Recent studies exploring the abilities of transformer-based protein language models have highlighted their performance on the task of remote homology detection, but have not provided datasets or evaluation procedures geared toward properly measuring performance on this task. With the goal of obtaining more informative and reproducible results, we offer a detailed procedure for constructing datasets and evaluating remote homology detection performance in a way that allows detailed analyses to be performed that shed light on the remote homology detection performance throughout the “twilight zone” of low sequence similarity. Using the proposed procedures, we found that three stateof-the-art protein language models exhibit diminishing performance
when the pairwise sequence similarity between the query sequence and other proteins is restricted to below 35%
identity.
more »
« less
In the twilight zone of protein sequence homology: do protein language models learn protein structure?
Protein language models based on the transformer architecture are increasingly improving performance on protein prediction tasks, including secondary structure, subcellular localization, and more. Despite being trained only on protein sequences, protein language models appear to implicitly learn protein structure. This paper investigates whether sequence representations learned by protein language models encode structural information and to what extent.
We address this by evaluating protein language models on remote homology prediction, where identifying remote homologs from sequence information alone requires structural knowledge, especially in the “twilight zone” of very low sequence identity. Through rigorous testing at progressively lower sequence identities, we profile the performance of protein language models ranging from millions to billions of parameters in a zero-shot setting. Our findings indicate that while transformer-based protein language models outperform traditional sequence alignment methods, they still struggle in the twilight zone. This suggests that current protein language models have not sufficiently learned protein structure to address remote homology prediction when sequence signals are weak.
We believe this opens the way for further research both on remote homology prediction and on the broader goal of learning sequence- and structure-rich representations of protein molecules. All code, data, and models are made publicly available.
- PAR ID:
- 10535670
- Publisher / Repository:
- Oxford University Press
- Date Published:
- Journal Name:
- Bioinformatics Advances
- Volume:
- 4
- Issue:
- 1
- ISSN:
- 2635-0041
- Format(s):
- Medium: X
- Sponsoring Org:
- National Science Foundation
More Like this
-
-
more » « less
Abstract Motivation As fewer than 1% of proteins have protein function information determined experimentally, computationally predicting the function of proteins is critical for obtaining functional information for most proteins and has been a major challenge in protein bioinformatics. Despite the significant progress made in protein function prediction by the community in the last decade, the general accuracy of protein function prediction is still not high, particularly for rare function terms associated with few proteins in the protein function annotation database such as the UniProt.
Results We introduce TransFew, a new transformer model, to learn the representations of both protein sequences and function labels [Gene Ontology (GO) terms] to predict the function of proteins. TransFew leverages a large pre-trained protein language model (ESM2-t48) to learn function-relevant representations of proteins from raw protein sequences and uses a biological natural language model (BioBert) and a graph convolutional neural network-based autoencoder to generate semantic representations of GO terms from their textual definition and hierarchical relationships, which are combined together to predict protein function via the cross-attention. Integrating the protein sequence and label representations not only enhances overall function prediction accuracy, but delivers a robust performance of predicting rare function terms with limited annotations by facilitating annotation transfer between GO terms.
Availability and implementation https://github.com/BioinfoMachineLearning/TransFew.
-
more » « less
Abstract Motivation Quality assessment (QA) of predicted protein tertiary structure models plays an important role in ranking and using them. With the recent development of deep learning end-to-end protein structure prediction techniques for generating highly confident tertiary structures for most proteins, it is important to explore corresponding QA strategies to evaluate and select the structural models predicted by them since these models have better quality and different properties than the models predicted by traditional tertiary structure prediction methods.
Results We develop EnQA, a novel graph-based 3D-equivariant neural network method that is equivariant to rotation and translation of 3D objects to estimate the accuracy of protein structural models by leveraging the structural features acquired from the state-of-the-art tertiary structure prediction method—AlphaFold2. We train and test the method on both traditional model datasets (e.g. the datasets of the Critical Assessment of Techniques for Protein Structure Prediction) and a new dataset of high-quality structural models predicted only by AlphaFold2 for the proteins whose experimental structures were released recently. Our approach achieves state-of-the-art performance on protein structural models predicted by both traditional protein structure prediction methods and the latest end-to-end deep learning method—AlphaFold2. It performs even better than the model QA scores provided by AlphaFold2 itself. The results illustrate that the 3D-equivariant graph neural network is a promising approach to the evaluation of protein structural models. Integrating AlphaFold2 features with other complementary sequence and structural features is important for improving protein model QA.
Availability and implementation The source code is available at https://github.com/BioinfoMachineLearning/EnQA.
Supplementary information Supplementary data are available at Bioinformatics online.
-
more » « less
Abstract Motivation Sequence-based protein–protein interaction (PPI) prediction represents a fundamental computational biology problem. To address this problem, extensive research efforts have been made to extract predefined features from the sequences. Based on these features, statistical algorithms are learned to classify the PPIs. However, such explicit features are usually costly to extract, and typically have limited coverage on the PPI information.
Results We present an end-to-end framework, PIPR (Protein–Protein Interaction Prediction Based on Siamese Residual RCNN), for PPI predictions using only the protein sequences. PIPR incorporates a deep residual recurrent convolutional neural network in the Siamese architecture, which leverages both robust local features and contextualized information, which are significant for capturing the mutual influence of proteins sequences. PIPR relieves the data pre-processing efforts that are required by other systems, and generalizes well to different application scenarios. Experimental evaluations show that PIPR outperforms various state-of-the-art systems on the binary PPI prediction problem. Moreover, it shows a promising performance on more challenging problems of interaction type prediction and binding affinity estimation, where existing approaches fall short.
Availability and implementation The implementation is available at https://github.com/muhaochen/seq_ppi.git.
Supplementary information Supplementary data are available at Bioinformatics online.
-
more » « less
Abstract Motivation Protein structure prediction has been greatly improved by deep learning, but the contribution of different information is yet to be fully understood. This article studies the impacts of two kinds of information for structure prediction: template and multiple sequence alignment (MSA) embedding. Templates have been used by some methods before, such as AlphaFold2, RoseTTAFold and RaptorX. AlphaFold2 and RosetTTAFold only used templates detected by HHsearch, which may not perform very well on some targets. In addition, sequence embedding generated by pre-trained protein language models has not been fully explored for structure prediction. In this article, we study the impact of templates (including the number of templates, the template quality and how the templates are generated) on protein structure prediction accuracy, especially when the templates are detected by methods other than HHsearch. We also study the impact of sequence embedding (generated by MSATransformer and ESM-1b) on structure prediction.
Results We have implemented a deep learning method for protein structure prediction that may take templates and MSA embedding as extra inputs. We study the contribution of templates and MSA embedding to structure prediction accuracy. Our experimental results show that templates can improve structure prediction on 71 of 110 CASP13 (13th Critical Assessment of Structure Prediction) targets and 47 of 91 CASP14 targets, and templates are particularly useful for targets with similar templates. MSA embedding can improve structure prediction on 63 of 91 CASP14 (14th Critical Assessment of Structure Prediction) targets and 87 of 183 CAMEO targets and is particularly useful for proteins with shallow MSAs. When both templates and MSA embedding are used, our method can predict correct folds (TMscore > 0.5) for 16 of 23 CASP14 FM targets and 14 of 18 Continuous Automated Model Evaluation (CAMEO) targets, outperforming RoseTTAFold by 5% and 7%, respectively.
Availability and implementation Available at https://github.com/xluo233/RaptorXFold.
Supplementary information Supplementary data are available at Bioinformatics online.
