Interpretation of Kcat prediction values in Enzeptional optimization output

[lucaspalmeira]

Description

When using Enzeptional for enzyme sequence optimization with the XGBoost model, the predicted Kcat values in the output include both low positive numbers (e.g., 0.499) and negative values (e.g., -0.874), which appear physically invalid for Kcat measurements.

From the framework documentation and previous papers, I understand that:

• The training process uses logarithmic transformation on Kcat data to improve linearity
• The scaler.pkl file is applied to reverse transformations during prediction
• However, it's unclear whether the final scores from SequenceScorer are already in the original scale or if manual conversion (e.g., 10^x) is required

Questions

1. Scale Interpretation: Are the Kcat values in the optimization output already converted back to the original (non-logarithmic) scale, or do they require additional transformation?

2. Negative Values: How should negative Kcat values be interpreted? Do they indicate:

   • Invalid sequences?
   • Pipeline errors?
   • Artifacts of the model prediction range?
   • Values below a certain detection threshold?

3. Expected Value Range: What is the expected valid range for Kcat predictions, and how should outliers or physically impossible values be handled?

Context

Example output values observed:

• Positive but low: 0.499, 0.123, 0.876
• Negative: -0.874, -1.234, -0.567

Additional Information

• Using Enzeptional via GT4SD library
• XGBoost model for Kcat prediction
• Following the example from: examples/enzeptional/example_enzeptional.py

Any clarification on the proper interpretation of these output values would be greatly appreciated for correct analysis of optimization results.