Source-linked AI summary
CliffRank: A Dual-Branch Framework for Activity-Cliff Ranking Prediction
Kewei Li, Rongying Zhang, Peiyu Yang, Zhongjian Wang, Qiuchen Zhao, Lan Huang, Fengfeng Zhou
TL;DR
Activity-cliff ranking is difficult because local structural changes can cause large activity differences and high-quality mechanistic data remain limited. CliffRank addresses this by combining absolute-activity regression with listwise ranking and Pairwise Preference Consistency across two predictors. It achieved the strongest cross-dataset ranking averages reported for the AMP and small-molecule tasks, while PPC effects remained setting-dependent.
Problem
Activity-cliff ranking remains difficult because local structural changes can cause large activity differences and high-quality data that resolve underlying mechanisms remain limited.
Method
CliffRank trains two parallel predictors with mean squared error, a thresholded listwise loss, and Pairwise Preference Consistency to align relative ordering in preference-probability space.
Results
CliffRank achieved the highest cross-dataset mean Spearman correlation and Recall@50 on AMP tasks, and higher mean Spearman correlation with Recall@50 matching ACANet-PNA on small-molecule tasks.
Takeaways & Limitations
The results support using ranking-consistency learning to extract more information from available activity labels, while its practical value depends on dataset, backbone, initialization, and PPC scheduling.
Takeaways & Limitations
No initialization or PPC activation strategy was best across all targets and metrics.
Abstract
from arXiv · showhide
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 effectively, we combine absolute-activity regression with ranking-consistency learning. CliffRank trains two parallel predictors with mean squared error, a thresholded listwise loss, and Pairwise Preference Consistency (PPC), which aligns relative ordering in the preference-probability space. On three antimicrobial peptide datasets, CliffRank with ESM2-t12 achieved the highest mean Spearman correlation of 0.5393 and mean Recall@50 of 21.4, although the leading method varied across individual datasets. On three small-molecule datasets, CliffRank with PNA, where PPC was activated after 120 epochs, achieved the highest mean Spearman correlation of 0.6890, while its mean Recall@50 of 30.4 matched that of ACANet-PNA. The PPC results also define its practical limits. Asymmetric initialization improved the MolCLR-GIN averages but did not improve every target. For PNA without pretrained weights, delayed PPC improved selected metrics, but no schedule was best for both mean Spearman correlation and mean Recall@50. Future work should evaluate more targets and antimicrobial peptide systems, develop adaptive PPC schedules, and incorporate protein or membrane context when available.
Introduction
Activity-cliff prediction is challenged by sparse, noisy, and mechanistically incomplete data, motivating methods that extract more ranking information from available activity labels. The proposed DualCliff framework combines dual predictors and ranking-consistency objectives, achieving stronger cross-dataset averages while showing dataset- and metric-dependent behavior.
- Activity cliffs arise when local structural changes produce large activity differences, but unseen MMP pairs remain difficult for QSAR models.
- Small-molecule activity-cliff prediction is constrained by training-set coverage, experimental error, structural-representation limits, noisy protein–ligand data, and possible split leakage.
- High-quality mechanistic data are difficult to obtain for AMP–membrane systems because realistic membranes require complex sampling and computational modeling.
- DualCliff uses the AMPCliff activity-cliff split and combines absolute-activity supervision with listwise ranking and Pairwise Preference Consistency, which constrains predicted order.
- Across AMP tasks, DualCliff with ESM2-t12 achieved the highest cross-dataset mean Spearman correlation and Recall@50, while small-molecule gains were concentrated in mean ranking performance.
- PPC behavior depended on initialization, dataset, and metric: asymmetry or delayed activation helped selected settings, but no strategy was uniformly best.
Related Work
The paper situates its dual-branch approach among ranking, metric-learning, hyperbolic, consistency-regularization, and margin-based methods. It distinguishes PPC from related objectives and reports that a tested contrastive constraint did not consistently help AMP ranking.
- Metric-learning comparison: A tested metric-learning contrastive constraint did not consistently improve AMP ranking performance and was excluded from the final model.The related adaptive-margin design places samples according to activity differences in latent space.
- Hyperbolic and protein-guided methods: HyperSeek’s protein-guided architecture could not be reproduced for AMP because the study’s datasets lack paired target-protein or pocket information.The implementation therefore retained a ligand-only backbone and Lorentz-space representation while removing protein-dependent branches.
- Consistency regularization: The method draws on joint dual-branch optimization and consistency regularization but does not use JoCoR’s small-loss sample selection.This separates DualCliff’s consistency mechanism from JoCoR’s noisy-label selection strategy.
- Dual-branch and preference consistency: PPC jointly trains two branches to align preference probabilities rather than directly matching raw scores.The design combines concurrent dual-branch optimization with sigmoid preference probabilities and symmetric MSE.
- Relation-based supervision: Unlike Margin-MSE, PPC uses concurrently trained branches instead of a pretrained frozen teacher providing offline margins.Both methods match relationships across samples, but their supervision and training setup differ.
Materials and Methods
The framework learns two parallel activity predictors using absolute regression, pairwise preference consistency, and thresholded listwise ranking objectives. Experiments evaluate AMP and small-molecule activity-cliff benchmarks with modality-matched backbones, controlled splits, and aligned random seeds.
- Model setup: DualCliff learns two parallel predictors and orients labels so larger values indicate stronger activity or potency.AMP labels use −log10(MIC), while small-molecule labels are likewise oriented toward stronger potency.
- Absolute activity regression: The averaged branch-wise MSE anchors both predictors to the ground-truth absolute activity scale.This preserves quantitative potency calibration alongside ranking objectives.
- Pairwise preference consistency: PPC converts branch score differences for unordered sample pairs into preference probabilities and minimizes their squared disagreement.It encourages relative-order agreement even when branch score magnitudes differ.
- Thresholded listwise ranking: The thresholded listwise loss makes each sufficiently high-activity anchor compete against weaker samples in the same batch.The activity-gap threshold removes nearly tied examples and focuses supervision on larger differences.
- Combined objective: The overall objective is a weighted sum of MSE, PPC, and listwise losses, with weights differing between AMP and small-molecule tasks.Both weights are 1 for AMP and 0.1 for small molecules; the listwise threshold is 0.2.
- Datasets and evaluation: AMP experiments use three datasets and five aligned seeds, while small-molecule experiments use three AC datasets with AC pairs assigned to test sets.Backbones include ESM2 for AMP and PNA or MolCLR-GIN for small molecules, with asymmetric branch initialization in dual-branch runs.
Main results
CliffRank showed strong cross-dataset ranking performance on antimicrobial peptide and small-molecule tasks, but its advantages varied across datasets, metrics, and training configurations.
- Antimicrobial peptide results: 0.5393 mean Spearman correlation and 21.4 mean Recall@50 were achieved by CliffRank with ESM2-t12 across three antimicrobial peptide datasets.Its per-dataset leaders varied, so the cross-dataset advantage was not uniform across individual tasks.
- Ablation study: 0.5358 mean Spearman correlation and 21.0 mean Recall@50 were the best joint AMP averages under the complete DualCliff objective.Adding dual-model MSE and PPC improved the averages relative to the single-model MSE baseline, while batching effects varied by setting.
- Ablation study: 0.5907 mean Spearman correlation was reached by the complete objective with MolCLR-GIN, while similarity-based batching reduced correlation to 0.5457 but increased Recall@50 to 28.3.For PNA, configuration averages were close, with delayed PPC reaching 0.6890 mean Spearman correlation and PPC alone preserving rather than improving the average.
Conclusions
DualCliff combines absolute-activity regression with listwise and preference-consistency learning for activity-cliff ranking. Across six datasets, it improved cross-dataset ranking performance, but PPC benefits depended on task, backbone, initialization, and training schedule.
- Method: DualCliff combines pointwise MSE, thresholded listwise loss, and PPC to align two predictors’ relative ordering in preference-probability space.PPC avoids forcing the predictors’ numerical outputs to match.
- Results: DualCliff achieved the highest cross-dataset mean Spearman correlation and Recall@50 on the antimicrobial-peptide tasks.
- Results: On small-molecule tasks, DualCliff’s main advantage was higher mean Spearman correlation, while mean Recall@50 matched ACANet-PNA.
- PPC behavior: PPC and listwise supervision improved ranking in selected settings, with gains depending on dataset, backbone, and sampling strategy.
- PPC behavior: Initialization asymmetry and delayed activation were beneficial in some settings, but no strategy was best across all targets and metrics.Initialization asymmetry improved MolCLR-GIN averages while showing target-specific effects; delayed PPC improved selected PNA metrics without a universally best schedule.
- Future work: Future work should test more targets and antimicrobial-peptide systems, develop adaptive PPC schedules, incorporate protein or membrane context, and quantify predictive uncertainty.