAM segments; ~6 s vs 12 s all-at-once for 6SUP). Performance and memory benchmarks

2 Methods evaluated

The pipeline is straightforward, reproducible, and lightweight:

1. Represent atoms as nodes and add edges for pairs within cutoff (7 Å default) to build an atom contact graph.
2. Compute Graphlet Degree Vectors (GDV) using graphlets up to size four, producing 15 orbit counts per atom.
3. Log transform and per-protein normalize GDV features and RMSF targets; train ordinary least squares multiple linear regression (15 features) with 10-fold cross-validation across three independent ATLAS MD replicates.
4. Produce per-atom RMSF predictions and evaluate Spearman correlation at residue Cα level. Methods summary

3 Strengths

• Parsimony: 15 interpretable GDV features with explicit linear coefficients (equation provided) — interpretable mapping between local packing topology and flexibility. Regression coefficients
• Speed and accessibility: runs on commodity hardware; by-parts scheme makes very large proteins tractable (practical for rapid exploratory uses and integration into pipelines). Hardware and runtime
• Reproducibility intent: code and scripts released (GitHub repo FastProtFlex) and cryo-EM MD data accessible via Figshare. Code and data links

4 Limitations, blind spots, and risks of over-claiming

• Training bias — model trained on a filtered ATLAS globular-like subset (N=1052) using TM-score and radius of gyration cutoffs; non-globular topologies (rod-like, extended, membrane proteins) are underrepresented and show failure cases (example 4KE2 had correlation −0.236). This restricts distributional generality. Training selection and outliers
• Limited long-range coupling — GDV captures local graph topology (up to 4-node graphlets) and therefore misses long-range or collective inter-domain motions that dominate flexibility in many multi-domain assemblies. The authors demonstrate this (1C96) where MD shows domain contrast that GDV underestimates. Inter-domain limitation
• Input model sensitivity — GDV uses raw coordinates; modeled/misplaced residues (MODELLER additions) produce artefactual local contact density and degraded predictions (example 1BY2). Hence, garbage-in→garbage-out for poor structural models. Sensitivity to model errors
• Comparison scope — comparisons to other methods (CABS-flex, RMSF-net) are informative but not exhaustive: different tools use different inputs (pLDDT, density maps) and different output normalization; side-by-side equivalence is therefore imperfect. Method comparisons caveat

5 Reproducibility and resources

The authors provide R scripts and a GitHub repo (FastProtFlex) with FUNCTION_GDV.r and predict.r and example PDB input (usage documented). Cryo-EM MD RMSF files are archived on Figshare which the authors used as an external benchmark. These materials materially support reproducibility; running the pipeline requires only R and commonly available packages. Code and dataset availability

6 Recommendations to improve the model and the manuscript

1. Explicitly quantify out-of-distribution performance: test on membrane proteins, intrinsically disordered proteins, and extended rod-like proteins (report distributions, not only curated benchmarks).
2. Augment GDV features with sparse long-range descriptors (e.g., path-length-based graphlets, global centralities, or low-rank diffusion distances) or add a simple second-stage model (e.g., gradient-boosted trees on summary global features) to capture inter-domain coupling while retaining speed.
3. Provide per-protein calibration plots (predicted vs MD RMSF quantiles) and error distributions by secondary structure and solvent accessibility to show where the model systematically under/over-predicts.
4. Release precomputed GDV matrices for ATLAS training subset to accelerate independent replication and downstream method development.

Each recommendation addresses a clear blind spot: distributional shift, long-range coupling, systematic bias analysis, and reproducibility throughput.

7 Bottom line

The GDV linear model is an elegant, fast, and interpretable approach that credibly predicts per-atom flexibility for globular proteins using only coordinates and graph topology, achieving competitive Spearman correlations versus MD and established methods while running in near real time on standard hardware; however, its limitations are concrete and documented: sensitivity to non-globular topology, inter-domain motions, and model input quality. It is a valuable addition to the toolbox (rapid screening, annotation pipelines, and educational uses), but it should not yet replace physics-based or density-informed methods when long-range collective motions or experimental data are essential. Paper conclusion