Use DynamicsPLM Finetuning Online
Commercially Available DynamicsPLM Finetuning No-Code Web Server
DynamicsPLM Finetuning: Conformation-Aware Protein Language Models for Function and Interaction Prediction
DynamicsPLM is a structure-informed protein language model that conditions on an ensemble of conformations rather than a single static structure. Each protein is represented by many conformations, and per-residue embeddings from those conformations are fused into one representation that is then fine-tuned for downstream prediction tasks. The method is described in a bioRxiv preprint.
Most structure-aware protein language models read a single structure, usually a predicted or crystallographic snapshot. Proteins are dynamic, and function often depends on which states a protein can adopt. The authors argue that ground-truth state labels are scarce, so a model that sees an ensemble can capture information a single structure misses.
Feature | Detail |
|---|---|
Inputs | Amino acid sequence plus AlphaFold2-predicted 3D structure |
Conformation generator | RocketSHP, kept frozen; other generators are said to be compatible |
Structure tokens | Each conformation is discretised with a VQ-VAE-pretrained autoencoder |
Encoder | Transformer layers initialised from SaProt-650M weights |
Evaluation tasks | HumanPPI, metal ion binding, Enzyme Commission (EC) prediction, DeepLoc localization |
How DynamicsPLM Combines Conformational Ensembles
Generate an ensemble: A frozen conformation generator produces multiple states for the input protein.
Tokenize each state: Every conformation is converted into structure tokens by a VQ-VAE-pretrained autoencoder.
Fuse across conformations: A dynamics embedding layer performs per-residue weighted fusion, with weights derived from token frequencies in the ensemble and normalized to sum to one. The authors say simple averaging of conformer embeddings gave lower performance and more computational overhead.
Fine-tune with task heads: Only the dynamics embedding layer and the PLM encoders are trained, with a task-specific classification head per downstream task.
Test splits cap pairwise sequence identity at 30%, and test proteins with at least 30% identity to the RocketSHP pretraining corpus were removed to limit leakage.
Benchmark Results Versus Single-Structure Baselines
Full test sets, compared with SaProt
Task | Test size | Gain over strongest baseline |
|---|---|---|
HumanPPI (accuracy) | 180 | +4.0 points |
Metal ion binding (accuracy) | 665 | +1.9 points |
DeepLoc (accuracy) | 2,747 | +1.6 points |
EC (Fmax) | 1,604 | +0.9 points |
The authors report Cohen's d above 0.8 across tasks and p below 0.05 using a paired t-test with Holm-Bonferroni correction, over 5 fine-tuning seeds.
Proteins with experimentally observed conformational diversity
On a dynamic subset built from CoDNaS-Q, gains were larger: HumanPPI +11.11 points (n = 18), metal ion binding +3.85 points (n = 101), EC Fmax +6.5 points (n = 134) and DeepLoc +6.25 points (n = 16). All four were reported as significant at p below 0.05, though several subsets are small. Case studies show corrected predictions for ATG10-ATG7 and MDM4-GCSAM interactions relative to SaProt.
What is Tamarind Bio?
Tamarind Bio is a no-code bioinformatics platform built to give life scientists and researchers access to powerful computational tools. Many cutting-edge machine learning models are hard to deploy and use. Tamarind provides an intuitive, web-based environment that removes the complexity of high-performance computing, software dependencies and command-line interfaces.
The platform is designed for biologists, chemists and other researchers who may not have a background in programming or cloud infrastructure but want to run models on their own data. Key features include:
A user-friendly graphical interface for setting up and launching experiments
A robust API for integration into existing research pipelines
An automated system for managing and scaling computational resources
Tamarind treats information and data security as a top priority, as detailed in its Trust Center and Terms of Service.
Accelerating Discovery with DynamicsPLM on Tamarind Bio
Protein-protein interaction prediction: Classify whether two human proteins interact, as in the HumanPPI benchmark.
Enzyme function annotation: Predict Enzyme Commission classes for proteins that sample multiple states.
Metal ion binding: Predict whether a protein binds metal ions.
Subcellular localization: Predict where a protein is located in the cell, as in DeepLoc.
How to Use DynamicsPLM Finetuning on Tamarind Bio
Log in: Sign in at tamarind.bio and open the DynamicsPLM Finetuning tool.
Provide task data: Upload protein sequences with structures and labels for your classification task, in the format the tool requests.
Select the task: Choose the downstream task type if exposed.
Set training parameters: Adjust fine-tuning settings such as epochs or seeds if exposed.
Run: Launch the fine-tuning job.
Download outputs: Retrieve the fine-tuned model and predictions.
Interpret: Compare performance with a single-structure baseline on your own held-out set.
Things to Keep in Mind
Ground-truth state labels are scarce, so it is often unknown which conformations are biologically active for a given label.
Performance depends on an external conformer generator whose coverage and biases may vary across protein families.
Backbone RMSD is an imprecise surrogate for functional state.
Dynamic-subset results rely on small samples, and case studies are illustrative rather than systematic.
Source: bioRxiv preprint on DynamicsPLM, DOI 10.1101/2025.10.06.680789 (version 1). No code repository was stated in the text consulted.