CliffRank Predicts Activity-Cliff Rankings with Dual-Branch AI Framework

Kewei Li, Rongying Zhang, Peiyu Yang, Zhongjian Wang, Qiuchen Zhao, Lan Huang, Fengfeng Zhou· September 3, 2026 View original

Key takeaways

  • CliffRank is a novel AI framework for predicting activity-cliff rankings in drug discovery.
  • It uses a dual-branch approach combining regression and ranking-consistency learning.
  • The framework demonstrated strong performance on both peptide and small-molecule datasets.
  • Future work includes adaptive schedules and incorporating biological context for further improvement.

Who benefits

PharmaceuticalsBiotechnologyChemical ManufacturingHealthcare

Summary

CliffRank is a new dual-branch AI framework that improves the prediction of activity-cliff rankings by combining absolute-activity regression with ranking-consistency learning. It achieved strong performance on antimicrobial peptide and small-molecule datasets, demonstrating its effectiveness in drug discovery applications.

This research introduces CliffRank, an innovative AI framework designed to enhance the prediction of activity-cliff rankings, a critical challenge in drug discovery. Activity cliffs occur when minor structural changes in molecules lead to significant shifts in their biological activity, making them difficult to model with traditional methods due to limited high-quality data. CliffRank addresses this by employing a dual-branch architecture that integrates absolute-activity regression with a novel ranking-consistency learning approach, including Pairwise Preference Consistency (PPC). The framework was rigorously tested on multiple datasets, including antimicrobial peptides and small molecules. On antimicrobial peptide datasets, CliffRank, particularly when integrated with ESM2-t12, showed superior performance in Spearman correlation and Recall@50 metrics. For small-molecule datasets, CliffRank with PNA also achieved high Spearman correlation, matching other leading methods in Recall@50. The study also explored the practical limits of PPC and the impact of initialization strategies, suggesting areas for future refinement like adaptive PPC schedules and incorporating contextual information.

Why it matters

Professionals in drug discovery and computational chemistry can leverage this framework to more accurately identify and understand activity cliffs, accelerating the design of new compounds with desired properties.

How to implement this in your domain

  1. 1Evaluate CliffRank's open-source implementation for specific drug discovery projects.
  2. 2Integrate the dual-branch learning approach into existing molecular modeling pipelines.
  3. 3Experiment with different pre-trained models (e.g., ESM2-t12, PNA) to optimize performance for target molecules.
  4. 4Develop adaptive Pairwise Preference Consistency (PPC) schedules tailored to specific datasets.
  5. 5Incorporate protein or membrane context into the model when available for improved accuracy.

Original post by Kewei Li, Rongying Zhang, Peiyu Yang, Zhongjian Wang, Qiuchen Zhao, Lan Huang, Fengfeng Zhou

"arXiv:2609.01673v1 Announce Type: new Abstract: Activity-cliff ranking remains difficult because local structural changes can cause large activity differences, while high-quality data that resolve the underlying mechanisms remain limited. To use available activity labels more eff…"

View on X

Originally posted by Kewei Li, Rongying Zhang, Peiyu Yang, Zhongjian Wang, Qiuchen Zhao, Lan Huang, Fengfeng Zhou on X · view source

Want to go deeper?

Turn these trends into skills with Learnijoy's hands-on AI & tech courses.

Explore courses