Step away from using pchembl values

[LindeSchoenmaker]

got a lot of feedback on the interpretability of pchembl value. It is possible to convert the predicted pchembl to the half maximal effect concentration. First step;

• ask feedback of people with cheminformatics background before implementing this

[Chris-Evelo]

With some help of ChatGPT I got this:

Understanding the Role and Limitations of pChEMBL in QSAR Models
For communication within the VHP4Safety project

What is pChEMBL?
The pChEMBL value is a standardized measure used in chemical biology databases (like ChEMBL) to represent compound potency. It converts various activity measurements — such as IC₅₀, EC₅₀, Kᵢ, or Kd — into a single negative log molar format. For example:

An IC₅₀ of 1 μM becomes a pChEMBL of 6

A Kd of 10 nM becomes a pChEMBL of 8

This transformation allows for easier comparison and integration across assays and datasets, which is why it’s commonly used as the endpoint in QSAR (Quantitative Structure–Activity Relationship) models.

How is it used in QSAR?
In QSAR models, the pChEMBL value often acts as:

A quantitative target variable to predict the potency of new compounds

A proxy for biological activity, especially in early-stage compound prioritization or virtual screening

A central value in building toxicity or risk evaluation models, such as predicting off-target interactions or potential hazard

Why pChEMBL ≠ Absolute Potency
Despite its usefulness, it’s important to recognize that the pChEMBL value doesn’t always represent true biological potency in a uniform way. That’s because:

Assay diversity:
pChEMBL values come from many types of assays — purified enzyme systems, cell-based assays, binding studies, or functional outputs — each with different biological relevance.

Conflated mechanisms:
The value may reflect direct binding, functional inhibition, or indirect effects (e.g. redox imbalance, metabolic interactions).

Experimental conditions vary:
Factors like enzyme concentration, redox state, co-factors, and even adsorption to labware (e.g., untreated glass surfaces) can shift apparent potency, particularly in low-concentration assays.

A Real-World Illustration (without overstating)
In earlier academic work, pChEMBL-like data derived from inhibition assays on glutathione S-transferase P1-1 (GSTP1) showed that apparent inhibition by some compounds could vary widely depending on:

Whether assays were cell-free or cell-based

Enzyme concentration

Presence of oxidative stress

Interaction with treated vs. untreated glassware

These variations, while not always visible in the summarized pChEMBL value, could strongly influence how that compound appears in a model.

Why This Matters for VHP4Safety
Projects like VHP4Safety aim to integrate chemical and biological data for better safety prediction. That means:

Awareness of uncertainty is essential when interpreting model results

pChEMBL is valuable, but must be contextualized — especially for regulatory or mechanistic interpretation

If possible, metadata about assay type and biological relevance should be included or flagged during model development and communication

Takeaway Message
pChEMBL is a powerful tool for QSAR modeling — offering a unified measure of potency across diverse datasets. But like any summary statistic, it comes with hidden layers of experimental and biological complexity. Being transparent about these limitations strengthens both the interpretability and credibility of predictive models in safety science.

Let me know if you'd like a slide version, shortened summary, or references to add!

You said:
I will share it with her on GitHub and see is she needs anything else. Thanks!

ChatGPT said:
Perfect — I hope it helps her communicate things clearly! If she needs a version tailored for a presentation, a diagram to go with it, or references to back up specific points, just let me know. Happy to help with any follow-up!