Embeddings from protein language models predict conservation and variant effects

Abstract The emergence of SARS-CoV-2 variants stressed the demand for tools allowing to interpret the effect of single amino acid variants (SAVs) on protein function. While Deep Mutational Scanning (DMS) sets continue to expand our understanding of the mutational landscape of single proteins, the results continue to challenge analyses. Protein Language Models (pLMs) use the latest deep learning (DL) algorithms to leverage growing databases of protein sequences. These methods learn to predict missing or masked amino acids from the context of entire sequence regions. Here, we used pLM representations (embeddings) to predict sequence conservation and SAV effects without multiple sequence alignments (MSAs). Embeddings alone predicted residue conservation almost as accurately from single sequences as ConSeq using MSAs (two-state Matthew Correlation Coefficient – MCC - for ProtT5 embeddings of 0.596 ± 0.006 vs. 0.608 ± 0.006 for ConSeq). Inputting the conservation prediction along with BLOSUM62 substitution scores and pLM mask reconstruction probabilities into a simplistic logistic regression (LR) ensemble for Variant Effect Scoring without alignments (VESPA) predicted SAV effect magnitude without any optimization on DMS data. Comparing predictions for a standard set of 39 DMS experiments to other methods (incl. ESM-1v, DeepSequence, and GEMME) revealed our approach as competitive with the state-of-the-art (SOTA) methods using MSA input. No method outperformed all others, neither consistently nor in a statistically significant manner, independently of the performance measure applied (incl. two-state accuracy: Q2, MCC, Spearman and Pearson correlation). Lastly, we investigated binary effect predictions on DMS experiments for four human proteins. Overall, embedding-based methods have become competitive with methods relying on MSAs for SAV effect prediction at a fraction of the costs in computing/energy. Our method predicted SAV effects for the entire human proteome (~ 20k proteins) within 40 minutes on one Nvidia Quadro RTX 8000. All methods and data sets are freely available for local and online execution through bioembeddings.com, https://github.com/Rostlab/VESPA, and PredictProtein.

Download Full-text

Embeddings from protein language models predict conservation and variant effects

Human Genetics ◽

10.1007/s00439-021-02411-y ◽

2021 ◽

Author(s):

Céline Marquet ◽

Michael Heinzinger ◽

Tobias Olenyi ◽

Christian Dallago ◽

Kyra Erckert ◽

...

Keyword(s):

Protein Function ◽

Pearson Correlation ◽

Performance Measure ◽

Language Models ◽

Single Amino Acid ◽

Data Sets ◽

Sequence Alignments ◽

Multiple Sequence ◽

Multiple Sequence Alignments ◽

Human Proteins

AbstractThe emergence of SARS-CoV-2 variants stressed the demand for tools allowing to interpret the effect of single amino acid variants (SAVs) on protein function. While Deep Mutational Scanning (DMS) sets continue to expand our understanding of the mutational landscape of single proteins, the results continue to challenge analyses. Protein Language Models (pLMs) use the latest deep learning (DL) algorithms to leverage growing databases of protein sequences. These methods learn to predict missing or masked amino acids from the context of entire sequence regions. Here, we used pLM representations (embeddings) to predict sequence conservation and SAV effects without multiple sequence alignments (MSAs). Embeddings alone predicted residue conservation almost as accurately from single sequences as ConSeq using MSAs (two-state Matthews Correlation Coefficient—MCC—for ProtT5 embeddings of 0.596 ± 0.006 vs. 0.608 ± 0.006 for ConSeq). Inputting the conservation prediction along with BLOSUM62 substitution scores and pLM mask reconstruction probabilities into a simplistic logistic regression (LR) ensemble for Variant Effect Score Prediction without Alignments (VESPA) predicted SAV effect magnitude without any optimization on DMS data. Comparing predictions for a standard set of 39 DMS experiments to other methods (incl. ESM-1v, DeepSequence, and GEMME) revealed our approach as competitive with the state-of-the-art (SOTA) methods using MSA input. No method outperformed all others, neither consistently nor statistically significantly, independently of the performance measure applied (Spearman and Pearson correlation). Finally, we investigated binary effect predictions on DMS experiments for four human proteins. Overall, embedding-based methods have become competitive with methods relying on MSAs for SAV effect prediction at a fraction of the costs in computing/energy. Our method predicted SAV effects for the entire human proteome (~ 20 k proteins) within 40 min on one Nvidia Quadro RTX 8000. All methods and data sets are freely available for local and online execution through bioembeddings.com, https://github.com/Rostlab/VESPA, and PredictProtein.

Download Full-text

Embeddings from protein language models predict conservation and variant effects

10.21203/rs.3.rs-584804/v3 ◽

2021 ◽

Author(s):

Céline Marquet ◽

Michael Heinzinger ◽

Tobias Olenyi ◽

Christian Dallago ◽

Michael Bernhofer ◽

...

Keyword(s):

Protein Function ◽

Pearson Correlation ◽

Performance Measure ◽

Language Models ◽

Single Amino Acid ◽

Data Sets ◽

Sequence Alignments ◽

Multiple Sequence ◽

Multiple Sequence Alignments ◽

Human Proteins

Abstract The emergence of SARS-CoV-2 variants stressed the demand for tools allowing to interpret the effect of single amino acid variants (SAVs) on protein function. While Deep Mutational Scanning (DMS) sets continue to expand our understanding of the mutational landscape of single proteins, the results continue to challenge analyses. Protein Language Models (pLMs) use the latest deep learning (DL) algorithms to leverage growing databases of protein sequences. These methods learn to predict missing or masked amino acids from the context of entire sequence regions. Here, we used pLM representations (embeddings) to predict sequence conservation and SAV effects without multiple sequence alignments (MSAs). Embeddings alone predicted residue conservation almost as accurately from single sequences as ConSeq using MSAs (two-state Matthews Correlation Coefficient – MCC - for ProtT5 embeddings of 0.596±0.006 vs. 0.608±0.006 for ConSeq). Inputting the conservation prediction along with BLOSUM62 substitution scores and pLM mask reconstruction probabilities into a simplistic logistic regression (LR) ensemble for Variant Effect Score Prediction without Alignments (VESPA) predicted SAV effect magnitude without any optimization on DMS data. Comparing predictions for a standard set of 39 DMS experiments to other methods (incl. ESM-1v, DeepSequence, and GEMME) revealed our approach as competitive with the state-of-the-art (SOTA) methods using MSA input. No method outperformed all others, neither consistently nor statistically significantly, independently of the performance measure applied (Spearman and Pearson correlation). Lastly, we investigated binary effect predictions on DMS experiments for four human proteins. Overall, embedding-based methods have become competitive with methods relying on MSAs for SAV effect prediction at a fraction of the costs in computing/energy. Our method predicted SAV effects for the entire human proteome (~20k proteins) within 40 minutes on one Nvidia Quadro RTX 8000. All methods and data sets are freely available for local and online execution through bioembeddings.com, https://github.com/Rostlab/VESPA, and PredictProtein.

Download Full-text

Embeddings from protein language models predict conservation and variant effects

10.21203/rs.3.rs-584804/v1 ◽

2021 ◽

Author(s):

Céline Marquet ◽

Michael Heinzinger ◽

Tobias Olenyi ◽

Christian Dallago ◽

Michael Bernhofer ◽

...

Keyword(s):

Protein Function ◽

Language Models ◽

Single Amino Acid ◽

Sequence Alignments ◽

Multiple Sequence ◽

Multiple Sequence Alignments ◽

Human Proteins ◽

Entire Sequence ◽

Embedding Methods ◽

Better Than

Abstract The emergence of SARS-CoV-2 variants stressed the demand for tools allowing to interpret the effect of single amino acid variants (SAVs) on protein function. While Deep Mutational Scanning (DMS) sets continue to expand our understanding of the mutational landscape of single proteins, the results continue to challenge analyses. Protein Language Models (LMs) use the latest deep learning (DL) algorithms to leverage growing databases of protein sequences. These methods learn to predict missing or marked amino acids from the context of entire sequence regions. Here, we explored how to benefit from learned protein LM representations (embeddings) to predict SAV effects. Although we have failed so far to predict SAV effects directly from embeddings, this input alone predicted residue conservation almost as accurately from single sequences as using multiple sequence alignments (MSAs) with a two-state per-residue accuracy (conserved/not) of Q2=80% (embeddings) vs. 81% (ConSeq). Considering all SAVs at all residue positions predicted as conserved to affect function reached 68.6% (Q2: effect/neutral; for PMD) without optimization, compared to an expert solution such as SNAP2 (Q2=69.8). Combining predicted conservation with BLOSUM62 to obtain variant-specific binary predictions, DMS experiments of four human proteins were predicted better than by SNAP2, and better than by applying the same simplistic approach to conservation taken from ConSeq. Thus, embedding methods have become competitive with methods relying on MSAs for SAV effect prediction at a fraction of the costs in computing/energy. This allowed prediction of SAV effects for the entire human proteome (~20k proteins) within 17 minutes on a single GPU.

Download Full-text

Protein embeddings and deep learning predict binding residues for various ligand classes

10.1101/2021.09.03.458869 ◽

2021 ◽

Author(s):

Maria Littmann ◽

Michael Heinzinger ◽

Christian Dallago ◽

Konstantin Weissenow ◽

Burkhard Rost

Keyword(s):

Nucleic Acids ◽

Metal Ions ◽

Small Molecules ◽

Protein Function ◽

Language Model ◽

Data Sets ◽

Sequence Alignments ◽

Multiple Sequence ◽

Human Proteins ◽

Binding Residues

AbstractOne important aspect of protein function is the binding of proteins to ligands, including small molecules, metal ions, and macromolecules such as DNA or RNA. Despite decades of experimental progress many binding sites remain obscure. Here, we proposed bindEmbed21, a method predicting whether a protein residue binds to metal ions, nucleic acids, or small molecules. The Artificial Intelligence (AI)-based method exclusively uses embeddings from the Transformer-based protein Language Model (pLM) ProtT5 as input. Using only single sequences without creating multiple sequence alignments (MSAs), bindEmbed21DL outperformed existing MSA-based methods. Combination with homology-based inference increased performance to F1=48±3% (95% CI) and MCC=0.46±0.04 when merging all three ligand classes into one. All results were confirmed by three independent data sets. Focusing on very reliably predicted residues could complement experimental evidence: For the 25% most strongly predicted binding residues, at least 73% were correctly predicted even when ignoring the problem of missing experimental annotations. The new method bindEmbed21 is fast, simple, and broadly applicable - neither using structure nor MSAs. Thereby, it found binding residues in over 42% of all human proteins not otherwise implied in binding and predicted about 6% of all residues as binding to metal ions, nucleic acids, or small molecules.

Download Full-text

Protein embeddings and deep learning predict binding residues for various ligand classes

Scientific Reports ◽

10.1038/s41598-021-03431-4 ◽

2021 ◽

Vol 11 (1) ◽

Author(s):

Maria Littmann ◽

Michael Heinzinger ◽

Christian Dallago ◽

Konstantin Weissenow ◽

Burkhard Rost

Keyword(s):

Nucleic Acids ◽

Metal Ions ◽

Small Molecules ◽

Protein Function ◽

Language Model ◽

Data Sets ◽

Sequence Alignments ◽

Multiple Sequence ◽

Human Proteins ◽

Binding Residues

AbstractOne important aspect of protein function is the binding of proteins to ligands, including small molecules, metal ions, and macromolecules such as DNA or RNA. Despite decades of experimental progress many binding sites remain obscure. Here, we proposed bindEmbed21, a method predicting whether a protein residue binds to metal ions, nucleic acids, or small molecules. The Artificial Intelligence (AI)-based method exclusively uses embeddings from the Transformer-based protein Language Model (pLM) ProtT5 as input. Using only single sequences without creating multiple sequence alignments (MSAs), bindEmbed21DL outperformed MSA-based predictions. Combination with homology-based inference increased performance to F1 = 48 ± 3% (95% CI) and MCC = 0.46 ± 0.04 when merging all three ligand classes into one. All results were confirmed by three independent data sets. Focusing on very reliably predicted residues could complement experimental evidence: For the 25% most strongly predicted binding residues, at least 73% were correctly predicted even when ignoring the problem of missing experimental annotations. The new method bindEmbed21 is fast, simple, and broadly applicable—neither using structure nor MSAs. Thereby, it found binding residues in over 42% of all human proteins not otherwise implied in binding and predicted about 6% of all residues as binding to metal ions, nucleic acids, or small molecules.

Download Full-text

Benchmarking Statistical Multiple Sequence Alignment

10.1101/304659 ◽

2018 ◽

Cited By ~ 1

Author(s):

Michael Nute ◽

Ehsan Saleh ◽

Tandy Warnow

Keyword(s):

Sequence Alignment ◽

Multiple Sequence Alignment ◽

Structural Alignment ◽

Estimation Method ◽

Simulated Data ◽

Protein Sequences ◽

Data Sets ◽

Sequence Alignments ◽

Multiple Sequence ◽

Simulated Data Sets

AbstractThe estimation of multiple sequence alignments of protein sequences is a basic step in many bioinformatics pipelines, including protein structure prediction, protein family identification, and phylogeny estimation. Statistical co-estimation of alignments and trees under stochastic models of sequence evolution has long been considered the most rigorous technique for estimating alignments and trees, but little is known about the accuracy of such methods on biological benchmarks. We report the results of an extensive study evaluating the most popular protein alignment methods as well as the statistical co-estimation method BAli-Phy on 1192 protein data sets from established benchmarks as well as on 120 simulated data sets. Our study (which used more than 230 CPU years for the BAli-Phy analyses alone) shows that BAli-Phy is dramatically more accurate than the other alignment methods on the simulated data sets, but is among the least accurate on the biological benchmarks. There are several potential causes for this discordance, including model misspecification, errors in the reference alignments, and conflicts between structural alignment and evolutionary alignments; future research is needed to understand the most likely explanation for our observations. multiple sequence alignment, BAli-Phy, protein sequences, structural alignment, homology

Download Full-text

SPOT-1D-LM: Reaching Alignment-profile-based Accuracy in Predicting Protein Secondary and Tertiary Structural Properties without Alignment.

10.1101/2021.10.16.464622 ◽

2021 ◽

Author(s):

Jaspreet Singh ◽

Kuldip Paliwal ◽

Jaswinder Singh ◽

Yaoqi Zhou

Keyword(s):

Structural Properties ◽

Solvent Accessibility ◽

Protein Structures ◽

Language Models ◽

Sequence Information ◽

Sequence Alignments ◽

Multiple Sequence ◽

Multiple Sequence Alignments ◽

Structural And Functional Properties ◽

Sequence Profiles

Protein language models have emerged as an alternative to multiple sequence alignment for enriching sequence information and improving downstream prediction tasks such as biophysical, structural, and functional properties. Here we show that a combination of traditional one-hot encoding with the embeddings from two different language models (ProtTrans and ESM-1b) allows a leap in accuracy over single-sequence based techniques in predicting protein 1D secondary and tertiary structural properties, including backbone torsion angles, solvent accessibility and contact numbers. This large improvement leads to an accuracy comparable to or better than the current state-of-the-art techniques for predicting these 1D structural properties based on sequence profiles generated from multiple sequence alignments. The high-accuracy prediction in both secondary and tertiary structural properties indicates that it is possible to make highly accurate prediction of protein structures without homologous sequences, the remaining obstacle in the post AlphaFold2 era.

Download Full-text

Contrastive learning on protein embeddings enlightens midnight zone at lightning speed

10.1101/2021.11.14.468528 ◽

2021 ◽

Author(s):

Michael Heinzinger ◽

Maria Littmann ◽

Ian Sillitoe ◽

Nicola Bordin ◽

Christine Orengo ◽

...

Keyword(s):

Structure Prediction ◽

Sequence Similarity ◽

3D Structure ◽

Three Dimensional ◽

Hierarchical Classification ◽

Language Models ◽

Sequence Alignments ◽

Sequence Comparisons ◽

Multiple Sequence ◽

3D Structures

Thanks to the recent advances in protein three-dimensional (3D) structure prediction, in particular through AlphaFold 2 and RoseTTAFold, the abundance of protein 3D information will explode over the next year(s). Expert resources based on 3D structures such as SCOP and CATH have been organizing the complex sequence-structure-function relations into a hierarchical classification schema. Experimental structures are leveraged through multiple sequence alignments, or more generally through homology-based inference (HBI) transferring annotations from a protein with experimentally known annotation to a query without annotation. Here, we presented a novel approach that expands the concept of HBI from a low-dimensional sequence-distance lookup to the level of a high-dimensional embedding-based annotation transfer (EAT). Secondly, we introduced a novel solution using single protein sequence representations from protein Language Models (pLMs), so called embeddings (Prose, ESM-1b, ProtBERT, and ProtT5), as input to contrastive learning, by which a new set of embeddings was created that optimized constraints captured by hierarchical classifications of protein 3D structures. These new embeddings (dubbed ProtTucker) clearly improved what was historically referred to as threading or fold recognition. Thereby, the new embeddings enabled the intrusion into the midnight zone of protein comparisons, i.e., the region in which the level of pairwise sequence similarity is akin of random relations and therefore is hard to navigate by HBI methods. Cautious benchmarking showed that ProtTucker reached much further than advanced sequence comparisons without the need to compute alignments allowing it to be orders of magnitude faster. Code is available at https://github.com/Rostlab/EAT .

Download Full-text

Protein language model embeddings for fast, accurate, alignment-free protein structure prediction

10.1101/2021.07.31.454572 ◽

2021 ◽

Author(s):

Konstantin Weissenow ◽

Michael Heinzinger ◽

Burkhard Rost

Keyword(s):

Protein Structure ◽

Structure Prediction ◽

Prediction Models ◽

Language Model ◽

Structural Features ◽

Language Models ◽

Evolutionary Information ◽

Major Advance ◽

Sequence Alignments ◽

Multiple Sequence

All state-of-the-art (SOTA) protein structure predictions rely on evolutionary information captured in multiple sequence alignments (MSAs), primarily on evolutionary couplings (co-evolution). Such information is not available for all proteins and is computationally expensive to generate. Prediction models based on Artificial Intelligence (AI) using only single sequences as input are easier and cheaper but perform so poorly that speed becomes irrelevant. Here, we described the first competitive AI solution exclusively inputting embeddings extracted from pre-trained protein Language Models (pLMs), namely from the transformer pLM ProtT5, from single sequences into a relatively shallow (few free parameters) convolutional neural network (CNN) trained on inter-residue distances, i.e. protein structure in 2D. The major advance originated from processing the attention heads learned by ProtT5. Although these models required at no point any MSA, they matched the performance of methods relying on co-evolution. Although not reaching the very top, our lean approach came close at substantially lower costs thereby speeding up development and each future prediction. By generating protein-specific rather than family-averaged predictions, these new solutions could distinguish between structural features differentiating members of the same family of proteins with similar structure predicted alike by all other top methods.

Download Full-text

Proteochemometric Models Using Multiple Sequence Alignments and a Subword Segmented Masked Language Model

10.26434/chemrxiv.14604720 ◽

2021 ◽

Author(s):

Héléna Alexandra Gaspar ◽

Mohamed Ahmed ◽

Thomas Edlich ◽

Benedek Fabian ◽

Zsolt Varszegi ◽

...

Keyword(s):

Language Processing ◽

Language Model ◽

Predictive Performance ◽

Language Models ◽

Cytochrome P450 Enzymes ◽

Sequence Alignments ◽

Multiple Sequence ◽

Multiple Sequence Alignments ◽

Combine Information ◽

Protein Datasets

<div>Proteochemometric (PCM) models of protein-ligand activity combine information from both the ligands and the proteins to which they bind. Several methods inspired by the field of natural language processing (NLP) have been proposed to represent protein sequences. </div><div>Here, we present PCM benchmark results on three multi-protein datasets: protein kinases, rhodopsin-like GPCRs (ChEMBL binding and functional assays), and cytochrome P450 enzymes. Keeping ligand descriptors fixed, we evaluate our own protein embeddings based on subword-segmented language models trained on mammalian sequences against pre-existing NLP-based descriptors, protein-protein similarity matrices derived from multiple sequence alignments (MSA), dummy protein one-hot encodings, and a combination of NLP-based and MSA-based descriptors. Our results show that performance gains over one-hot encodings are small and combining NLP-based and MSA-based descriptors increases predictive performance consistently across different splitting strategies. This work has been presented at the 3rd RSC-BMCS / RSC-CICAG Artificial Intelligence in Chemistry in September 2020.</div>

Download Full-text