Protein Language Modeling beyond static folds reveals sequence-encoded flexibility

Finn H. Lüth, Victor Mihaila, Milot Mirdita, Martin Steinegger, Burkhard Rost, Michael HeinzingerbioRxiv 2026 · DOI

TL;DR

This paper introduces ProtProfileMD, a novel method that uses protein language models to predict protein dynamics directly from amino acid sequences. By capturing a protein's "dynamic fingerprint," ProtProfileMD significantly improves remote homology detection, a crucial task in understanding protein function and evolution.

What Problem Does This Paper Address?

Proteins are not static entities; they are dynamic machines that explore various conformations to perform their functions. However, most computational methods for analyzing proteins, including homology detection, rely on static 3D structures. This overlooks the crucial role of protein dynamics, which can range from subtle movements to large-scale conformational changes. While molecular dynamics (MD) simulations can capture these dynamics, they are computationally expensive and difficult to scale to entire proteomes. This paper aims to bridge this gap by developing a method that can infer protein dynamics from sequences alone, making dynamic information accessible for large-scale analyses.

Methods

The core of ProtProfileMD lies in leveraging pre-trained protein language models (pLMs), specifically ProstT5. The authors introduce a new representation called "FlexProfiles," which are per-residue probability distributions over discrete structural states (the 3Di alphabet from Foldseek). These FlexProfiles are derived from molecular dynamics simulations.

The process involves:

  1. Generating Ground Truth FlexProfiles: MD trajectories of proteins are discretized into sequences of 3Di tokens. These sequences are then aggregated into Position Probability Matrices (PPMs) representing the probability of each 3Di token at each residue position.
  2. Training ProtProfileMD: A parameter-efficient finetuning technique called Low-Rank Adaptation (LoRA) is applied to ProstT5. The pLM is trained to predict these ground-truth 3Di-FlexProfiles directly from amino acid sequences. The training objective is to minimize the Kullback-Leibler (KL) divergence between the predicted and ground-truth FlexProfiles.
  3. Homology Detection: The predicted 3Di-FlexProfiles are then used with Foldseek for remote homology detection, allowing for comparisons based on dynamic information rather than just static structure.

Datasets

  • mdCATH: This dataset, comprising simulations of 5,398 protein domains from the CATH database, was used for generating ground-truth 3Di-FlexProfiles. Each domain was simulated with five replicas at five different temperatures. The authors chose to train their model exclusively on data from the lowest temperature (320K) due to increasing variance and unfolding at higher temperatures.
  • SCOPe: This dataset was used for evaluating the performance of ProtProfileMD in remote homology detection across different hierarchical levels (Family, Superfamily, Fold).

Key Insights and Ideas

  • 3Di captures dynamics: Despite being a coarse-grained descriptor, the 3Di alphabet is surprisingly effective at capturing aspects of protein conformational changes observed in MD simulations.
  • FlexProfiles as a "dynamic fingerprint": The concept of FlexProfiles provides a protein-specific representation of its dynamic landscape, analogous to how sequence profiles represent evolutionary conservation.
  • pLMs can learn dynamics: Protein language models, when finetuned with FlexProfiles, can learn to predict these dynamic signatures directly from amino acid sequences, bypassing the need for computationally expensive MD simulations.
  • Dynamics enhance homology detection: Incorporating dynamic information through FlexProfiles significantly improves the sensitivity of remote homology detection, enabling the identification of more distant evolutionary relationships.

Key Results

  • Correlation with MD observables: Predicted 3Di-FlexProfiles show strong correlations with established MD observables:
  • Weighted Hamming distance of 3Di sequences correlates well with RMSD (Pearson correlation of 0.65-0.79).
  • Shannon entropy of FlexProfiles correlates well with RMSF (Pearson correlation of 0.76-0.87).
  • Improved Homology Detection: ProtProfileMD, when used with Foldseek for homology detection, shows improved sensitivity across all SCOPe classification levels (Family, Superfamily, Fold) compared to methods relying solely on static sequences or static 3Di representations. For example, at the SCOPe Family level, ProtProfileMD achieved a sensitivity of 0.875, compared to 0.341 for MMseqs2 (amino acid search) and 0.864 for Foldseek with ProstT5-predicted 3Di.
  • High Throughput: The method achieves high throughput, with processing times comparable to existing pLM-based sequence analysis tools.

Limitations and Future Work

  • Training Data Scale: The mdCATH dataset, while large, consists of single domains and has limitations in spatial and temporal scale for MD-derived motions. Expanding the training data with more diverse and extensive simulations is suggested.
  • Structural Alphabet Resolution: The 3Di alphabet, with 20 states, might be insufficient to capture the full spectrum of dynamic variability. Exploring more refined structural alphabets could be beneficial.
  • Profile-Profile Comparison: Current search implementations (like Foldseek) do not fully support profile-profile comparison, limiting the potential of FlexProfiles. Developing such capabilities would further enhance the method's utility.
  • Broader Benchmarks: Applying the method to more challenging benchmarks, especially those involving more intrinsically disordered proteins or proteins with complex conformational changes, is a future direction.
  • Novel Applications: The authors envision ProtProfileMD enabling new applications such as proteome-scale annotation and searching for proteins or motifs that lack rigid structures.

Summary updated

Abstract

Abstract Motivation Proteins function through motion. Yet, most discoveries still commence with static representations of protein structures. Here, we investigated the feasibility of leveraging protein dynamics to improve homology detection. Results We introduce ProtProfileMD, a sequence-to-3D-probability model that predicts, from an amino acid sequence, a profile of discrete structural representations capturing protein dynamics. We applied supervised parameter-efficient finetuning of the ProstT5 protein Language Model (pLM) to predict per-residue distributions over Foldseek’s 3Di alphabet derived from motions observed in molecular dynamics. This original result reveals that the 3Di tokens, despite being coarse-grained descriptors of 3D structure, still offer sufficient resolution to capture aspects of conformational changes. This is evidenced by a correlation between fluctuations in the 3D protein structure over the course of a molecular dynamics trajectory and the entropy of 3Di states. Based on this insight, we introduce a proof-of-concept for making remote homology detection of proteins more sensitive by leveraging a protein’s distinctive dynamic fingerprint captured by our model. Our method recovers flexibility signals with a fidelity that is biologically relevant, improving search and complementing protein structure predictions, for example, by flagging flexible, disordered, or other functionally relevant regions. Availability and Implementation ProtProfileMD is available at github.com/finnlueth/ProtProfileMD . The associated training data and model weights are available at huggingface.co/datasets/finnlueth/ProtProfileMD and huggingface.co/finnlueth/ProtProfileMD .
BibTeX
@misc{luth2026protein,
  author = {Finn H. Lüth and Victor Mihaila and Milot Mirdita and Martin Steinegger and Burkhard Rost and Michael Heinzinger},
  title = {{Protein Language Modeling beyond static folds reveals sequence-encoded flexibility}},
  year = {2026},
  doi = {10.64898/2026.01.21.700698},
  howpublished = {bioRxiv},
}
Figure