QSAR-derived affinity fingerprints (part 2): modeling performance for potency prediction

Area of Science:

• Computational chemistry
• Cheminformatics
• Drug discovery

Background:

• Affinity fingerprints capture small molecule activity across assays, aiding in understanding bioactivities of diverse compounds.
• Traditional structure-based models have limitations for complex biological endpoints like toxicity and cancer cell line sensitivity.

Purpose of the Study:

• To propose and validate a framework for calculating QSAR-derived affinity fingerprints (QAFFP) as compound descriptors.
• To model in vitro compound activity using computationally predicted bioactivity profiles.
• To evaluate the predictive power of QAFFP for various biological endpoints.

Main Methods:

• A framework for calculating QAFFP was developed using 1360 QSAR models from ChEMBL database (Ki, Kd, IC50, EC50 data).
• QAFFP were benchmarked using IC50 data for 18 cancer cell lines and 25 protein targets.
• The study complements prior work evaluating QAFFP in similarity searching, scaffold hopping, and bioactivity classification.

Main Results:

• QAFFP as descriptors resulted in prediction errors in the ~0.65-0.95 pIC50 range, comparable to ChEMBL data uncertainty (0.76-1.00 pIC50).
• QAFFP showed slightly lower predictive power than Morgan2 fingerprints and physicochemical descriptors.
• Including low-predictive QSAR models did not enhance QAFFP predictive power.

Conclusions:

• QAFFP provide a method to encode and relate compounds based on bioactivity similarity.
• The predictive performance of QAFFP can be improved by using more diverse and biologically relevant QSAR models.
• Publicly available datasets and Python code facilitate further research and application of QAFFP.