What Makes a Good Protein Structure? A Data-Driven Look at 45,000 Models

By shrijeetverma13 · June 29, 2026

This analysis examines nearly 46,000 protein structures to understand what predicts quality. We found that most individual features are weak predictors of…

The analysis shows virtually no relationship between F8 (count metric) and average RMSD. With a Pearson correlation of exactly 0.000, changes in F8 values do not predict or explain changes in RMSD. A line chart was generated showing average RMSD across 20 binned ranges of F8, making it easy to see that the trend is essentially flat overall.

The analysis successfully produced visualizations showing how RMSD values distribute across three quality bands: High Quality (RMSD ≤ 5), Medium Quality (5 < RMSD ≤ 10), and Low Quality (RMSD 10). A bar chart displays the count of structures in each band, and a histogram shows the full RMSD value distribution colored by quality band.

Among the nine features (F1–F9), F3 shows the strongest linear correlation with RMSD (r = 0.374), making it the most influential predictor. Most other features have relatively weak correlations, suggesting that RMSD prediction may involve non-linear relationships. Two bar charts and correlation tables were generated to visualize these relationships clearly.

There are 13 feature pairs with correlations of r ≥ 0.8, indicating significant redundancy among your features. A heatmap and bar chart have been generated to visualize these relationships. The strongest correlation is between F1 and F5 (r = 0.998), meaning they are nearly identical in information content. Features F1, F4, F5, F6, and F9 are the most involved in redundant pairs, with F1 appearing most frequently.

The analysis identified 1,300 records (out of 45,730 total) that show extreme F5 or F7 feature values yet have surprisingly low RMSD scores. These anomalies were flagged using a 90th percentile threshold for F5 (≥ 2,150,446) or F7 (≥ 5,369), combined with RMSD in the bottom 25% (≤ 2.31). The most extreme example is row 39194, with an RMSD of exactly 0.000 despite F5 reaching 3,734,152 and F7 at 8,165 — a striking contradiction. Two scatter plots were generated highlighting these anomalous records in red against the normal data in gray, for both F5 vs RMSD and F7 vs RMSD.

The analysis compared feature profiles between high-quality structures (RMSD < 2) and others across 9 features (F1–F9). Two visualizations were generated: a grouped bar chart showing normalized mean feature values for both groups side by side, and a horizontal bar chart highlighting the percentage difference in mean feature values between high-quality and other structures.

The IQR method was applied to F1, F2, and F5 independently. All three features show a similar proportion of outliers, ranging from 2.35% to 2.73% of rows. A side-by-side box plot was generated showing the distribution and outliers for all three features. The cross-tabulation between F5 outliers and F1/F2 outliers was partially computed, with the statistical summaries successfully produced for each feature.

The full correlation matrix for features F1 through F9 has been computed and visualized as an annotated heatmap. The heatmap uses a red-blue color scale where deep blue indicates strong positive correlations and deep red indicates strong negative ones, with each cell labeled with its exact coefficient. The analysis reveals significant multicollinearity concerns across the feature set.

The analysis successfully binned RMSD values into 5 quantile-based tiers (Very Low, Low, Medium, High, Very High) and computed the mean and standard deviation of F2, F4, and F7 for each tier. Two grouped bar charts were generated: one showing the raw mean values with error bars (standard deviation), and a second normalized (0–1 scale) chart for easy cross-feature comparison. These visualizations clearly reveal how F2, F4, and F7 change systematically as RMSD quality levels shift from best (Very Low) to worst (Very High).

Here's a breakdown of the distributions for F3, F4, and F6 based on their computed statistics:

F3 is the most well-behaved of the three — it's approximately normally distributed with a mean of 0.30, median of 0.30, and a tight standard deviation of 0.063. Its skewness (0.24) and excess kurtosis (0.09) are both near zero, and it shows a unimodal pattern.

F4 is right-skewed (skewness = 1.23), meaning a long tail stretches to the right. The mean (103.5) is noticeably higher than the median (87.7), confirming the pull from high-value outliers. It also has heavy tails (excess kurtosis = 1.33) and a possible multimodal pattern — suggesting the data may come from more than one underlying group.

F6 behaves similarly to F4 — right-skewed (skewness = 1.12) with a mean of 145.6 versus a median of 126.2. It also shows heavy tails (excess kurtosis = 1.08) and a possible multimodal pattern, indicating potential subpopulations in the data.

In summary: F3 is clean and normal; F4 and F6 are skewed, heavy-tailed, and potentially multimodal — worth investigating further for subgroups or outliers.