Computational and artificial intelligence-based methods for antibody development

Probably not a substantial claim: may not pass S3 (load-bearing).

Each antibody chain is encoded by multiple gene segments (V, D, and J segments), which are spliced together using a V(D)J recombination process.

“Each of the chains is encoded by multiple gene segments (V, D, and J segments), which are spliced together using a V(D)J recombination process [7].”

• cites [7] sourceNot checked yet

  Not checked in the free run.

Probably not a substantial claim: may not pass S4 (based).

De novo design of antibodies remains a challenging problem.

“Deep learning approaches are improving the performance of structure prediction of antibodies, including CDRs, while de novo design remains a challenging problem.”

Classified as probably not a substantial claim.

Large snapshots of the antibody repertoire can now be obtained using next-generation sequencing.

“Large snapshots of this repertoire can now be obtained using next-generation sequencing (NGS) approaches.”

Probably not a substantial claim: may not pass S3 (load-bearing).

Through V(D)J recombination, a diverse range of antibody sequences can be produced.

“Through this process, a diverse range of antibody sequences can be produced.”

Classified as probably not a substantial claim.

Information about antibody sequences, structures, and their associated properties has been informative to the design of novel antibodies.

“Thus, information about antibody sequences, structures, and their associated properties has been informative to the design of novel antibodies.”

Probably not a substantial claim: may not pass S3 (load-bearing).

Efforts have been made to create standardized, publicly available repositories for antibody sequencing data.

“Efforts have been made to create standardized, publicly available repositories for these sequencing data [8,9].”

• cites [8] sourceNot checked yet

  Not checked in the free run.

• cites [9] sourceNot checked yet

  Not checked in the free run.

Probably not a substantial claim: may not pass S4 (based).

Antibody abilities and properties are determined by antibody sequences and structures.

“These abilities and properties are determined by antibody sequences and structures.”

Probably not a substantial claim: may not pass S3 (load-bearing).

Prediction of drug-target interactions in small-molecule drug discovery is largely driven by significant improvement in the performance of computational methods.

“This is analogous to small-molecule drug discovery, where computational methods have already made significant inroads [5,6]; prediction of drug-target interactions in particular is largely driven by the significant improvement in the performance of computational methods.”

• cites [5] sourceNot checked yet

  Not checked in the free run.

• cites [6] sourceNot checked yet

  Not checked in the free run.

Probably not a substantial claim: may not pass S4 (based).

A large suite of predictors of different biophysical and other properties of antibodies have been developed.

“A large suite of predictors of different biophysical and other properties of antibodies have been developed.”

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

These databases have provided researchers with easy access to a vast number of sequences and created opportunities for large-scale data mining.

“These databases have provided researchers with easy access to a vast number of sequences and created opportunities for large-scale data mining.”

May combine two claims.

The observed antibody space (OAS) collates variable fragment (Fv) sequences and contains nearly 2 billion sequences spanning 68 different studies.

“The observed antibody space (OAS) collates variable fragment (Fv) sequences and contains nearly 2 billion sequences spanning 68 different studies [8].”

• cites [8] sourceNot checked yet

  Not checked in the free run.

Classified as probably not a substantial claim.

Databases hold large repertoires of antibody sequences.

“Databases hold large repertoires of antibody sequences, but only limited structural data; data on biophysical properties is also available.”

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

In the human immune system, antibodies are produced by B cells and the repertoire is estimated at approximately 10^13 unique sequences.

“In the human immune system, antibodies are produced by B cells and its repertoire is estimated at approximately 10 13 unique sequences [7].”

• cites [7] sourceNot checked yet

  Not checked in the free run.

Probably not a substantial claim: may not pass S3 (load-bearing).

Data on biophysical properties of antibodies is available.

“Databases hold large repertoires of antibody sequences, but only limited structural data; data on biophysical properties is also available.”

Probably not a substantial claim: may not pass S3 (load-bearing).

Several deep learning models have been trained on the OAS database as a means of discerning and generating humanized antibody sequences.

“Several DL models have been trained on the OAS database as a means of discerning and generating humanized antibody sequences (details are described in the 'Computational design of antibodies' section).”

Probably not a substantial claim: may not pass S3 (load-bearing).

Therapeutic antibodies account for four of the ten top therapeutics by sales in 2021.

“Therapeutic antibodies are highly successful biotherapeutics and account for four of the ten top therapeutics by sales in 2021 [1].”

• cites [1] sourceNot checked yet

  Not checked in the free run.

May combine two claims.

Traditional experimental antibody discovery approaches are time-consuming and laborious and have several limitations, including difficulty specifying the antibody-binding epitope and issues obtaining antibodies manufacturable at scale.

“However, these approaches are timeconsuming and laborious and have several limitations, including difficulties in specifying the antibody-binding side (epitope), as well as issues with obtaining antibodies that can be manufactured at scale (Box 1).”

May combine two claims.

TAP provides guideline cut-offs for each metric and those cut-off values are collectively used for filtering out antibody candidates with poor developability.

“TAP also provides guideline cut-offs for each metric and those cut-off values are collectively used for filtering out antibody candidates with poor developability [23].”

• cites [23] sourceNot checked yet

  Not checked in the free run.

Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based). May combine two claims.

While the aforementioned tools predict overall developability, there are other tools that estimate one specific property for therapeutic antibody candidates.

“While the aforementioned tools predict the overall developability, there are other tools that estimate one specific property for therapeutic antibody candidates.”

Probably not a substantial claim: may not pass S3 (load-bearing).

The developer of AbDiver identified that it could find suitable profiles for 742 therapeutic antibodies.

“Moreover, the developer identified that it could find suitable profiles for 742 therapeutic antibodies.”

• cites [10] sourceNot checked yet

  Not checked in the free run.

Probably not a substantial claim: may not pass S4 (based).

Recently, computational tools or methods based on either classic statistics or ML have been developed and can be used to rapidly predict the developability of antibody candidates.

“Recently, computational tools or methods based on either classic statistics or ML have been developed and can be used to rapidly predict the developability of antibody candidates.”

Probably not a substantial claim: may not pass S3 (load-bearing).

Antibody structures in the Protein Data Bank (PDB) have been extracted and compiled into various datasets.

“Antibody structures in the Protein Data Bank (PDB) have been extracted and compiled into various datasets [16].”

• cites [16] sourceNot checked yet

  Not checked in the free run.

Probably not a substantial claim: may not pass S3 (load-bearing).

As of July 2022, Thera-SAbDab is tracking 748 unique therapeutics, including monoclonal antibody and bispecific therapeutics.

“As of July 2022, this database is tracking 748 unique therapeutics, including monoclonal antibody and bispecific therapeutics.”

• cites [15] sourceNot checked yet

  Not checked in the free run.

Probably not a substantial claim: may not pass S4 (based).

Most initial hits coming out of screening campaigns tend to have poor developability, leading to challenges later in their development.

“Most initial hits coming out of screening campaigns tend to have poor developability, leading to challenges later in their development.”

Probably not a substantial claim: may not pass S2 (checkable), S4 (based).

Developability is crucial for the development of novel antibody therapeutics.

“Apart from binding specificity and affinity, developability is crucial for the development of novel antibody therapeutics.”

• AssumptionRests on an assumption: Developability properties influence the likelihood that an antibody candidate can be advanced to clinical use, given proper efficacy.

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

AbDiver employs publicly curated B cell receptor NGS datasets and compares designed sequences to natural repertoires.

“For example, AbDiver is a tool that employs the publicly curated B cell receptor NGS datasets and compares designed sequences to natural repertoires [10].”

• cites [10] sourceNot checked yet

  Not checked in the free run.

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

Researchers can use information obtained from antibody structures to increase binding affinity, or to develop methods to predict the epitope and paratope.

“Researchers can use the information obtained from antibody structures to increase binding affinity, or to develop methods to predict the epitope and paratope [12][13][14][15].”

• cites [12][13][14][15]Not checked yet

  Not checked in the free run.

May combine two claims.

Antibody binders to many targets can be obtained from existing well-established experimental approaches, such as immunization or directed evolution, but substantial limitations remain.

“Antibody binders to many targets can be obtained from existing well-established experimental approaches, such as immunization or directed evolution, but there still remain substantial limitations.”

Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

Several main properties comprise developability: intrinsic immunogenicity, aggregation/insolubility, viscosity, and half-life span.

“Several main properties comprise developability: intrinsic immunogenicity, aggregation/insolubility, viscosity, and half-life span.”

Probably not a substantial claim: may not pass S3 (load-bearing).

The six CDR loops, three on each VH and VL, are involved in antigen binding.

“The six CDR loops, three on each of the variable domain of the heavy chain (VH) and the light chain (VL), are involved in antigen binding.”

Probably not a substantial claim: may not pass S3 (load-bearing).

Thera-SAbDab is a therapeutic structural antibody database including antibody- and nanobody-related biotherapeutics.

“Additionally, the Therapeutic Structural Antibody Database (Thera-SAbDab) is a therapeutic structural antibody database including antibody-and nanobodyrelated biotherapeutics [15].”

• cites [15] sourceNot checked yet

  Not checked in the free run.

Classified as probably not a substantial claim. May combine two claims.

The 3D structure of the antibody determines how it interacts with an antigen and governs its binding properties.

“Therefore, the 3D structure of the antibody determines how it interacts with an antigen and governs its binding properties.”

Probably not a substantial claim: may not pass S3 (load-bearing).

RAPID includes more than 300 million clones extracted from human heavy chain repertoire sequencing.

“Another tool, the repertoire sequencing dataset analysis platform with an integrated antibody database (RAPID), includes more than 300 million clones extracted from human HC repertoire sequencing [11].”

• cites [11]Not checked yet

  Not checked in the free run.

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

According to specific criteria, antibody structures from the PDB are listed in SAbDab, a database containing 12,367 Fv region structures as of June 2022, and abYsis.

“Similarly, according to specific criteria, antibody structures from the PDB are listed in the Structural Antibody Database (SAbDab) [13], a database containing 12 367 Fv region structures as of June 2022, and abYsis [14].”

• cites [13] sourceNot checked yet

  Not checked in the free run.

• cites [14] sourceNot checked yet

  Not checked in the free run.

Probably not a substantial claim: may not pass S3 (load-bearing).

Chen et al. built an ML pipeline to predict antibody developability using a sequence dataset of 2400 antibodies from the SAbDab database.

“For example, Chen et al. [24] built an ML pipeline to predict antibody developability using a sequence dataset of 2400 antibodies from the SAbDab database.”

• cites [24]Not checked yet

  Not checked in the free run.

Probably not a substantial claim: may not pass S3 (load-bearing).

Additional dedicated datasets summarize curated information about properties and structure of antibodies, including SKEMPI and AB-Bind.

“Additional dedicated datasets summarize curated information about properties and structure of antibodies.These databases include the structural database of kinetics and energetics of mutant protein interactions (SKEMPI) [17], and the antibody binding mutational database (AB-Bind) [18].”

• cites [17] sourceNot checked yet

  Not checked in the free run.

• cites [18] sourceNot checked yet

  Not checked in the free run.

Probably not a substantial claim: may not pass S3 (load-bearing).

Fv regions from antibodies have been updated in the Antibody Structure Database (AbDb) using information from the Summary of Antibody Crystal Structures (SACS) dataset.

“Fv regions from the antibody have been updated in the Antibody Structure Database (AbDb) using information from the Summary of Antibody Crystal Structures (SACS) dataset [12].”

• cites [12] sourceNot checked yet

  Not checked in the free run.

Probably not a substantial claim: may not pass S3 (load-bearing).

Antibodypedia provides validation information on commercially available antibodies.

“Moreover, Antibodypedia provides validation information on commercially available antibodies, another useful resource for dataset curation [21].”

• cites [21] sourceNot checked yet

  Not checked in the free run.

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

To predict the solubility and aggregation propensity of protein, several prediction tools have been developed, such as Camsol and SOLart, with SOLart currently holding state-of-the-art performance.

“To predict the solubility and aggregation propensity of protein, several prediction tools have been developed, such as the Camsol [25], and SOLart [26], with SOLart currently holding state-of-the-art performance.”

• cites [25] sourceNot checked yet

  Not checked in the free run.

• cites [26] sourceNot checked yet

  Not checked in the free run.

May combine two claims.

Aggregation of antibody therapeutics can lead to precipitation and shortened storage period of drugs before administration, while aggregation in vivo can increase the immunogenicity of the drug.

“Aggregation of antibody therapeutics can lead to precipitation and shortened storage period of drugs before administration, while aggregation in vivo can increase the immunogenicity of the drug [22].”

• cites [22] sourceNot checked yet

  Not checked in the free run.

Probably not a substantial claim: may not pass S3 (load-bearing).

While antibody datasets remain relatively small on their own, these repositories can be used in conjunction with other datasets through techniques such as transfer learning.

“While these datasets remain relatively small on their own, these repositories can be used in conjunction with other datasets through techniques such as transfer learning (a process in which information learned from one task is applied to a related task, i.e., refining a model trained on data for one task with additional data from another).”

Probably not a substantial claim: may not pass S4 (based).

Most screening campaigns only yield binders to one given epitope, which may not be the one needed to achieve the desired biological or therapeutic effect.

“Most screening campaigns only yield binders to one given epitope, which may not be the one needed to achieve the desired biological or therapeutic effect.”

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

There are more focused repositories of antibodies, such as CoV-AbDab for anticoronavirus antibodies, and sdAb-DB for single-domain antibodies and related classes of proteins.

“Furthermore, there are more focused repositories of antibodies, such as CoV-AbDab [19] for anticoronavirus antibodies, and sdAb-DB [20] for single-domain antibodies and related classes of proteins.”

• cites [19] sourceNot checked yet

  Not checked in the free run.

• cites [20] sourceNot checked yet

  Not checked in the free run.

Probably not a substantial claim: may not pass S4 (based).

It remains quite difficult to target specific epitopes on the antigen.

“First of all, it remains quite difficult to target specific epitopes on the antigen.”

Probably not a substantial claim: may not pass S4 (based). May combine two claims.

Once a candidate antibody binder has been obtained, a structure of the antibody/antigen complex is highly useful for further engineering, but requires X-ray crystallography or NMR, which is typically time-consuming and low throughput.

“Moreover, once a candidate antibody binder has been obtained, a structure of the antibody/antigen complex is highly useful for further engineering, but requires either X-ray crystallography or NMR, which is typically timeconsuming and low throughput.”

Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

Hydrophobicity can relate to aggregation propensity, solubility, viscosity, self-interaction, and protein stability, making it useful to predict potential downstream risks.

“Hydrophobicity can relate to aggregation propensity, solubility, viscosity, self-interaction, and protein stability, making it useful to predict potential downstream risks.”

Probably not a substantial claim: may not pass S4 (based).

Only a relatively small subset of antibodies has the biophysical and other properties needed to become therapeutics (developability).

“Second, only a relatively small subset of antibodies has the biophysical and other properties needed to become therapeutics (e.g., solubility in high concentrations, manufacturability at large scale); these properties are referred to as 'developability' of an antibody.”

Classified as probably not a substantial claim.

Immunogenic response occurs when the immune system recognizes the B and T cell epitopes of antibody therapeutics, thereby eliciting the production of anti-drug antibodies.

“Immunogenic response occurs when the immune system recognizes the B and T cell epitopes of antibody therapeutics, thereby eliciting the production of anti-drug antibodies.”

Probably not a substantial claim: may not pass S3 (load-bearing).

IEDB-AR is a comprehensive website that provides several computational tools focused on the prediction and analysis of B and T cell epitopes.

“Immune epitope database analysis resource (IEDB-AR) [31] is a comprehensive website that provides several computational tools focused on the prediction and analysis of B and T cell epitopes.”

• cites [31] sourceNot checked yet

  Not checked in the free run.

May combine two claims.

They provide an open-source platform, named BioPhi, incorporating the Transformer, and an evaluation method to aid antibody humanization.

“They provide an open-source platform, named BioPhi [33], incorporating the Transformer, and an evaluation method to aid antibody humanization.”

• cites [33] sourceNot checked yet

  Not checked in the free run.

May combine two claims.

Although these results present a significant advancement of antibody structure modeling, more work is needed and the reliable modeling of an antibody/antigen complex structure remains an unsolved problem.

“Although these results present a significant advancement of antibody structure modeling (largely driven by the adoption of DL techniques), more work is needed and the reliable modeling of an antibody/antigen complex structure remains an unsolved problem.”

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

Antibody structures have conventionally been predicted using either physics-based modeling, such as MD simulations, or homology-based modeling, or a combined method, such as MODELLER.

“Antibody structures have conventionally been predicted using either physics-based modeling, such as MD simulations [38], or homologybased modeling [39], or a combined method, such as MODELLER [40].”

• cites [38] sourceNot checked yet

  Not checked in the free run.

• cites [39] sourceNot checked yet

  Not checked in the free run.

• cites [40] sourceNot checked yet

  Not checked in the free run.

May combine two claims.

Grinshpun et al. collected 64 mAbs that have been approved or are under Phase 2 and 3 clinical trials, and their published clearance data, and trained a random forest classifier to distinguish antibodies with different clearance levels.

“To predict clearance rate through comprehensive evaluation of these factors, Grinshpun et al. [35] collected 64 mAbs that have been approved or are under Phase 2 and 3 clinical trials, and their published clearance data, and trained a random forest classifier to distinguish antibodies with different clearance levels.”

• cites [35] sourceNot checked yet

  Not checked in the free run.

May combine two claims.

DeepAb achieved a prediction accuracy for H3 loop at RMSD of 2.33 Å while ABlooper and ABodyBuilder achieved RMSD of 2.49 Å and 3.25 Å, respectively.

“In the end, DeepAb achieved a prediction accuracy for H3 loop at rootmean-square deviation (RMSD) of 2.33 Å while ABlooper and ABodyBuilder achieved RMSD of 2.49 Å and 3.25 Å, respectively [41].”

• cites [41] sourceNot checked yet

  Not checked in the free run.

May combine two claims.

For five of the six CDRs (H1, H2, L1, L2, and L3), structural diversity is limited and the loops typically follow canonical conformations; thus, these loops have significantly more structural constraints, allowing for easier prediction.

“For five of the six CDRs (H1, H2, L1, L2, and L3), structural diversity is limited and the loops typically follow canonical conformations; thus, these loops have significantly more structural constraints, allowing for easier prediction.”

Probably not a substantial claim: may not pass S3 (load-bearing).

They randomly masked or mutated amino acids in the natural human Fv sequences from OAS database and required Sapiens to predict the original amino acids based on the remaining sequences.

“They randomly masked or mutated amino acids in the natural human Fv sequences from OAS database and required the Sapiens to predict the original amino acids based on the remaining sequences.”

May combine two claims.

Antibody therapeutics are often derived from mice or other non-human organisms, which are therefore potential epitopes recognized by B and T cells.

“Antibody therapeutics are often derived from mice or other non-human organisms, which are therefore potential epitopes recognized by B and T cells (B and T cell epitopes).”

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

Pharmacokinetic clearance determines the half-life of an antibody and therefore determines how long an antibody drug can remain in the body after administration.

“Pharmacokinetic (PK) clearance determines the half-life of an antibody and therefore determines how long an antibody drug can remain in the body after administration.”

May combine two claims.

Methods to predict immunogenic sequences and humanize them have been developed.

“As such, methods to predict immunogenic sequences and humanize them have been developed.”

Classified as probably not a substantial claim.

When performing large screens, independently determining all antibody loop structures and their interface with the antigen is impractical.

“When performing large screens, independently determining all antibody loop structures and their interface with the antigen is impractical.”

May combine two claims.

CDR-H3 loops have high conformational diversity, even in cases where sequence similarity is high, and so CDR-H3 modeling is a significantly less constrained, more difficult problem.

“Conversely, CDR-H3 loops have high conformational diversity, even in cases where sequence similarity is high [37], and so CDR-H3 modeling is a significantly less constrained, more difficult problem.”

• cites [37] sourceNot checked yet

  Not checked in the free run.

Classified as probably not a substantial claim. May combine two claims.

Antibody structure information is critical to developing an understanding of the characteristics of designed antibodies, such as specificity and affinity, and so experimental prediction of the structures of designs is crucial for model development.

“Antibody structure information is critical to developing an understanding of the characteristic of designed antibodies, such as specificity and affinity, and so experimental prediction of the structures of designs is crucial for model development.”

May combine two claims.

The Deane group used nearly 2 billion antibody sequences available in OAS across organisms, labeling all human sequences as positive and others as negative, and trained a random forest classifier model to produce a 'humanness score'.

“Moreover, the Deane group used nearly 2 billion antibody sequences, available in OAS, across organisms where all human sequences were labeled as positive class, while others were labeled as negative class, and trained a random forest classifier model to produce a 'humanness score'.”

• cites [32] sourceNot checked yet

  Not checked in the free run.

May combine two claims.

The output score not only distinguishes human V genes from non-human sequences in the variable region but also indicates the level of immunogenicity risks.

“The output score not only distinguishes human V genes from non-human sequences in the variable region but also indicates the level of immunogenicity risks [32].”

• cites [32] sourceNot checked yet

  Not checked in the free run.

Probably not a substantial claim: may not pass S3 (load-bearing).

Labute et al. reported structure-based charge calculations to predict PK clearance rate.

“Also, Labute et al. [36] reported structure-based charge calculations to predict PK clearance rate.”

• cites [36] sourceNot checked yet

  Not checked in the free run.

Probably not a substantial claim: may not pass S3 (load-bearing).

AlphaFold2 has proven to be revolutionary in predicting protein structure from sequence.

“AlphaFold2 [46] has proven to be revolutionary in predicting protein structure from sequence.”

• cites [46] sourceNot checked yet

  Not checked in the free run.

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

The random forest classifier consists of 200 decision trees and each decision tree individually separates the input sequences into classes (human or other species) using their features.

“The random forest classifier consists of 200 decision trees and each decision tree individually separates the input sequences into classes (human or other species) using their features.”

• cites [32] sourceNot checked yet

  Not checked in the free run.

Probably not a substantial claim: may not pass S3 (load-bearing).

ABodyBuilder is an automated antibody homology modeling pipeline that follows four steps: template selection, orientation prediction, CDR loop modeling, and side chain prediction.

“For example, ABodyBuilder [41], an automated antibody homology modeling pipeline, follows four steps: template selection, orientation prediction, CDR loop modeling, and side chain prediction.”

• cites [41] sourceNot checked yet

  Not checked in the free run.

May combine two claims.

Such methods can be used to guide the engineering of antibody candidates and avoid 'developability traps'.

“Such methods can be used to guide the engineering of antibody candidates and avoid 'developability traps' (e.g., antibodies with very poor aggregation or viscosity properties).”

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

The Transformer used attention mechanisms to put different weights in different amino acids in the input sequence, revealing the different dependencies between residues and whole sequences.

“The Transformer used attention mechanisms to put different weights in different amino acids in the input sequence, revealing the different dependencies between residues and whole sequences.”

May combine two claims.

The Gray group recently established a protein-language-based model called IgFold, which can predict structures faster (one structure under 1 min).

“The Gray group recently established a protein-language-based model called IgFold, which can predict structures faster (one structure under 1 min).”

• cites [45]Not checked yet

  Not checked in the free run.

May combine two claims.

Immunogenicity negatively affects therapeutic efficacy and impacts safety through the induction of adverse drug reactions.

“Immunogenicity negatively affects therapeutic efficacy and impacts safety through the induction of adverse drug reactions [30].”

• cites [30] sourceNot checked yet

  Not checked in the free run.

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

AlphaFold2 currently struggles to predict the structures of orphan proteins, and MSA generation proves to be a major computational bottleneck.

“Currently it struggles to predict the structures of orphan proteins, and MSA generation proves to be a major computational bottleneck.”

Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

The performance of AlphaFold2-Multimer is still not optimal.

“However, the performance of AlphaFold2-Multimer is still not optimal.”

Probably not a substantial claim: may not pass S3 (load-bearing).

Since then, there have been great advances in NLP due to developments of Transformer-based models such as BERT.

“Since then, there have been great advances in NLP due to developments of Transformer-based models such as BERT [57].”

• cites [57]Not checked yet

  Not checked in the free run.

Probably not a substantial claim: may not pass S3 (load-bearing).

Shin et al. applied causal dilated convolutions to learn long-range relationships in sequences in an autoregressive model.

“In an early method, Shin et al. [56] applied causal dilated convolutions to learn long-range relationships in sequences in an autoregressive model (i.e., a model that predicts each amino acid one at a time, using the previous predictions as the next model input).”

• cites [56] sourceNot checked yet

  Not checked in the free run.

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

AlphaFold2 functions by taking in an input multiple sequence alignment (MSA) which provides important structural information via evolutionary history, as well as template PDB structures which provide further structural constraints to the model.

“It functions by taking in an input multiple sequence alignment (MSA) which provides important structural information via evolutionary history, as well as template PDB structures which provide further structural constraints to the model.”

Probably not a substantial claim: may not pass S3 (load-bearing).

In the second category, there are DL models that attempt to learn overall features of antibodies from their sequence alone.

“In the second category, there are DL models that attempt to learn overall features of antibodies from their sequence alone (Table 3).”

Probably not a substantial claim: may not pass S3 (load-bearing).

The BERT architecture was directly trained on the Uniref100 and BFD-100 datasets to create the ProtBERT model for protein sequence prediction.

“Such models were then applied to protein sequence: the BERT architecture was directly trained on the Uniref100 and BFD-100 datasets to create the ProtBERT model [58] for protein sequence prediction.”

• cites [58] sourceNot checked yet

  Not checked in the free run.

Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

Structure-based models do not allow for conditioning on specific epitopes.

“Furthermore, these structure-based models do not allow for conditioning on specific epitopes, which is also not desirable.”

Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

Invariant point attention (IPA), also introduced in AlphaFold2, allows 3D atom coordinates in a global frame to be generated.

“Invariant point attention (IPA), also introduced in AlphaFold2, allows 3D atom coordinates in a global frame to be generated.”

Probably not a substantial claim: may not pass S3 (load-bearing).

The drawback to backbone only generation is that they rely on external tools from Rosetta to design sequences that would fold into the specified structure.

“The drawback to backbone only generation is that they rely on external tools from Rosetta to design sequences that would fold into the specified structure.”

Probably not a substantial claim: may not pass S2 (checkable).

Since binding to antigen could lead to conformational changes, it is critical to extend current models to include antigen information when predicting Fv or CDR-H3 loop structures.

“Finally, since binding to antigen could lead to conformational changes, it is critical to extend current models to include antigen information when predicting Fv or CDR-H3 loop structures.”

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

In ESM-1b, 250 million protein sequences were used to train a language model which learned embeddings that encoded important biological properties, which could be easily recovered using trained linear layers.

“Similarly, in ESM-1b [59], 250 million protein sequences were used to train a language model which learned embeddings that encoded important biological properties, which could be easily recovered using trained linear layers.”

• cites [59] sourceNot checked yet

  Not checked in the free run.

May combine two claims.

DLAB localizes the interface of an antibody and antigen and uses a 3D convolution-based neural network to generate binding scores.

“For example, DLAB [52] localizes the interface of an antibody and antigen and uses a 3D convolution-based neural network to generate binding scores.”

• cites [52] sourceNot checked yet

  Not checked in the free run.

May combine two claims.

Limitations of these models include the necessary discretization; since the correct fineness is not known a priori, there is the potential for significant information loss.

“Limitations of these models include the necessary discretization (segmenting the protein structure into 3D grid of user-defined fineness, where each grid element may contain atom density information about different types, such as carbon and oxygen); since the correct fineness is not known a priori, there is the potential for significant information loss.”

May combine two claims.

Both ProtBERT and ESM-1b were highly successful at capturing many or most properties of protein sequences.

“Both ProtBERT and ESM-1b were highly successful at capturing many or most properties of protein sequences.”

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

AlphaFold2 extensively uses the attention mechanism that dominates NLP, as well as geometrically inspired attention variants such as triangle self-attention.

“AlphaFold2 extensively uses the attention mechanism that dominates NLP, as well as geometrically inspired attention variants such as triangle self-attention.”

May combine two claims.

IG-VAE model builds on previous work from the Huang group, generates complete backbones including all 3D coordinates by implementing a variational autoencoder trained to accurately reconstruct 3D coordinates, torsional angles, and distance maps from a learned latent space.

“For example, the IG-VAE model [53], which builds on previous work from the Huang group [54], generates complete backbones including all 3D coordinates by implementing a variational autoencoder trained to accurately reconstruct 3D coordinates, torsional angles, and distance maps from a learned latent space.”

• cites [53] sourceNot checked yet

  Not checked in the free run.

• cites [54]Not checked yet

  Not checked in the free run.

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

Dataset limitations include being dependent on a specific docking algorithm for data generation or having clonally related sequences in the dataset.

“Furthermore, dataset limitations include being dependent on a specific docking algorithm for data generation or having clonally related sequences in the dataset.”

May combine two claims.

The AntiBERTy model is an antibody-specific BERT-based model trained on 558 million antibody sequences.

“The AntiBERTy model [62], an antibody specific BERT-based model, was trained on 558 million antibody sequences, with embeddings that cluster into directed evolution trajectories and the ability to detect paratope”

• cites [62]Not checked yet

  Not checked in the free run.

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

At their core, such models learn interdependencies of antibody sequence; the rise of autoregressive or generative models then allowed for models that can generate 'antibody-like' sequences.

“At their core, such models learn interdependencies of antibody sequence; the rise of autoregressive or generative models then allowed for models that can generate 'antibody-like' sequences.”

Probably not a substantial claim: may not pass S3 (load-bearing).

Shuai et al. used the Immunoglobulin Language Model to great success in generating immunoglobulin sequences.

“Shuai et al. [63] used the Immunoglobulin Language Model (IgLM), to great success in generating immunoglobulin sequences.”

• cites [63]Not checked yet

  Not checked in the free run.

Probably not a substantial claim: may not pass S3 (load-bearing).

Prior to recent ML-based techniques, classical antibody design techniques involved classical sampling and scoring methods.

“Prior to the recent development of ML-based techniques for antibody design, more classical techniques involved the use of classical sampling and scoring methods.”

Probably not a substantial claim: may not pass S3 (load-bearing).

Chroma created proteins with several thousand amino acids.

“Very recently, RFdiffusion [76] and Chroma [77], both score-based generative models, have been able to generate all-atom coordinates for novel proteins, with Chroma creating proteins with several thousand amino acids.”

• cites [77]Not checked yet

  Not checked in the free run.

May combine two claims.

At the moment, purely sequence-based models are either antigen agnostic or need outside experimental data.

“However, at the moment, purely sequence-based models are either antigen agnostic or need outside experimental data and would benefit from incorporating structural information during the generation process.”

Probably not a substantial claim: may not pass S3 (load-bearing).

RosettaAntibodyDesign relies on two algorithmic loops referred to as the outer and inner loops of design.

“RosettaAntibodyDesign (RAbD) [89] relies on two algorithmic loops referred to the outer and inner loops of design.”

• cites [89] sourceNot checked yet

  Not checked in the free run.

Probably not a substantial claim: may not pass S3 (load-bearing).

Nimrod et al. proposed a strategy of re-epitoping using similarity to epitopes of known antibody-antigen complexes for antibody design.

“Nimrod et al. [92] proposed a strategy of re-epitoping [i.e., using similarity to epitopes of known antibodyantigen complexes for their antibody design strategy.”

• cites [92] sourceNot checked yet

  Not checked in the free run.

Probably not a substantial claim: may not pass S3 (load-bearing).

Guided search can be used to guide antibody design in structure-based generative frameworks.

“This type of guided search, based on Markov chain Monte Carlo methods or gradient updates, is often used for conditional generation [84] and can be used to guide antibody design in structure-based generative frameworks.”

• cites [84]Not checked yet

  Not checked in the free run.

Probably not a substantial claim: may not pass S3 (load-bearing).

Guided search based on Markov chain Monte Carlo methods or gradient updates is often used for conditional generation.

“This type of guided search, based on Markov chain Monte Carlo methods or gradient updates, is often used for conditional generation [84] and can be used to guide antibody design in structure-based generative frameworks.”

• cites [84]Not checked yet

  Not checked in the free run.

Probably not a substantial claim: may not pass S3 (load-bearing).

The problem of predicting a protein sequence that folds into a particular structure, inverse protein folding, has seen continuous improvements.

“The problem of predicting a protein sequence that folds into a particular structure (inverse protein folding) has, for example, seen continuous improvements; state-of-the-art approaches integrate contemporary architectures that capture protein invariances while reasoning about protein residues in local and global coordinate frames [79][80][81].”

• cites [79]Not checked yet

  Not checked in the free run.

• cites [80]Not checked yet

  Not checked in the free run.

• cites [81] sourceNot checked yet

  Not checked in the free run.

Probably not a substantial claim: may not pass S3 (load-bearing).

State-of-the-art inverse protein folding approaches integrate contemporary architectures that capture protein invariances while reasoning about protein residues in local and global coordinate frames.

“The problem of predicting a protein sequence that folds into a particular structure (inverse protein folding) has, for example, seen continuous improvements; state-of-the-art approaches integrate contemporary architectures that capture protein invariances while reasoning about protein residues in local and global coordinate frames [79][80][81].”

• cites [79]Not checked yet

  Not checked in the free run.

• cites [80]Not checked yet

  Not checked in the free run.

• cites [81] sourceNot checked yet

  Not checked in the free run.

Probably not a substantial claim: may not pass S3 (load-bearing).

The inverse protein folding problem directly relates to the design of a CDR sequence given the backbone structure.

“This problem directly relates to the design of a CDR sequence given the backbone structure.”

May combine two claims.

Codon optimization can result in an improvement in protein expression yield, antibody assembly, affinity, and biological activity.

“Codon optimization can not only result in an improvement in protein expression yield, but also antibody assembly, affinity, and biological activity [85,86].”

• cites [85] sourceNot checked yet

  Not checked in the free run.

• cites [86] sourceNot checked yet

  Not checked in the free run.