Protein Benchmark for ProtEnrich

The paper is under review.

[Github Repo] | [Model Collection] | [Cite]

Abstract

Protein language models effectively capture evolutionary and functional signals from sequence data but lack explicit representation of the biophysical properties that govern protein structure and dynamics. Existing multimodal approaches attempt to integrate such physical information through direct fusion, often requiring multimodal inputs at inference time and distorting the sequence embedding space. Consequently, a fundamental challenge of how to incorporate structural and dynamical knowledge into sequence representations without disrupting their established semantic organization remains a field-of-research. We introduce ProtEnrich, a representation learning framework based on a residual multimodal enrichment paradigm. ProtEnrich decomposes sequence embeddings into two complementary latent subspaces, an anchor subspace that preserves sequence semantics, and an alignment subspace that encodes biophysical relationships. By converting multimodal information derived from ProstT5 and RocketSHP to a low-energy residual component, our approach injects physical representation while maintaining the original sequence embedding. Across eight diverse protein foundational models trained on 550,120 SwissProt proteins with AlphaFold structures, enriched embeddings improved zero-shot remote homology retrieval, increasing Precision@10 and MRR by up to 0.13 and 0.11, respectively. Downstream performance also improved on structure-dependent tasks, reducing fluorescence prediction error by up to 16% and increasing metal ion binding AUCROC by up to 2.4 points, while requiring only sequence input at inference.

Dataset Details

This repository provides the benchmarks used in ProtEnrich paper.

Pretraining

This dataset comprises 550,120 proteins from SwissProt, representing the pretraining database used in the ProtEnrich paper. If you use this dataset, please cite our paper.

Out-of-distribution

This dataset comprises 10,000 proteins from TrEMBL, representing the proteins out-of-pretraining (out-of-distribution) set. If you use this dataset, please cite our paper.

Fluorescence

The dataset consists of 21,446 training proteins, 5,362 validation proteins, and 27,217 test proteins. Fluorescence values range from 1.28 to 4.12. The objective is to predict the fluorescence intensity of mutant green fluorescent proteins. If you use this dataset, please cite this paper.

Localization Binary

The dataset consists of 5,735 training proteins, 1,009 validation proteins, and 1,728 test proteins. The objective is to determine whether a protein is membrane-bound. If you use this dataset, please cite this paper.

Localization Multiclass

The dataset consists of 9,324 training proteins, 1,658 validation proteins, and 2,742 test proteins. The objective is to determine ten possible subcellular compartments. If you use this dataset, please cite this paper.

Metal Ion Binding

The dataset consists of 5,066 training proteins, 662 validation proteins, and 665 test proteins. The objective is to predict whether a protein contains a metal ion binding site. If you use this dataset, please cite this paper.

Protein-Protein Interaction

The dataset consists of 5,131 proteins forming 26,317 training interactions, 215 proteins with 234 validation interactions, and 180 proteins with 180 test interactions. The objective is to determine whether two human proteins interact. If you use this dataset, please cite this paper.

Protein Function - Biological Process

The dataset consists of 80,185 training proteins, 10,127 validation proteins, and 9,910 test proteins, with 595 possible GO terms. The objective is to predict GO terms considering Biological Process (BP). If you use this dataset, please cite our paper.

Protein Function - Cellular Component

The dataset consists of 85,314 training proteins, 10,753 validation proteins, and 10,467 test proteins, with 560 possible GO terms. The objective is to predict GO terms considering Cellular Component (CC). If you use this dataset, please cite our paper.

Protein Function - Molecular Function

The dataset consists of 75,072 training proteins, 9,295 validation proteins, and 9,235 test proteins, with 550 possible GO terms. The objective is to predict GO terms considering Molecular Function (MF). If you use this dataset, please cite our paper.

Solubility

The dataset consists of 62,478 training proteins, 6,942 validation proteins, and 1,999 test proteins. The objective is to predict whether a protein is soluble. If you use this dataset, please cite this paper.

Thermostability

The dataset consists of 5,025 training proteins, 636 validation proteins, and 1,329 test proteins. Thermostability values range from 40.20 to 67.00. The objective is to predict the melting temperature of human proteins. If you use this dataset, please cite this paper.

Remote Homology - Family

The dataset consists of 9,084 proteins. Each protein has a family identifier. If you use this dataset, please cite this paper.

Remote Homology - Superfamily

The dataset consists of 10,524 proteins. Each protein has a superfamily identifier. If you use this dataset, please cite this paper.

Remote Homology - Fold

The dataset consists of 10,892 proteins. Each protein has a fold identifier. If you use this dataset, please cite this paper.

Dataset Usage

Pretraining

Use the code below to load the pretraining dataset.

from datasets import load_dataset 

dataset_dict = load_dataset("SaeedLab/ProtEnrich", data_dir="pretraining")

Out-of-distribution

Use the code below to load the out-of-distribution dataset.

from datasets import load_dataset 

dataset_dict = load_dataset("SaeedLab/ProtEnrich", data_dir="out-of-distribution")

Fluorescence

Use the code below to load the fluorescence dataset.

from datasets import load_dataset 

dataset_dict = load_dataset("SaeedLab/ProtEnrich", data_dir="fluorescence")

Localization Binary

Use the code below to load the localization binary dataset.

from datasets import load_dataset 

dataset_dict = load_dataset("SaeedLab/ProtEnrich", data_dir="localization_bin")

Localization Multiclass

Use the code below to load the localization multiclasse dataset.

from datasets import load_dataset 

dataset_dict = load_dataset("SaeedLab/ProtEnrich", data_dir="localization_multi")

Metal Ion Binding

Use the code below to load the metal ion binding dataset.

from datasets import load_dataset 

dataset_dict = load_dataset("SaeedLab/ProtEnrich", data_dir="metal_ion_binding")

Protein-Protein Interaction

Use the code below to load the protein-protein interaction (ppi) dataset.

from datasets import load_dataset 

dataset_dict = load_dataset("SaeedLab/ProtEnrich", data_dir="ppi")

Protein Function - Biological Process

Use the code below to load the protein function for biological process dataset.

from datasets import load_dataset 
from huggingface_hub import hf_hub_download
import ast

def parse_terms(example):
  example["terms"] = ast.literal_eval(example["terms"])
  return example

dataset = load_dataset("SaeedLab/ProtEnrich", data_dir="protein_function_bp")
dataset["train"] = dataset["train"].map(parse_terms)
dataset["validation"] = dataset["validation"].map(parse_terms)
dataset["test"] = dataset["test"].map(parse_terms)

ic_values = load_dataset("SaeedLab/ProtEnrich", data_files={"ic": "protein_function_bp/ic.parquet"})['ic'].to_pandas()
obo_path = hf_hub_download(repo_id="SaeedLab/ProtEnrich", filename=f"protein_function_bp/go-basic.obo", repo_type="dataset")

Protein Function - Cellular Component

Use the code below to load the protein function for cellular component dataset.

from datasets import load_dataset 
from huggingface_hub import hf_hub_download
import ast

def parse_terms(example):
  example["terms"] = ast.literal_eval(example["terms"])
  return example

dataset = load_dataset("SaeedLab/ProtEnrich", data_dir="protein_function_cc")
dataset["train"] = dataset["train"].map(parse_terms)
dataset["validation"] = dataset["validation"].map(parse_terms)
dataset["test"] = dataset["test"].map(parse_terms)

ic_values = load_dataset("SaeedLab/ProtEnrich", data_files={"ic": "protein_function_cc/ic.parquet"})['ic'].to_pandas()
obo_path = hf_hub_download(repo_id="SaeedLab/ProtEnrich", filename=f"protein_function_cc/go-basic.obo", repo_type="dataset")

Protein Function - Molecular Function

Use the code below to load the protein function for molecular function dataset.

from datasets import load_dataset 
from huggingface_hub import hf_hub_download
import ast

def parse_terms(example):
  example["terms"] = ast.literal_eval(example["terms"])
  return example

dataset = load_dataset("SaeedLab/ProtEnrich", data_dir="protein_function_mf")
dataset["train"] = dataset["train"].map(parse_terms)
dataset["validation"] = dataset["validation"].map(parse_terms)
dataset["test"] = dataset["test"].map(parse_terms)

ic_values = load_dataset("SaeedLab/ProtEnrich", data_files={"ic": "protein_function_mf/ic.parquet"})['ic'].to_pandas()
obo_path = hf_hub_download(repo_id="SaeedLab/ProtEnrich", filename=f"protein_function_mf/go-basic.obo", repo_type="dataset")

Solubility

Use the code below to load the solubility dataset.

from datasets import load_dataset 

dataset_dict = load_dataset("SaeedLab/ProtEnrich", data_dir="solubility")

Thermostability

Use the code below to load the thermostability dataset.

from datasets import load_dataset 

dataset_dict = load_dataset("SaeedLab/ProtEnrich", data_dir="thermostability")

Remote Homology - Family

Use the code below to load the remote homology - family dataset.

from datasets import load_dataset 

dataset_dict = load_dataset("SaeedLab/ProtEnrich", data_dir="remote_homology_family")

Remote Homology - Superfamily

Use the code below to load the remote homology - superfamily dataset.

from datasets import load_dataset 

dataset_dict = load_dataset("SaeedLab/ProtEnrich", data_dir="remote_homology_superfamily")

Remote Homology - Fold

Use the code below to load the remote homology - fold dataset.

from datasets import load_dataset 

dataset_dict = load_dataset("SaeedLab/ProtEnrich", data_dir="remote_homology_fold")

Citation

The paper is under review. As soon as it is accepted, we will update this section.

License

This model and associated code are released under the CC-BY-NC-ND 4.0 license and may only be used for non-commercial, academic research purposes with proper attribution. Any commercial use, sale, or other monetization of this model and its derivatives, which include models trained on outputs from the model or datasets created from the model, is prohibited and requires prior approval. Downloading the model requires prior registration on Hugging Face and agreeing to the terms of use. By downloading this model, you agree not to distribute, publish or reproduce a copy of the model. If another user within your organization wishes to use the model, they must register as an individual user and agree to comply with the terms of use. Users may not attempt to re-identify the deidentified data used to develop the underlying model. If you are a commercial entity, please contact the corresponding author.

Contact

For any additional questions or comments, contact Fahad Saeed (fsaeed@fiu.edu).