The DOI matches its Crossref record; 130 of the 133 references it cites were found in a registry or an index; none of the 53 sources its claims cite is checked for support yet.
Learning the protein language: Evolution, structure, and function
Open accessAbstract
Language models have recently emerged as a powerful machine-learning approach for distilling information from massive protein sequence databases. From readily available sequence data alone, these models discover evolutionary, structural, and functional organization across protein space. Using language models, we can encode amino-acid sequences into distributed vector representations that capture their structural and functional properties, as well as evaluate the evolutionary fitness of sequence variants. We discuss recent advances in protein language modeling and their applications to downstream protein property prediction problems. We then consider how these models can be enriched with prior biological knowledge and introduce an approach for encoding protein structural knowledge into the learned representations. The knowledge distilled by these models allows us to improve downstream function prediction through transfer learning. Deep protein language models are revolutionizing protein biology. They suggest new ways to approach protein and therapeutic design. However, further developments are needed to encode strong biological priors into protein language models and to increase their accessibility to the broader community.
Overview
AI snapshot
Written from the paper’s full text.
TL;DR
Protein language models learn evolutionary, structural, and functional organization from large sequence databases, and adding structure supervision improves downstream prediction through transfer learning. The authors present a multi-task LSTM trained on 76 million sequences plus 28,000 structures and show gains in structural similarity, transmembrane prediction, and sequence-to-phenotype prediction.
Key findings
- A multi-task LSTM (MT-LSTM) trained with masked language modeling, residue-residue contact prediction, and SCOP structural similarity prediction learned embeddings that “dramatically outperform” sequence comparison methods and even outperform TMalign for structural similarity; structure supervision was critical for fine-grained organization, and MT-LSTM outperformed the prior SSA-LSTM.
- For transmembrane region prediction, BiLSTM+CRF models using MT-LSTM embeddings significantly outperformed DLM-LSTM embeddings (paired t-test, p=0.044) and also outperformed existing predictors and SSA-LSTM.
- For sequence-to-phenotype prediction on deep mutational scanning datasets, MT-LSTM embeddings with Gaussian process regression achieved an average Spearman correlation of 0.65, significantly outperforming ECNet’s 0.60 (paired t-test, p=0.006).
- The paper notes that language models predicted SARS-CoV-2 spike mutations with high antibody-escape potential, but as of publication viral escape had not yet been experimentally verified.
Methods
- The authors trained a three-layered bidirectional LSTM with skip connections and 1024 hidden units in each direction per layer. The MT-LSTM combined three losses: masked language modeling (10% of amino acids randomly masked), residue-residue contact prediction (contacts defined as residues within 8 Å), and SCOP structural similarity prediction via ordinal regression; MT-LSTM loss weights were λmasked=0.5, λcontact=0.9, λsimilarity=0.1, while DLM-LSTM used only λmasked=1.
- Training data were 76,215,872 UniRef90 sequences (July 2018) and 28,010 SCOPe ASTRAL 2.06 protein sequences with structures, split into 22,408 training and 5,602 testing sequences. Models were trained for 1,000,000 parameter updates with Adam at learning rate 0.0001 on a single NVIDIA V100 GPU; training took roughly 13 days for DLM-LSTM and 51 days for MT-LSTM.
- Evaluations included structural similarity on 100,000 randomly sampled SCOPe ASTRAL test pairs, transmembrane prediction on the TOPCONS2 dataset of 5,154 proteins using 10-fold cross validation, and sequence-to-phenotype prediction on deep mutational scanning datasets using 5-fold cross validation and Gaussian process regression with an RBF kernel.
Limitations
- The paper states that further developments are needed to encode strong biological priors into protein language models and to increase their accessibility to the broader community; it also states that current models make heavily simplified phylogenetic assumptions, treating sequences as independent draws from a star phylogeny.
- The paper states practical limitations including high compute costs, murky design guidelines, ad hoc model development, quadratic scaling of transformers with sequence length (requiring long sequences to be excluded or truncated), and poor learning of long-range dependencies by RNNs.
- Evident from the text: the new results come from a single MT-LSTM architecture and specific benchmarks, and the SARS-CoV-2 escape predictions had not been experimentally verified as of publication.
Model-written summary. Check it against the paper before relying on it.
Integrity signals
Registry
Cites
Statistics
Verification
Claims the paper makes
0 of 53 cited sources checked · 0 supported · 0 flagged
A result the paper reports from its own data — it cites no other source for it, so there is none to check. The quote shows it is in the paper.
In the paper
“We also find that increasing model size improves transfer learning performance.”
§ Learning the Protein Language: Evolution, Structure and Function > Transfer learning improves downstream applications
Tests✓ Asserted: passes✓ Checkable: passes✓ Load-bearing: passes✓ Based: passesA result the paper reports from its own data — it cites no other source for it, so there is none to check. The quote shows it is in the paper.
In the paper
“However, this organization is improved when we include structure supervision in the language model training (Figure 4b).”
§ Learning the Protein Language: Evolution, Structure and Function > Multi-task language models capture the semantic organization of proteins
Tests✓ Asserted: passes✓ Checkable: passes✓ Load-bearing: passes✓ Based: passesA result the paper reports from its own data — it cites no other source for it, so there is none to check. The quote shows it is in the paper.
In the paper
“Furthermore, representations learned by our MT-LSTM model significantly outperform (paired t-test, p=0.044) the embeddings learned by our DLM-LSTM model on this application (Figure 5b).”
§ Learning the Protein Language: Evolution, Structure and Function > Transfer learning improves downstream applications
Tests✓ Asserted: passes✓ Checkable: passes✓ Load-bearing: passes✓ Based: passesIn the paper
“Furthermore, the multi-task learning approach outperforms a two-step learning approach presented previously (SSA-LSTM)27.”
§ Learning the Protein Language: Evolution, Structure and Function > Multi-task language models capture the semantic organization of proteins
What the sources say
- Not checked yetcites [27] Bepler T & Berger B Learning protein sequence embeddings using information from structure. in International Conference on Learning Representations (2019). PMC12947628 41768824 source
Tests✓ Asserted: passes✓ Checkable: passes✓ Load-bearing: passes✓ Based: passesA result the paper reports from its own data — it cites no other source for it, so there is none to check. The quote shows it is in the paper.
In the paper
“Because we embed sequences into a semantic representation space, we can find structurally related proteins even though their sequences are not closely related (Figure 4c, Supplemental Table 1).”
§ Learning the Protein Language: Evolution, Structure and Function > Multi-task language models capture the semantic organization of proteins
Tests✓ Asserted: passes✓ Checkable: passes✓ Load-bearing: passes✓ Based: passesA result the paper reports from its own data — it cites no other source for it, so there is none to check. The quote shows it is in the paper.
In the paper
“We observe that even though the DLM-LSTM model was trained using only sequence information, protein sequences still organize roughly by structure in embedding space.”
§ Learning the Protein Language: Evolution, Structure and Function > Multi-task language models capture the semantic organization of proteins
Tests✓ Asserted: passes✓ Checkable: passes✓ Load-bearing: passes✓ Based: passesA result the paper reports from its own data — it cites no other source for it, so there is none to check. The quote shows it is in the paper.
In the paper
“Because TMscore is not symmetric, we calculate TMscore for both comparison directions and average them together for each protein pair. We found this outperformed other methods of combining the two scores.”
§ Learning the Protein Language: Evolution, Structure and Function > Methods > Protein structural similarity prediction evaluation
Strength:plain· not yet calibratedCausal:direct· not yet calibratedTests✓ Asserted: passes✓ Checkable: passes? Load-bearing: unclear✓ Based: passesIn the paper
“The key insight was that the sequence representations (vector embeddings) learned by the language model were powerful features for solving other prediction problems.”
§ Learning the Protein Language: Evolution, Structure and Function > Transfer learning improves downstream applications
What the sources say
- Not checked yetcites [27] Bepler T & Berger B Learning protein sequence embeddings using information from structure. in International Conference on Learning Representations (2019). PMC12947628 41768824 source
Tests✓ Asserted: passes✓ Checkable: passes✓ Load-bearing: passes✓ Based: passesIn the paper
“Interestingly, because so many sequences are available, these models continue to benefit from increased size 32.”
§ Learning the Protein Language: Evolution, Structure and Function > Protein language models distill information from massive protein sequence databases
What the sources say
- Not checked yetcites [32] Rives A Biological structure and function emerge from scaling unsupervised learning to 250 million protein sequences. bioRxiv (2019). PMC8053943 33876751 source
Strength:plain· not yet calibratedCausal:direct· not yet calibratedTests✓ Asserted: passes✓ Checkable: passes? Load-bearing: unclear✓ Based: passesIn the paper
“Learned protein representations improve predictive ability for downstream prediction problems through transfer learning (Figure 5a).”
§ Learning the Protein Language: Evolution, Structure and Function > Transfer learning improves downstream applications
What the sources say
- Not checked yetcites [27] Bepler T & Berger B Learning protein sequence embeddings using information from structure. in International Conference on Learning Representations (2019). PMC12947628 41768824 source
Tests✓ Asserted: passes✓ Checkable: passes✓ Load-bearing: passes✓ Based: passes- Rests on an assumption: Many problems in biology are small data problems; quantitative protein characterization assays are rarely high throughput and methods are needed that can generalize given only 10s to 100s of experimental measurements.
- Rests on an assumption: We are often interested in extrapolating from data collected over a small region of protein sequence space to other sequences, often with little to no homology.
- It may state two separate claims as one. 13
- The quote only partly states this claim. 1
- Probably not a substantial claim. 8
A result the paper reports from its own data — it cites no other source for it, so there is none to check. The quote shows it is in the paper.
In the paper
“We find that our learned semantic embeddings dramatically outperform the sequence comparison methods and even outperform structure comparison with TMalign when predicting structural similarity.”
§ Learning the Protein Language: Evolution, Structure and Function > Multi-task language models capture the semantic organization of proteins
Tests✓ Asserted: passes✓ Checkable: passes✓ Load-bearing: passes✓ Based: passesIt may state two separate claims as one.
A result the paper reports from its own data — it cites no other source for it, so there is none to check. The quote shows it is in the paper.
In the paper
“The knowledge distilled by these models allows us to improve downstream function prediction through transfer learning.”
- Also stated in: § Multi-task language models capture the semantic organization of proteins · boosted
“Here, we will demonstrate that training protein language models with self-supervision on a large amount of natural sequence data and with structure supervision on a smaller set of sequence, structure pairs enriches the learned representations and translates into improvements in downstream prediction problems (Figure 3).”
- Also stated in: § Transfer learning improves downstream applications · plain
“Learned protein representations improve predictive ability for downstream prediction problems through transfer learning (Figure 5a).”
§ Learning the Protein Language: Evolution, Structure and Function > Abstract
Tests✓ Asserted: passes✓ Checkable: passes✓ Load-bearing: passes✓ Based: passesIt may state two separate claims as one.
- Rests on an assumption: Many problems in biology are small data problems; quantitative protein characterization assays are rarely high throughput and methods are needed that can generalize given only 10s to 100s of experimental measurements.
- Rests on an assumption: We are often interested in extrapolating from data collected over a small region of protein sequence space to other sequences, often with little to no homology.
- Also stated in: § Multi-task language models capture the semantic organization of proteins · boosted
A result the paper reports from its own data — it cites no other source for it, so there is none to check. The quote shows it is in the paper.
In the paper
“We find that we can predict the phenotypes of unobserved sequence variants across datasets better than existing methods (Figure 5c).”
§ Learning the Protein Language: Evolution, Structure and Function > Transfer learning improves downstream applications
Tests✓ Asserted: passes✓ Checkable: passes✓ Load-bearing: passes✓ Based: passesIt may state two separate claims as one. The quote only partly states this claim.
- Rests on an assumption: Many problems in biology are small data problems; quantitative protein characterization assays are rarely high throughput and methods are needed that can generalize given only 10s to 100s of experimental measurements.
- Rests on an assumption: We are often interested in extrapolating from data collected over a small region of protein sequence space to other sequences, often with little to no homology.
A result the paper reports from its own data — it cites no other source for it, so there is none to check. The quote shows it is in the paper.
In the paper
“We find that the BiLSTM+CRFs with our new embeddings (DLM-LSTM and MT-LSTM) outperform existing transmembrane predictors and a BiLSTM+CRF using our previous smaller embedding model (SSA-LSTM).”
§ Learning the Protein Language: Evolution, Structure and Function > Transfer learning improves downstream applications
Tests✓ Asserted: passes✓ Checkable: passes✓ Load-bearing: passes✓ Based: passesIt may state two separate claims as one.
A result the paper reports from its own data — it cites no other source for it, so there is none to check. The quote shows it is in the paper.
In the paper
“Interestingly, we observe that the structural supervision component is critical for learning well organized embeddings at a fine-grained level, because the DLM-LSTM representations alone do not perform well at this task (Supplemental Table 1).”
§ Learning the Protein Language: Evolution, Structure and Function > Multi-task language models capture the semantic organization of proteins
Tests✓ Asserted: passes✓ Checkable: passes✓ Load-bearing: passes✓ Based: passesIt may state two separate claims as one.
In the paper
“This parallels the general trend in natural language processing, where the number of parameters, rather than specific architectural choices, is the best indicator of model performance 52.”
§ Learning the Protein Language: Evolution, Structure and Function > Protein language models distill information from massive protein sequence databases
What the sources say
- Not checked yetcites [52] Kaplan J Scaling Laws for Neural Language Models. arXiv [cs.LG] (2020).
Strength:plain· not yet calibratedCausal:correlational· not yet calibratedTests✓ Asserted: passes✓ Checkable: passes✗ Load-bearing: fails✓ Based: passesProbably not a substantial claim. The paper’s conclusions may not depend on it.
None of the sources cited for this claim have been checked yet.
In the paper
“Although structure is more informative of function than sequence 57,58 and structure is encoded by sequence, predicting structure remains hard, particularly due to the relative paucity of structural relative to sequence data.”
§ Learning the Protein Language: Evolution, Structure and Function > Supervision encodes biological meaning
What the sources say
- Not checked yetcites [57] Zhang C & Kim S-H Overview of structural genomics: from structure to function. Curr. Opin. Chem. Biol 7, 28– 32 (2003). 12547423 source
- Not checked yetcites [58] Shin DH Structure-based inference of molecular functions of proteins of unknown function from Berkeley Structural Genomics Center. J. Struct. Funct. Genomics 8, 99– 105 (2007). 17764033 source
Tests✓ Asserted: passes✓ Checkable: passes✓ Load-bearing: passes✓ Based: passesIt may state two separate claims as one.
None of the sources cited for this claim have been checked yet.
In the paper
“They also learn summary representations, powerful features that can be used to better capture sequence relationships and link sequence to function via transfer learning 27,28,32-35.”
§ Learning the Protein Language: Evolution, Structure and Function > Introduction
What the sources say
- Not checked yetcites [27] Bepler T & Berger B Learning protein sequence embeddings using information from structure. in International Conference on Learning Representations (2019). PMC12947628 41768824 source
- Not checked yetcites [28] Alley EC, Khimulya G, Biswas S, AlQuraishi M & Church GM Unified rational protein engineering with sequence-based deep representation learning. Nat. Methods 16, 1315– 1322 (2019). 31636460 PMC7067682 source
- Not checked yetcites [32] Rives A Biological structure and function emerge from scaling unsupervised learning to 250 million protein sequences. bioRxiv (2019). PMC8053943 33876751 source
- Not checked yetcites [33] Rao R Evaluating Protein Transfer Learning with TAPE. Adv. Neural Inf. Process. Syst 32, 9689– 9701 (2019). 33390682 PMC7774645 source
- Not matchedcites [34] Luo Y Evolutionary Context-Integrated Deep Sequence Modeling for Protein Engineering. in Research in Computational Molecular Biology 261– 263 (Springer International Publishing, 2020).
Couldn’t be matched to a paper in Crossref, so it can’t be checked.
- Not checked yetcites [35] Hie B, Bryson BD & Berger B Leveraging Uncertainty in Machine Learning Accelerates Biological Discovery and Design. Cell Syst 11, 461– 477.e9 (2020). 33065027 source
Tests✓ Asserted: passes✓ Checkable: passes✓ Load-bearing: passes✓ Based: passesIt may state two separate claims as one.
None of the sources cited for this claim have been checked yet.
In the paper
“Although it is generally believed that proteins with similar sequences form similar structures, there are also interesting examples of highly similar protein sequences having radically different structures and functions 60,61 and of sequences that can form multiple folds 62.”
§ Learning the Protein Language: Evolution, Structure and Function > Supervision encodes biological meaning
What the sources say
- Not checked yetcites [60] Kosloff M & Kolodny R Sequence-similar, structure-dissimilar protein pairs in the PDB. Proteins 71, 891– 902 (2008). 18004789 PMC2673347 source
- Not checked yetcites [61] Wei KY Computational design of closely related proteins that adopt two well-defined but structurally divergent folds. Proc. Natl. Acad. Sci. U. S. A 117, 7208– 7215 (2020). 32188784 PMC7132107 source
- Not checked yetcites [62] James LC & Tawfik DS Conformational diversity and protein evolution--a 60-year-old hypothesis revisited. Trends Biochem. Sci 28, 361– 368 (2003). 12878003 source
Strength:plain· not yet calibratedCausal:correlational· not yet calibratedTests✓ Asserted: passes✓ Checkable: passes? Load-bearing: unclear✓ Based: passesIt may state two separate claims as one.
The paper states this without a source or data behind it here, so there is nothing to check it against. The quote shows it is in the paper.
In the paper
“For example, unrelated protein sequences can form the same structural fold and, therefore, are semantically similar. However, we cannot deduce this relationship from sequences alone. Supervision is required to learn that these sequences belong to the same semantic category.”
§ Learning the Protein Language: Evolution, Structure and Function > Supervision encodes biological meaning
Tests✓ Asserted: passes✓ Checkable: passes✓ Load-bearing: passes✓ Based: passesIt may state two separate claims as one.
A conclusion the paper draws from its own results — it cites no other source for it, so there is none to check. The quote shows it is in the paper.
In the paper
“These phenomena suggest that there are aspects of protein biology that may not be discoverable by statistical sequence models alone. Supervision that represents known protein structure, function, and other prior knowledge may be necessary to encode distant sequence relationships into learned embeddings.”
§ Learning the Protein Language: Evolution, Structure and Function > Supervision encodes biological meaning
Tests✓ Asserted: passes✓ Checkable: passes✓ Load-bearing: passes✓ Based: passesIt may state two separate claims as one.
- Rests on an assumption: Proteins with similar structures and functions still exist below 30% sequence identity, but they can no longer be detected from sequence similarity alone and it is unclear whether their functions are conserved.
- Rests on an assumption: Evolutionary innovation requires that protein function can change with only a few mutations.
None of the sources cited for this claim have been checked yet.
In the paper
“Language models learn the probability of a sequence occurring and this can be directly applied to predict the fitness of sequence mutations 29-31 .”
§ Learning the Protein Language: Evolution, Structure and Function > Introduction
What the sources say
- Not checked yetcites [29] Riesselman AJ, Ingraham JB & Marks DS Deep generative models of genetic variation capture the effects of mutations. Nat. Methods 15, 816– 822 (2018). 30250057 PMC6693876 source
- Not checked yetcites [30] Hie B, Zhong E, Bryson B & Berger B Learning mutational semantics. Adv. Neural Inf. Process. Syst 33, (2020).
- Not checked yetcites [31] Hie B, Zhong ED, Berger B & Bryson B Learning the language of viral evolution and escape. Science 371, 284– 288 (2021). 33446556 source
Tests✓ Asserted: passes✓ Checkable: passes✓ Load-bearing: passes✓ Based: passesIt may state two separate claims as one.
A conclusion the paper draws from its own results — it cites no other source for it, so there is none to check. The quote shows it is in the paper.
In the paper
“Language models capture complex relationships between residues in protein sequences by condensing information from enormous protein sequence databases.”
§ Learning the Protein Language: Evolution, Structure and Function > Conclusions and Perspectives: Strong biological priors are key to improving protein language models
Tests✓ Asserted: passes✓ Checkable: passes✓ Load-bearing: passes✗ Based: failsProbably not a substantial claim. The paper may give no reason why it holds.
A conclusion the paper draws from its own results — it cites no other source for it, so there is none to check. The quote shows it is in the paper.
In the paper
“Fine-tuning pre-trained models can help with this problem but does not fully resolve it.”
§ Learning the Protein Language: Evolution, Structure and Function > Conclusions and Perspectives: Strong biological priors are key to improving protein language models
Tests✓ Asserted: passes✓ Checkable: passes✓ Load-bearing: passes✓ Based: passesProbably not a substantial claim. It may state two separate claims as one.
- Rests on an assumption: For individual protein families or application-specific protein datasets, brute force hyperparameter search is the gold standard for selecting model architectures and number of parameters.
A conclusion the paper draws from its own results — it cites no other source for it, so there is none to check. The quote shows it is in the paper.
In the paper
“High compute costs and murky design guidelines mean that developing new models is often an expensive, time consuming, and ad hoc process.”
§ Learning the Protein Language: Evolution, Structure and Function > Conclusions and Perspectives: Strong biological priors are key to improving protein language models
Strength:hedged· not yet calibratedCausal:direct· not yet calibratedTests✓ Asserted: passes✓ Checkable: passes✗ Load-bearing: fails✓ Based: passesProbably not a substantial claim. The paper’s conclusions may not depend on it.
None of the sources cited for this claim could be matched to a paper in Crossref, so there is nothing to check it against.
In the paper
“To compare methods, we follow Luo et al and perform 5-fold cross validation on each deep mutational scanning dataset 34”
§ Learning the Protein Language: Evolution, Structure and Function > Methods > Sequence-to-phenotype prediction and evaluation
What the sources say
- Not matchedcites Luo et al
Couldn’t be matched to a paper in Crossref, so it can’t be checked.
- Not matchedcites [34] Luo Y Evolutionary Context-Integrated Deep Sequence Modeling for Protein Engineering. in Research in Computational Molecular Biology 261– 263 (Springer International Publishing, 2020).
Couldn’t be matched to a paper in Crossref, so it can’t be checked.
Load-bearingStrength:plain· not yet calibratedTests✓ Asserted: passes✓ Checkable: passes✓ Load-bearing: passes✓ Based: passesProbably not a substantial claim.
- Not matched
What the paper did — it cites no source for this, so there is none to check. The quote shows it is in the paper.
In the paper
“We compare methods by 10-fold cross validation.”
§ Learning the Protein Language: Evolution, Structure and Function > Methods > Transmembrane region prediction training and evaluation
Load-bearingStrength:plain· not yet calibratedTests✓ Asserted: passes✓ Checkable: passes✓ Load-bearing: passes✓ Based: passesProbably not a substantial claim.
None of the sources cited for this claim have been checked yet.
In the paper
“New linear complexity attention mechanisms may help to alleviate this limitation 95,96.”
§ Learning the Protein Language: Evolution, Structure and Function > Conclusions and Perspectives: Strong biological priors are key to improving protein language models
What the sources say
- Not matchedcites [95] Choromanski KM Rethinking Attention with Performers. in International Conference on Learning Representations (2020).
Couldn’t be matched to a paper in Crossref, so it can’t be checked.
- Not checked yetcites [96] Wang S, Li BZ, Khabsa M, Fang H & Ma H Linformer: Self-Attention with Linear Complexity. (2020). source
Strength:hedged· not yet calibratedCausal:conditional· not yet calibratedTests✓ Asserted: passes✓ Checkable: passes✗ Load-bearing: fails✓ Based: passesProbably not a substantial claim. The paper’s conclusions may not depend on it.
- Rests on an assumption: Transformers scale quadratically with sequence length.
- Not matched
In the paper
“Recently, Luo et al. presented a method for combining language model-based representations with local evolutionary context-based representations (ECNet) and demonstrated that these representations were powerful for sequence-to-phenotype mapping on a panel of deep mutational scanning datasets 34.”
§ Learning the Protein Language: Evolution, Structure and Function > Transfer learning improves downstream applications
What the sources say
- Not matchedcites [34] Luo Y Evolutionary Context-Integrated Deep Sequence Modeling for Protein Engineering. in Research in Computational Molecular Biology 261– 263 (Springer International Publishing, 2020).
Couldn’t be matched to a paper in Crossref, so it can’t be checked.
Strength:boosted· not yet calibratedCausal:correlational· not yet calibratedTests✓ Asserted: passes✓ Checkable: passes✗ Load-bearing: fails✓ Based: passesProbably not a substantial claim. The paper’s conclusions may not depend on it. It may state two separate claims as one.
- Not matched
The paper relies on this without support.
Many problems in biology are small data problems; quantitative protein characterization assays are rarely high throughput and methods are needed that can generalize given only 10s to 100s of experimental measurements.
“A key challenge in biology is that many problems are small data problems. Quantitative protein characterization assays are rarely high throughput and methods are needed that can generalize given only 10s to 100s of experimental measurements.”
The paper relies on this without support.
We are often interested in extrapolating from data collected over a small region of protein sequence space to other sequences, often with little to no homology.
“Furthermore, we are often interested in extrapolating from data collected over a small region of protein sequence space to other sequences, often with little to no homology.”
The paper relies on this without support.
Proteins with similar structures and functions still exist below 30% sequence identity, but they can no longer be detected from sequence similarity alone and it is unclear whether their functions are conserved.
“Often called the “twilight zone” of protein sequence homology, proteins with similar structures and functions still exist below this level, but they can no longer be detected from sequence similarity alone and it is unclear whether their functions are conserved.”
The paper relies on this without support.
Evolutionary innovation requires that protein function can change with only a few mutations.
“Evolutionary innovation requires that protein function can change with only a few mutations.”
The paper relies on this without support.
Proteins exist in physical space.
“Proteins are objects that exist in physical space.”
The paper relies on this without support.
Many fundamental evolutionary processes that give rise to the diversity of protein sequences observed today are understood.
“Similarly, we understand many of the fundamental evolutionary processes that give rise to the diversity of protein sequences observed today.”
The paper relies on this without support.
Physics and evolution are key properties of proteins.
“These two elements, physics and evolution, are the key properties of proteins and our models might benefit from being structured explicitly to incorporate evolutionary and physics-based inductive biases.”
The paper relies on this without support.
Biological sequences are not natural language.
“Biological sequences are not natural language”
The paper relies on this without support.
Filtering sequences based on percent identity is a conventional crude address to the phylogenetic assumption problem.
“Conventionally, this problem is crudely addressed by filtering sequences based on percent identity.”
The paper relies on this without support.
At the upper limit of large natural protein databases, larger models continue to yield improved performance.
“At the upper limit of large natural protein databases, larger models continue to yield improved performance.”
The paper relies on this without support.
For individual protein families or application-specific protein datasets, brute force hyperparameter search is the gold standard for selecting model architectures and number of parameters.
“However, for individual protein families or other application specific protein data sets, the gold standard is to select model architectures and number of parameters via brute force hyperparameter search methods.”
The paper relies on this without support.
Transformers scale quadratically with sequence length.
“Transformers scale quadratically with sequence length”
The paper relies on this without support.
Recurrent neural networks scale linearly with sequence length.
“This problem is less extreme for recurrent neural networks, which scale linearly with sequence length”
The paper relies on this without support.
Increasing model size, compute power, and dataset size will continue to improve performance of protein language models.
“Increasing model size, compute power, and dataset size will only continue to improve performance of protein language models.”
The paper relies on this without support.
Protein language models are transforming computational protein biology today due to their ease of use and widespread applicability.
“Already, these methods are transforming computational protein biology today due to their ease of use and widespread applicability.”
The paper relies on this without support.
Augmenting language models with protein-specific properties such as structure and function offers an already successful route towards richer representations and novel biology.
“Furthermore, augmenting language models with protein specific properties such as structure and function offers one already successful route towards even richer representations and novel biology.”
Claims other papers make about this one
Each is checked against this paper’s own text: does it say what the citing paper says it does?
In the citing paper
“Instead of using multi-headed self-attention (MHSA), we adopt the Gated Attention Unit (GAU) (line 8 in Algorithm 1), which has shown great promise as an alternative to MHSAs with smaller memory consumption and faster convergence rate (25).”
From High-resolution de novo structure prediction from primary sequence (2022) · cites this paper as [25]
In the citing paper
“The original GAU in (25) also suggests that relu 2 (•) performs better compared to softmax in terms of both computation speed and convergence rate.”
From High-resolution de novo structure prediction from primary sequence (2022) · cites this paper as [25]
Citations
Cites
Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A. N., Kaiser, L., & Polosukhin, I. (2017) · arXiv · doi:10.48550/arxiv.1706.03762 · Cited by 120,404
Hochreiter, S., & Schmidhuber, J. (1997) · Neural Computation · doi:10.1162/neco.1997.9.8.1735 · Cited by 102,318
Adam: A Method for Stochastic Optimization
Kingma, D. P., & Ba, J. (2014) · arXiv · doi:10.48550/arxiv.1412.6980 · Cited by 98,946
BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding
Devlin, J., Chang, M.-W., Lee, K., & Toutanova, K. (2018) · arXiv · doi:10.48550/arxiv.1810.04805 · Cited by 73,173
BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding
Devlin, J., Chang, M.-W., Lee, K., & Toutanova, K. (2019) · Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers) · doi:10.18653/v1/n19-1423 · Cited by 54,325
Berman, H. M. (2000) · Nucleic Acids Research · doi:10.1093/nar/28.1.235 · Cited by 40,196
ImageNet Large Scale Visual Recognition Challenge
Russakovsky, O., Deng, J., Su, H., Krause, J., Satheesh, S., Ma, S., Huang, Z., Karpathy, A., Khosla, A., Bernstein, M., et al. (2015) · International Journal of Computer Vision · doi:10.1007/s11263-015-0816-y · Cited by 39,380
Language Models are Few-Shot Learners
Brown, T. B., Mann, B., Ryder, N., Subbiah, M., Kaplan, J., Dhariwal, P., Neelakantan, A., Shyam, P., Sastry, G., Askell, A., et al. (2020) · arXiv · doi:10.48550/arxiv.2005.14165 · Cited by 23,372
MRBAYES: Bayesian inference of phylogenetic trees.
Huelsenbeck, J. P., & Ronquist, F. (2001) · Bioinformatics · doi:10.1093/bioinformatics/17.8.754 · Cited by 21,937
GROMACS 4: Algorithms for Highly Efficient, Load-Balanced, and Scalable Molecular Simulation
Hess, B., Kutzner, C., van der Spoel, D., & Lindahl, E. R. (2008) · Journal of Chemical Theory and Computation · doi:10.1021/ct700301q · Cited by 15,830
Flagged first, then by citations
3 of the 133 references couldn’t be matched to a known paper.
Research works printed but not linked: 4
- Radford, 2019, Language models are unsupervised multitask learners, OpenAI blog, 1
- Remmert, 2011, HHblits: lightning-fast iterative protein sequence searching by HMM-HMM alignment, Nat. Methods, 9, 173, 10.1038/nmeth.1818 · DOI · Paper page
- Remmert, 2011, HHblits: lightning-fast iterative protein sequence searching by HMM-HMM alignment, Nat. Methods, 9, 173, 10.1038/nmeth.1818 · DOI · Paper page
- Tsirigos, 2015, The TOPCONS web server for consensus prediction of membrane protein topology and signal peptides, Nucleic Acids Res., 43, W401-7, 10.1093/nar/gkv485 · DOI
Cited by
Sorted by citations
Evolutionary-scale prediction of atomic level protein structure with a language model
Lin, Z., Akin, H., Rao, R., Hie, B., Zhu, Z., Lu, W., Smetanin, N., Verkuil, R., Kabeli, O., Shmueli, Y., et al. (2022) · doi:10.1101/2022.07.20.500902 · Cited by 5,990
On the Opportunities and Risks of Foundation Models
Bommasani, R., Hudson, D. A., Adeli, E., Altman, R., Arora, S., von Arx, S., Bernstein, M. S., Bohg, J., Bosselut, A., Brunskill, E., et al. (2021) · arXiv · doi:10.48550/arxiv.2108.07258 · Cited by 2,676
ProtGPT2 is a deep unsupervised language model for protein design
Ferruz, N., Schmidt, S., & Höcker, B. (2022) · Nature Communications · doi:10.1038/s41467-022-32007-7 · Cited by 795
“Supervised NLP methods, where the input sequences are trained jointly with their labels to produce predictive models, have been applied to various tasks, such as detecting structural similarity or predicting stability 3,4 .”
MethodEfficient evolution of human antibodies from general protein language models
Hie, B., Shanker, V. R., Xu, D., Bruun, T. U. J., Weidenbacher, P. A. B., Tang, S., Wu, W. C. H., Pak, J. E., & Kim, P. S. (2023) · Nature Biotechnology · doi:10.1038/s41587-023-01763-2 · Cited by 472
“1c) to learn patterns that are likely to occur in natural proteins 15–22 .”
High-resolution de novo structure prediction from primary sequence
Wu, R., Ding, F., Wang, R., Shen, R., Zhang, X., Luo, S., Su, C., Wu, Z., Xie, Q., Berger, B., et al. (2022) · doi:10.1101/2022.07.21.500999 · Cited by 392
“The OmegaFold model processes a protein sequence with a stack of GAU layers from ( 25 ) instaed of self-attention layers and multi-layer perceptrons.”
BackgroundMethodSelf-citationBefore and after AlphaFold2: An overview of protein structure prediction
Bertoline, L. M. F., Lima, A. N., Krieger, J. E., & Teixeira, S. K. (2023) · Frontiers in Bioinformatics · doi:10.3389/fbinf.2023.1120370 · Cited by 330
“A synthesis of the recent advances in protein language modeling and their applications to protein prediction problems can be found elsewhere (Bepler and Berger, 2021).”
BackgroundRhea, the reaction knowledgebase in 2022
Bansal, P., Morgat, A., Axelsen, K. B., Muthukrishnan, V., Coudert, E., Aimo, L., Hyka‐Nouspikel, N., Gasteiger, E., Kerhornou, A., Neto, T. B., et al. (2021) · Nucleic Acids Research · doi:10.1093/nar/gkab1016 · Cited by 251
Li, C., Stražar, M., Mohamed, A. M., Ávila-Pacheco, J., Walker, R. L., Lebar, T., Zhao, S., Lockart, J., Dame, A., Thurimella, K., et al. (2024) · Cell · doi:10.1016/j.cell.2024.03.014 · Cited by 269
Rapid in silico directed evolution by a protein language model with EVOLVEpro
Jiang, K., Yan, Z., Bernardo, M. D., Sgrizzi, S. R., Villiger, L., Kayabölen, A., Kim, B., Carscadden, J. K., Hiraizumi, M., Nishimasu, H., et al. (2024) · Science · doi:10.1126/science.adr6006 · Cited by 234
Sledzieski, S., Singh, R., Cowen, L., & Berger, B. (2021) · Cell Systems · doi:10.1016/j.cels.2021.08.010 · Cited by 235
Self-citationSingh, R., Sledzieski, S., Bryson, B. D., Cowen, L., & Berger, B. (2023) · Proceedings of the National Academy of Sciences · doi:10.1073/pnas.2220778120 · Cited by 216
“Specifically, we investigate the pretrained models Prose (30), ESM (76), and ProtBert (27), with default dimensions dt = 6165, 1280, and1024, respectively (SI Appendix, S2).”
Self-citationComprehensive Overview of Bottom-Up Proteomics Using Mass Spectrometry
Jiang, Y., Balaya, R. D. A., Schuster, D., Neely, B. A., Rosano, G. L., Volkmar, N., Momenzadeh, A., Peters-Clarke, T. M., Egbert, S., Kreimer, S. D., et al. (2024) · ACS Measurement Science Au · doi:10.1021/acsmeasuresciau.3c00068 · Cited by 217
Transformer-based deep learning for predicting protein properties in the life sciences
Chandra, A. A., Tünnermann, L., Löfstedt, T., & Gratz, R. (2023) · eLife · doi:10.7554/elife.82819 · Cited by 202
“…model representations are better in general compared, for instance, to MSAs but does show that this is a promising avenue to be explored further since there are recurrent evolutionary relations captured in the representations from such type of language models ( Bepler and Berger, 2021 ).”
BackgroundEmbeddings from protein language models predict conservation and variant effects
(2021) · doi:10.1007/s00439-021-02411-y · Cited by 174
“While we expect that the about twenty approaches that we have compared in several of our recent methods (including the following 13: ESM-1[b|v] Meier et al. 2021; Rives et al. 2021), ProSE[*|DLM|MT] (Bepler and Berger 2019b, 2021), Prot[Albert|Bert|Electra|Vec|T5|T5XL|T5XLNet|T 5XXL] (Elnaggar et al. 2021; Heinzinger…”
InfluentialWu, H., Liu, J., Jiang, T., Zou, Q., Qi, S., Cui, Z., Tiwari, P., & Ding, Y. (2023) · Neural Networks · doi:10.1016/j.neunet.2023.11.018 · Cited by 190
State-specific protein–ligand complex structure prediction with a multiscale deep generative model
Qiao, Z., Nie, W., Vahdat, A., Miller, T. F., & Anandkumar, A. (2024) · Nature Machine Intelligence · doi:10.1038/s42256-024-00792-z · Cited by 210
“Several works have demonstrated that deep generative models are capable of producing de novo-designed proteins with experimentally validated functions, including language models for protein sequence design 9,27,28 and diffusion models for protein backbone generation 29–32 .”
Hie, B., Yang, K., & Kim, P. S. (2022) · Cell Systems · doi:10.1016/j.cels.2022.01.003 · Cited by 151
Protein language model embeddings for fast, accurate, alignment-free protein structure prediction
Weissenow, K., Heinzinger, M., & Rost, B. (2021) · doi:10.1101/2021.07.31.454572 · Cited by 139
FLIP: Benchmark tasks in fitness landscape inference for proteinsPreprint
Dallago, C., Mou, J., Johnston, K. E., Wittmann, B. J., Bhattacharya, N., Goldman, S., Madani, A., & Yang, K. (2021) · bioRxiv (Cold Spring Harbor Laboratory) · doi:10.1101/2021.11.09.467890 · Cited by 142
“The vast majority of representation learning on protein sequences models entire sequences [27, 37, 15, 34].”
BackgroundDevelopment and use of machine learning algorithms in vaccine target selection
Bravi, B. (2024) · npj Vaccines · doi:10.1038/s41541-023-00795-8 · Cited by 145
Scientific Large Language Models: A Survey on Biological & Chemical Domains
Zhang, Q., Ding, K., Lv, T., Wang, X., Yin, Q., Zhang, Y., Jing, Y., Wang, Y., Li, X., Xiang, Z., et al. (2025) · ACM Computing Surveys · doi:10.1145/3715318 · Cited by 216
“Unlike surveys concentrating solely on molecule [53, 239], protein [20, 87, 213, 217], or genome data [44], we aim to provide a comprehensive view of language models for chemical and biological research.”
Machine learning modeling of family wide enzyme-substrate specificity screens
Goldman, S., Das, R., Yang, K. K., & Coley, C. W. (2022) · PLOS Computational Biology · doi:10.1371/journal.pcbi.1009853 · Cited by 120
Genomic language models: opportunities and challenges
Benegas, G., Ye, C., Albors, C., Li, J. C., & Song, Y. S. (2025) · Trends in Genetics · doi:10.1016/j.tig.2024.11.013 · Cited by 157
“…over the past decade, and protein language models (pLMs) trained on these immense data have achieved impressive performance on complex problems such as structure prediction [5] and variant effect prediction [6, 7], to name just a few examples (see [8, 9] for reviews on pLMs and their applications).”
Current and future directions in network biology
Žitnik, M., Li, M. M., Wells, A., Glass, K., Gysi, D. M., Krishnan, A., Murali, T. M., Predrag Radivojac, A., Roy, S., Baudot, A., et al. (2024) · Bioinformatics Advances · doi:10.1093/bioadv/vbae099 · Cited by 128
“How to simultaneously leverage network- and sequence-based link prediction remains an open problem, with valuable initial work [29].”
BackgroundMethodSparks of function by de novo protein design
Chu, A. E., Lu, T., & Huang, P. (2024) · Nature Biotechnology · doi:10.1038/s41587-024-02133-2 · Cited by 131
A new age in protein design empowered by deep learning
Khakzad, H., Igashov, I., Schneuing, A., Goverde, C. A., Bronstein, M. M., & Correia, B. E. (2023) · Cell Systems · doi:10.1016/j.cels.2023.10.006 · Cited by 121
Contrastive learning on protein embeddings enlightens midnight zone
(2022) · doi:10.1093/nargab/lqac043 · Cited by 105
“22 23 Abbreviations used: 3D, three-dimensional; BFD, Big Fantastic Database (11); CATH, hierarchical classification of protein 3D structures in Class, Architecture, 24 Topology and Homologous superfamily (1,2); DL, Deep Learning; EAT, Embedding-based Annotation Transfer; EI, evolutionary information; embeddings fixed…”
BackgroundMethodInfluentialTMbed: transmembrane proteins predicted through language model embeddings
Bernhofer, M., & Rost, B. (2022) · BMC Bioinformatics · doi:10.1186/s12859-022-04873-x · Cited by 109
Taxonomic classification of DNA sequences beyond sequence similarity using deep neural networks
Mock, F., Kretschmer, F., Kriese, A., Böcker, S., & Marz, M. (2022) · Proceedings of the National Academy of Sciences · doi:10.1073/pnas.2122636119 · Cited by 111
Machine learning to navigate fitness landscapes for protein engineering
Freschlin, C. R., Fahlberg, S. A., & Romero, P. A. (2022) · Current Opinion in Biotechnology · doi:10.1016/j.copbio.2022.102713 · Cited by 111
“One method to reduce the data requirement for training is to apply more informative protein representations to supervised learning tasks [14].”
MethodBioSeq-Diabolo: Biological sequence similarity analysis using Diabolo
Li, H., & Liu, B. (2023) · PLoS Computational Biology · doi:10.1371/journal.pcbi.1011214 · Cited by 103
“As one of the fundamental steps in the biological structure and function studies, analysing biological sequence similarities is the foundation of many tasks in bioinformatics, such as protein remote homology detection [1], protein fold recognition [2], protein structure and function prediction [3–5], non-coding RNA an…”
AlQuraishi, M., & Sorger, P. K. (2021) · Nature Methods · doi:10.1038/s41592-021-01283-4 · Cited by 96
Li, L., Gupta, E., Spaeth, J., Shing, L., Jaimes, R., Engelhart, E., Lopez, R., Caceres, R. S., Bepler, T., & Walsh, M. E. (2023) · Nature Communications · doi:10.1038/s41467-023-39022-2 · Cited by 96
“All models were trained until convergence of the cross-entropy loss value (which is evaluated on the validation data after every epoch), or until the maximum number of epochs, 10, was reached.”
Self-citationLarge language models improve annotation of prokaryotic viral proteins
Flamholz, Z. N., Biller, S. J., & Kelly, L. (2024) · Nature Microbiology · doi:10.1038/s41564-023-01584-8 · Cited by 92
NetGO 3.0: Protein Language Model Improves Large-Scale Functional Annotations
Wang, S., You, R., Liu, Y., Xiong, Y., & Zhu, S. (2023) · Genomics Proteomics & Bioinformatics · doi:10.1016/j.gpb.2023.04.001 · Cited by 91
“, contact maps and structural similarity prediction tasks) to enrich 107 protein language models [22].”
BackgroundTopsy-Turvy: integrating a global view into sequence-based PPI prediction
Singh, R., Devkota, K., Sledzieski, S., Berger, B., & Cowen, L. (2022) · Bioinformatics · doi:10.1093/bioinformatics/btac258 · Cited by 83
“First, we generate a feature-rich representation of each protein using a protein language model (PLM) (Bepler and Berger, 2019, 2021); next, these features are combined using a convolutional neural network to predict interaction.”
MethodSelf-citationLiu, S., Rodriguez, J. S., Munteanu, V., Ronkowski, C., Sharma, N. K., Alser, M., Andreace, F., Blekhman, R., Błaszczyk, D., Chikhi, R., et al. (2025) · Nature Reviews Methods Primers · doi:10.1038/s43586-024-00376-6 · Cited by 99
“The tools ProtBERT 187 and ProSE 188 rely on sequence embeddings.”
Democratizing protein language models with parameter-efficient fine-tuning
Sledzieski, S., Kshirsagar, M., Baek, M., Dodhia, R., Lavista Ferres, J., & Berger, B. (2024) · Proceedings of the National Academy of Sciences · doi:10.1073/pnas.2405840121 · Cited by 86
Self-citationShrivastava, A. D., Swainston, N., Samanta, S., Roberts, I., Muelas, M. W., & Kell, D. B. (2021) · Biomolecules · doi:10.3390/biom11121793 · Cited by 96
Generative artificial intelligence for de novo protein design
Winnifrith, A., Outeiral, C., & Hie, B. (2024) · Current Opinion in Structural Biology · doi:10.1016/j.sbi.2024.102794 · Cited by 103
Lindorff-Larsen, K., & Kragelund, B. B. (2021) · Journal of Molecular Biology · doi:10.1016/j.jmb.2021.167196 · Cited by 85
Improving protein succinylation sites prediction using embeddings from protein language model
Pokharel, S., Pratyush, P., Heinzinger, M., Newman, R. H., & KC, D. B. (2022) · Scientific Reports · doi:10.1038/s41598-022-21366-2 · Cited by 90
“Since large protein sequence databases represent vast data goldmines, recently various pLMs have been developed for distilling information from them 26 .”
BackgroundKaminski, K., Ludwiczak, J., Pawlicki, K., Alva, V., & Dunin-Horkawicz, S. (2023) · Bioinformatics · doi:10.1093/bioinformatics/btad579 · Cited by 84
A survey of geometric graph neural networks: data structures, models and applications
Han, J., Cen, J., Wu, L., Li, Z., Kong, X., Jiao, R., Yu, Z., Xu, T., Wu, F., Wang, Z., et al. (2025) · Frontiers of Computer Science · doi:10.1007/s11704-025-41426-w · Cited by 104
“Pretraining ProSE [18] Prot.”
Biophysics-based protein language models for protein engineering
Gelman, S., Johnson, B., Freschlin, C. R., Sharma, A., D’Costa, S., Peters, J., Gitter, A., & Romero, P. A. (2025) · Nature Methods · doi:10.1038/s41592-025-02776-2 · Cited by 77
“These models learn meaningful, low-dimensional representations that capture the semantic organization of protein space and have broad utility in protein engineering 2 .”
Connecting chemistry and biology through molecular descriptors
Fernández‐Torras, A., Comajuncosa-Creus, A., Duran‐Frigola, M., & Aloy, P. (2021) · Current Opinion in Chemical Biology · doi:10.1016/j.cbpa.2021.09.001 · Cited by 74
Harnessing generative AI to decode enzyme catalysis and evolution for enhanced engineering
Xie, W. J., & Warshel, A. (2023) · National Science Review · doi:10.1093/nsr/nwad331 · Cited by 69
“…it is worth considering the development of models specifically tailored to the unique requirements of biological research, including proposing biologically meaningful tokenization, structure-informed sequence embedding, and designing a good measure for sequence representation quality [18 ,74 ,85 ].”
Theoretical and Data-Driven Approaches for Biomolecular Condensates
Saar, K. L., Qian, D., Good, L. L., Morgunov, A. S., Collepardo‐Guevara, R., Best, R. B., & Knowles, T. P. J. (2023) · Chemical Reviews · doi:10.1021/acs.chemrev.2c00586 · Cited by 66
AI Methods for Antimicrobial Peptides: Progress and Challenges
Brizuela, C. A., Liu, G., Stokes, J. M., & de la Fuente‐Núñez, C. (2025) · Microbial Biotechnology · doi:10.1111/1751-7915.70072 · Cited by 80
Kant, S., Deepika, & Roy, S. (2025) · Discover Pharmaceutical Sciences · doi:10.1007/s44395-025-00007-3 · Cited by 86
“Despite challenges like computational costs and interpretability, future directions include multimodal protein models and energy-based optimization for de novo protein design [66, 67].”
Ma, W., Zhang, S., Li, Z., Jiang, M., Wang, S., Lu, W., Bi, X., Jiang, H., Zhang, H., & Wei, Z. (2022) · Journal of Chemical Information and Modeling · doi:10.1021/acs.jcim.2c00885 · Cited by 64
CodonBERT large language model for mRNA vaccinesPreprint
Li, S., Moayedpour, S., Li, R., Bailey, M., Riahi, S., Kogler-Anele, L., Miladi, M., Miner, J. C., Pertuy, F., Zheng, D., et al. (2024) · Genome Research · doi:10.1101/gr.278870.123 · Cited by 66
A review of deep learning methods for ligand based drug virtual screening
Wu, H., Liu, J., Zhang, R., Lu, Y., Cui, G., Cui, Z., & Ding, Y. (2024) · Fundamental Research · doi:10.1016/j.fmre.2024.02.011 · Cited by 68
“Understanding the sequence-structure-function relationship of proteins is a central issue in protein biology and is essential for investigating disease mechanisms, protein design and drug discovery [76] .”
BackgroundMachine learning approaches in predicting allosteric sites
Nerín-Fonz, F., & Cournia, Z. (2024) · Current Opinion in Structural Biology · doi:10.1016/j.sbi.2024.102774 · Cited by 62
Evaluating the representational power of pre-trained DNA language models for regulatory genomics
Tang, Z., Somia, N., Yu, Y., & Koo, P. K. (2025) · Genome biology · doi:10.1186/s13059-025-03674-8 · Cited by 71
Ieremie, I., Ewing, R. M., & Niranjan, M. (2022) · Bioinformatics · doi:10.1093/bioinformatics/btac104 · Cited by 61
“…such as the primary structure of the protein (Chen et al. , 2019; Hashemifar et al. , 2018; Li et al. , 2018), the 3D protein structure (Bepler and Berger, 2021), gene expression profiles (Chin et al. , 2010) and Gene Ontology (GO) annotation (Bandyopadhyay and Mallick, 2017; Jain and…”
Deciphering microbial gene function using natural language processing
Miller, D., Stern, A., & Burstein, D. (2022) · Nature Communications · doi:10.1038/s41467-022-33397-4 · Cited by 61
“Such applications have been used to model various protein characteristics 16 – 19 , discover antimicrobial peptides 20 , and even predict antigens leading to viral escape 21 .”
MethodIn vitro continuous protein evolution empowered by machine learning and automation
Yu, T., Boob, A. G., Singh, N., Su, Y., & Zhao, H. (2023) · Cell Systems · doi:10.1016/j.cels.2023.04.006 · Cited by 55
The convergence of AI and synthetic biology: the looming deluge
Groff-Vindman, C., Trump, B. D., Cummings, C. L., Smith, M., Titus, A. J., Oye, K., Prado, V., Turmus, E. K., & Linkov, I. (2025) · npj Biomedical Innovations. · doi:10.1038/s44385-025-00021-1 · Cited by 73
“Initially, machine learning and biodesign tools were used for biodesign tasks such as predicting protein structure from amino acid sequence 1,2 .”
Pantolini, L., Studer, G., Pereira, J., Durairaj, J., Tauriello, G., & Schwede, T. (2024) · Bioinformatics · doi:10.1093/bioinformatics/btad786 · Cited by 55
“Two examples of embedding-based alignment (EBA) methods were introduced by Bepler and Berger (2021) and Hamamsy et al. (2023).”
Stražar, M., Park, J., Abelin, J. G., Taylor, H. B., Pedersen, T. K., Plichta, D. R., Brown, E., Eraslan, B., Hung, Y.-M., Ortiz, K., et al. (2023) · Immunity · doi:10.1016/j.immuni.2023.05.009 · Cited by 56
Large language models for science and medicine
Telenti, A., Auli, M., Hie, B., Maher, C., Saria, S., & Ioannidis, J. P. A. (2024) · European Journal of Clinical Investigation · doi:10.1111/eci.14183 · Cited by 62
Linguistically inspired roadmap for building biologically reliable protein language models
Vu, M. H., Akbar, R., Robert, P. A., Świątczak, B., Sandve, G. K., Greiff, V., & Haug, D. T. T. (2023) · Nature Machine Intelligence · doi:10.1038/s42256-023-00637-1 · Cited by 59
“While our points echo previous discussions on the differences between natural language and protein sequences 4,5,12,24,25 , our approach is unique in the focus on building protein LMs that learn true biological principles and in drawing insights from linguistics to reach this goal.”
BackgroundMethodInfluentialKo, Y. S., Parkinson, J., Liu, C., & Wang, W. (2024) · Briefings in Bioinformatics · doi:10.1093/bib/bbae359 · Cited by 54
“In addition, while Topsy-Turvy utilizes the ProSE [29] protein language model’s N × 6165 embedding for each length N protein sequence, TUnA uses ESM-2’s N × 640 embedding, requiring ∼ 10 times less memory.”
BackgroundMethodProT-VAE: Protein Transformer Variational AutoEncoder for functional protein design
Sevgen, E., Moller, J., Lange, A., Parker, J., Quigley, S. D., Mayer, J., Srivastava, P., Gayatri, S., Hosfield, D. J., Dilks, C. M., et al. (2025) · Proceedings of the National Academy of Sciences · doi:10.1073/pnas.2408737122 · Cited by 59
Wei, B., Zhang, Y., & Gong, X. (2022) · Scientific Reports · doi:10.1038/s41598-022-23014-1 · Cited by 53
“To utilize the raw drug molecular SMILES string and protein sequence as inputs to the DeepLPI model, we first encode them into numeric vector representations using the pre-trained embedding models Mol2Vec 17 and ProSE 18 , respectively.”
MethodInfluentialSETH predicts nuances of residue disorder from protein embeddings
Ilzhöfer, D., Heinzinger, M., & Rost, B. (2022) · Frontiers in Bioinformatics · doi:10.3389/fbinf.2022.1019597 · Cited by 52
Antibody design using deep learning: from sequence and structure design to affinity maturation
Joubbi, S., Micheli, A., Milazzo, P., Maccari, G., Ciano, G., Cardamone, D., & Medini, D. (2024) · Briefings in Bioinformatics · doi:10.1093/bib/bbae307 · Cited by 56
“This similarity has led to the development of NLP techniques tailored for encoding and using protein information [77, 78].”
MethodChen, J., Wang, J., Hu, Y., Li, X., Qian, Y., & Song, C.-L. (2025) · Frontiers in Bioengineering and Biotechnology · doi:10.3389/fbioe.2025.1506508 · Cited by 64
“These models are based on the distributional assumption that amino acids appearing in similar contexts tend to have similar meanings (Bepler and Berger, 2021).”
Structure-aware protein self-supervised learning
Chen, C., Zhou, J., Wang, F., Liu, X., & Dou, D. (2023) · Bioinformatics · doi:10.1093/bioinformatics/btad189 · Cited by 54
Are protein language models the new universal key?
Weißenow, K., & Rost, B. (2025) · Current Opinion in Structural Biology · doi:10.1016/j.sbi.2025.102997 · Cited by 54
Kurgan, L., Hu, G., Wang, K., Ghadermarzi, S., Zhao, B., Malhis, N., Erdős, G., Gsponer, J. A., Uversky, V. N., & Dosztányi, Z. (2023) · Nature Protocols · doi:10.1038/s41596-023-00876-x · Cited by 49
“Another relevant development are embeddings generated by the protein language models, which encode amino acids using numeric vectors that describe their surrounding sequence 173–175 .”
BackgroundEnhancing computational enzyme design by a maximum entropy strategy
Xie, W. J., Asadi, M., & Warshel, A. (2022) · Proceedings of the National Academy of Sciences · doi:10.1073/pnas.2122355119 · Cited by 50
Recent Advances in Machine Learning Variant Effect Prediction Tools for Protein Engineering
Horne, J., & Shukla, D. (2022) · Industrial & Engineering Chemistry Research · doi:10.1021/acs.iecr.1c04943 · Cited by 48
Evolutionary selection of proteins with two folds
(2023) · doi:10.1038/s41467-023-41237-2 · Cited by 41
“Nevertheless, the rapidgrowthofdiverse sequenced proteins 68 , recent advances in deep learning 69,70 , and increasingly accessible computational resources leave us optimistic that these challenges will be overcome.”
Nallapareddy, V., Bordin, N., Sillitoe, I., Heinzinger, M., Littmann, M., Waman, V. P., Sen, N., Rost, B., & Orengo, C. (2023) · Bioinformatics · doi:10.1093/bioinformatics/btad029 · Cited by 47
Manfredi, M., Savojardo, C., Martelli, P. L., & Casadio, R. (2022) · Bioinformatics · doi:10.1093/bioinformatics/btac678 · Cited by 48
“The learned representation captures important features of the proteins, including physicochemical, structural, functional , and evolutionary features (Bepler and Berger, 2021; Ofer et al ., 2021).”
BackgroundMethodInfluentialPLM-ARG: antibiotic resistance gene identification using a pretrained protein language model
Wu, J., Ouyang, J., Qin, H., Zhou, J., Roberts, R. A., Siam, R., Wang, L., Tong, W., Liu, Z., & Shi, T. (2023) · Bioinformatics · doi:10.1093/bioinformatics/btad690 · Cited by 48
“Different protein-language models were applied to develop predictive methods for extracting complex sequence–structure–function relationships (Bepler and Berger 2021, Ofer et al. 2021, Unsal et al. 2022).”
MethodPang, Y.-H., & Liu, B. (2024) · BMC Biology · doi:10.1186/s12915-023-01803-y · Cited by 38
“Its capability to capture semantic information of protein sequence, structure, and function [51] has demonstrated significant potential in a series of studies, including protein design [52–54] and protein function prediction [55].”
BackgroundA protein language model for exploring viral fitness landscapes
Ito, J., Strange, A. P., Liu, W., Joas, G., Lytras, S., Matsuno, K., Nao, N., Sawa, H., Mizuma, K., Kojima, I., et al. (2025) · Nature Communications · doi:10.1038/s41467-025-59422-w · Cited by 51
“We speculated that these challenges can be addressed using protein language models, an innovative technology that has become increasingly popular in the fi eld of natural language processing 31 .”
Synthetic molecular evolution of antimicrobial peptides
Chen, C. H., Bepler, T., Pepper, K., Fu, D., & Lu, T. K. (2022) · Current Opinion in Biotechnology · doi:10.1016/j.copbio.2022.102718 · Cited by 43
Self-citationPre‐Training of Equivariant Graph Matching Networks with Conformation Flexibility for Drug Binding
Wu, F., Jin, S., Jiang, Y., Jin, X., Tang, B., Niu, Z., Liu, X., Zhang, Q., Zeng, X., & Li, S. Z. (2022) · Advanced Science · doi:10.1002/advs.202203796 · Cited by 46
ADOPT: intrinsic protein disorder prediction through deep bidirectional transformers
Redl, I., Fisicaro, C., Dutton, O., Hoffmann, F., Henderson, L. D., Owens, B. M. J., Heberling, M. M., Paci, E., & Tamiola, K. (2023) · NAR Genomics and Bioinformatics · doi:10.1093/nargab/lqad041 · Cited by 42
“In a broader context, our work complements the most recent developments in NLP-based computational biology [see e.g. (66,67)].”
BackgroundMethodDeploying synthetic coevolution and machine learning to engineer protein-protein interactions
Yang, A., Jude, K. M., Lai, B., Minot, M., Kocyła, A., Glassman, C. R., Nishimiya, D., Kim, Y. S., Reddy, S. T., Khan, A. A., et al. (2023) · Science · doi:10.1126/science.adh1720 · Cited by 45
“One solution is the use of protein language models, which are self-supervisedmachinelearningmodelspretrainedonlargeprotein-sequencedatabases( 15 , 26 – 28 ).”
Improving antibody language models with native pairing
Burbach, S. M., & Briney, B. S. (2024) · Patterns · doi:10.1016/j.patter.2024.100967 · Cited by 40
“In essence, each person’s unique collection of affinity matured antibody genes constitutes a detailed molecular record of all previous pathogen encounters.”
Pakhrin, S. C., Pokharel, S., Aoki-Kinoshita, K. F., Beck, M. R., Dam, T. K., Caragea, D., & KC, D. B. (2023) · Glycobiology · doi:10.1093/glycob/cwad033 · Cited by 39
Engineering highly active nuclease enzymes with machine learning and high-throughput screening
Thomas, N., Belanger, D., Xu, C. A., Lee, H., Hirano, K., Iwai, K., Polic, V., Nyberg, K. D., Hoff, K. G., Frenz, L., et al. (2025) · Cell Systems · doi:10.1016/j.cels.2025.101236 · Cited by 41
Qian, H., Wang, Y., Zhou, X.-B., Gu, T., Wang, H., Lyu, H., Li, Z., Li, X., Zhou, H., Guo, C.-C., et al. (2025) · Nature Communications · doi:10.1038/s41467-025-58521-y · Cited by 35
“During the training process, the learning rate was set at 1×10 (cid:1) 5 and the Adam optimizer was utilized 44 .”
Using big sequencing data to identify chronic SARS-Coronavirus-2 infections
Harari, S., Miller, D., Fleishon, S., Burstein, D., & Stern, A. (2024) · Nature Communications · doi:10.1038/s41467-024-44803-4 · Cited by 37
“Recently, language models have gained widespread popularity due to their ability to tackle such large volumes of data and have impacted research in diverse areas of biological sciences, including virology 28 – 30 .”
Domain loss enabled evolution of novel functions in the snake three-finger toxin gene superfamily
Koludarov, I., Senoner, T., Jackson, T., Dashevsky, D., Heinzinger, M., Aird, S. D., & Rost, B. (2023) · Nature Communications · doi:10.1038/s41467-023-40550-0 · Cited by 38
“As this objective does not require labeled data, it allows mining the wealth of unlabeled data 17 – 19, 60 , i.e. , data In thiswork, we generated embeddings foreachproteinsequence using the pLM ProtT5-XL-UniRef50 18 (for simplicity referred to as ProtT5 ), which has been built in analogy to the…”
Yeung, W., Zhou, Z., Li, S., & Kannan, N. (2023) · Briefings in Bioinformatics · doi:10.1093/bib/bbac599 · Cited by 35
“Despite being trained on sequences alone, these models are capable of learning protein representations which encode structural, functional and evolutionary features [3].”
Wang, D., Pourmirzaei, M., Abbas, U. L., Zeng, S., Manshour, N., Esmaili, F., Poudel, B., Jiang, Y., Shao, Q., Chen, J., et al. (2025) · Advanced Science · doi:10.1002/advs.202404212 · Cited by 46
Exploring Scoring Function Space: Developing Computational Models for Drug Discovery
Bitencourt‐Ferreira, G., Villarreal, M. A., Quiroga, R., Biziukova, N. Y., Poroikov, V., Tarasova, O. A., & de Azevedo, W. F. (2023) · Current Medicinal Chemistry · doi:10.2174/0929867330666230321103731 · Cited by 34
CodonBERT: Large Language Models for mRNA design and optimizationPreprint
Li, S., Moayedpour, S., Li, R., Bailey, M., Riahi, S., Kogler-Anele, L., Miladi, M., Miner, J. C., Zheng, D., Wang, J., et al. (2023) · bioRxiv (Cold Spring Harbor Laboratory) · doi:10.1101/2023.09.09.556981 · Cited by 36
“Recently, this concept has been scaled to biological sequences (protein, DNA, and RNA) [32, 33, 34, 35, 36].”
BackgroundLeveraging protein language models for accurate multiple sequence alignments
(2023) · doi:10.1101/gr.277675.123 · Cited by 30
“Here, weconsidera new approach to determining the similarities between amino acids across proteins, based on protein language models (Rao et al. 2019; Bepler and Berger 2021; Rives et al. 2021; Chowdhury et al. 2022; Elnaggar et al. 2022).”
Hu, F., Hu, Y., Zhang, W., Huang, H., Pan, Y. F., & Yin, P. (2023) · Advanced Science · doi:10.1002/advs.202301223 · Cited by 33
Superior protein thermophilicity prediction with protein language model embeddings
Haselbeck, F., John, M., Zhang, Y., Pirnay, J., Werner, J., Costa, R. D., & Grimm, D. G. (2023) · NAR Genomics and Bioinformatics · doi:10.1093/nargab/lqad087 · Cited by 35
Gujral, O. S., Bafna, M., Alm, E. J., & Berger, B. (2025) · Proceedings of the National Academy of Sciences · doi:10.1073/pnas.2506316122 · Cited by 49
Self-citation
Metadata
- Authors
- Bepler, T., & Berger, B.
- Affiliations
- New York Structural Biology Center (US)
- MIT Computer Science and Artificial Intelligence Laboratory (US)
- Corresponding author
- Bepler, Tristan
- Berger, Bonnie
- Venue
- Cell Systems
- Publisher
- Elsevier BV
- Published
- 2021-06
- Volume
- 12
- Issue
- 6
- Pages
- 654-669.e3
- ISSN
- 2405-4712 (ISSN-L)
- PMC
- PMC8238390
- OpenAlex
- W3166142427
- Cited by
- 483
- References
- 133
- Citations by year
- 2026: 87 · 2025: 120 · 2024: 127 · 2023: 82 · 2022: 55 · 2021: 12
- FWCI
- 21.74
- Type
- journal article
- Language
- en
- Open access
- hybrid
- doi.org
- Licence
- Locations
- Cell Systems (publishedVersion, cc-by-nc-nd)
- PubMed (publishedVersion)
- PubMed Central (acceptedVersion)
- DSpace@MIT (Massachusetts Institute of Technology) (submittedVersion, cc-by-nc-nd)
- Funders
- NIH (R35 GM141861)
- Simons Foundation (SF349247)
- Topics
- Life Sciences › Biochemistry, Genetics and Molecular Biology › Molecular Biology › Machine Learning in Bioinformatics
- Life Sciences › Biochemistry, Genetics and Molecular Biology › Molecular Biology › Protein Structure and Dynamics
- Life Sciences › Biochemistry, Genetics and Molecular Biology › Molecular Biology › Fractal and DNA sequence analysis
- Keywords
- protein language models, protein sequence representation, protein structure prediction, protein function prediction, knowledge distillation, transfer learning, protein design, therapeutic design
- MeSH terms
- Machine Learning, Amino Acid Sequence, Language, Proteins, Databases, Protein
About these checks
Registry, retraction, Cites and statistics results are computed, not judged, with one exception: a printed reference that no index lists is read once by a model before it counts as a citation error. Claims are extracted by a model and kept only when their quote appears verbatim in the text; each cited source is then checked, with a model’s help, for whether it says what the claim relies on. Those verdicts can be wrong — read the claim’s quote and the cited source before relying on one.
A language model drew the claims from the paper on 2026-10-10, under version 3 of the claim definition (model deepseek-v4.1-flash, prompt extract-v3).
A second model labelled each claim on 2026-10-10 (model jev-1.13-20260917, definition revision gold-guideline-v0.2).
Statistics and arithmetic are recomputed, not judged.
Confidence bands are shown only for fields calibrated on our annotated test set. None is calibrated yet, so no band is shown.
What counts as a claim ›How we measure confidence ›How we check this ›
Are you an author, or do you see a problem with this page? Request a review
