Empirical Geometry-Guided Modeling Enables Robust, High-Throughput Collagen Structure Prediction
Protein structure prediction is increasingly dominated by large learned generative models, yet for proteins governed by strong structural constraints, much of the relevant conformational space may be captured by substantially more compact representations. Collagen provides a compelling test case: its repeating Gly-X-Y sequence, restricted backbone conformations, and conserved triple-helical…
In a groundbreaking development, researchers have created a Collagen-specific Deterministic Structure Modeler (CDSM) that significantly outperforms large generative learned models in predicting the structure of collagen. Collagen, a protein with a highly constrained repeating Gly-X-Y sequence, restricted backbone conformations, and conserved triple-helical topology, serves as an ideal test case for this new approach.
The CDSM extends the empirical geometric parameterization of THeBuScr into a robust, all-atom structure-prediction pipeline. When benchmarked against AlphaFold 3, Boltz-2, Chai-1, and Protenix-v1 on 80 experimentally resolved collagen triple-helical structures, CDSM's coverage increased from 8.8% to 93.8%. Across the 75 successfully predicted structures, CDSM closely replicated experimental backbone and global geometry, with fewer large-error predictions compared to the learned models.
Interestingly, when evaluated only on structures deposited after each learned model's training-data cutoff, CDSM's competitive advantage grew even more pronounced. Aggregate win rates for TM-score and backbone RMSD increased from 29% to 55% and from 40% to 65%, respectively. Despite these superior performance metrics, CDSM remains remarkably efficient, generating structures in just 2.4 seconds on a single CPU core at an estimated computational cost of $3 x 10^-5 per structure.
This cost is 400 to 790 times lower than the learned methods even when run on the cheapest compatible GPU.
The implications of this work are far-reaching. By identifying and explicitly encoding relevant geometric constraints, CDSM dramatically compresses the structure-prediction problem, retaining much of the accuracy at an order-of-magnitude lower computational cost. Moreover, this study highlights the potential for compact, executable scientific representations that can complement large learned models.
It also motivates future AI-driven searches over algorithms, representations, and empirical parameterizations to continually refine such models for other structural protein domains.
Written by urgent.news from bioRxiv's reporting — not their text. Machine-written — may contain errors; check the original before relying on it.
