Luis Serrano
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Luis Serrano has a total of 438 co-authors. Here are the top co-authors by number of shared publications:
- Maria Lluch-Senar
Pulmobiotics, Dr Aiguader 88, Barcelona, Spain
27 shared publications - Martin H. Schaefer
Department of Experimental Oncology, Istituto Europeo di Oncologia IRCCS, Milan, Italy
15 shared publications - Samuel Miravet-Verde
ETH Zurich
14 shared publications - Xavier Hernandez-Alias
School of Computation, Information and Technology, Technical University of Munich, Munich, Germany
13 shared publications - Hannah Benisty
Centre for Genomic Regulation (CRG), The Barcelona Institute of Science and Technology, Dr. Aiguader 88, Barcelona 08003, Spain
11 shared publications - Javier Delgado
Centre for Genomic Regulation (CRG), The Barcelona Institute for Science and Technology Barcelona Spain
11 shared publications
Latest Publications
- Open Access
- 22 October 2026
- Published by Frontiers Media SA in Frontiers in Microbiology
- Vol. 17, 1950185
Integrative multiomic characterization of the methylome and restriction-modification repertoire of Mycoplasma hyorhinis SK76
Restriction-modification (R-M) systems are widespread in Mycoplasma species, contributing to defense against foreign DNA and epigenetic regulation. However, the limited genetic tractability of Mollicutes hinders the functional characterization of R-M systems. Here, we characterized the R-M repertoire of Mycoplasma hyorhinis SK76 by integrating comparative genomics, structural modeling, quantitative proteomics, and genome-wide methylome profiling. We found that SK76 encodes Type I, Type II and Type III R-M systems with some duplicated components. For Type I, HsdS2 was predicted to recognize a bipartite TA-N₆-ATC motif, which was detected as a secondary 6mA-associated methylation signature, whereas the remaining HsdS variants displayed features consistent with reduced functionality. Among the Type II R-M system, Mhy3 and its cognate endonuclease End3 were associated with the 5mC motif GATC, while the remaining predicted methyltransferases lacked both protein expression and corresponding methylation signatures. Regarding Type III, Mod1 was identified as the principal active adenine methyltransferase responsible for the asymmetric GGAT-centered methylation pattern, whereas additional Mod homologues likely represent phase-variable or conditionally regulated elements with limited activity. Overall, the functional methylome of M. hyorhinis represents only a subset of its genomic R-M repertoire, underscoring the value of integrating genomic, proteomic and methylome data to define active epigenetic systems.
- 22 October 2026
- Published by Frontiers Media SA in Frontiers in Microbiology
- Vol. 17, 1950185
Open AccessIntegrative multiomic characterization of the methylome and restriction-modification repertoire of Mycoplasma hyorhinis SK76
Restriction-modification (R-M) systems are widespread in Mycoplasma species, contributing to defense against foreign DNA and epigenetic regulation. However, the limited genetic tractability of Mollicutes hinders the functional characterization of R-M systems. Here, we characterized the R-M repertoire of Mycoplasma hyorhinis SK76 by integrating comparative genomics, structural modeling, quantitative proteomics, and genome-wide methylome profiling. We found that SK76 encodes Type I, Type II and Type III R-M systems with some duplicated components. For Type I, HsdS2 was predicted to recognize a bipartite TA-N₆-ATC motif, which was detected as a secondary 6mA-associated methylation signature, whereas the remaining HsdS variants displayed features consistent with reduced functionality. Among the Type II R-M system, Mhy3 and its cognate endonuclease End3 were associated with the 5mC motif GATC, while the remaining predicted methyltransferases lacked both protein expression and corresponding methylation signatures. Regarding Type III, Mod1 was identified as the principal active adenine methyltransferase responsible for the asymmetric GGAT-centered methylation pattern, whereas additional Mod homologues likely represent phase-variable or conditionally regulated elements with limited activity. Overall, the functional methylome of M. hyorhinis represents only a subset of its genomic R-M repertoire, underscoring the value of integrating genomic, proteomic and methylome data to define active epigenetic systems.
- Open Access
- 15 September 2026
- Published by Wiley in Protein Science
- Vol. 35 (10) , e70779
A curated structural dataset of peptide–protein complexes reveals biases in existing datasets and principles of peptide binding
Peptide–protein interactions are fundamental to many biological processes, and peptide design is gaining interest due to its therapeutic potential. This has led to the emergence of various structural databases, such as PepBDB and Peptipedia, that classify peptides as polypeptides with fewer than 50 amino acids. These databases provide valuable starting points for studying peptide recognition by protein partners and are widely used for machine learning applications, docking, and scoring functions. However, under such a length definition for peptides, we have very different cases that are likely to confound the analysis of peptide binding, including miniproteins, peptide–peptide complexes, intramolecular peptide disulfide bonds, intermolecular disulfide bridges, and proteins undergoing internal cleavage, such as serpins, as well as non‐natural amino acids and covalently bound cofactors. Here, we present a rigorous classification of peptide–protein complexes to generate datasets suitable for comparative energetic analysis with a focus on peptides that are unstructured in the absence of their target protein. The analysis of this dataset shows that peptide binding is typically driven by a small number of hotspot residues mainly enriched in aromatic and bulky hydrophobic side chains. Their number of hotspots and their spatial organization depend on peptide length, secondary structure, and covalent constraints. Short peptides rely on central anchor regions, whereas longer peptides distribute hotspots more broadly, with helices showing periodic spacing and β‐strands relying more on backbone‐mediated stabilization. Disulfide bonds further decrease the number of hotspots per peptide length by either pre‐organizing the peptide or acting as covalent anchors. This work provides a curated resource and general principles for peptide recognition. It highlights the importance of structurally classifying peptide–protein complexes to avoid bias in downstream computational and machine‐learning applications.
- 15 September 2026
- Published by Wiley in Protein Science
- Vol. 35 (10) , e70779
Open AccessA curated structural dataset of peptide–protein complexes reveals biases in existing datasets and principles of peptide binding
Peptide–protein interactions are fundamental to many biological processes, and peptide design is gaining interest due to its therapeutic potential. This has led to the emergence of various structural databases, such as PepBDB and Peptipedia, that classify peptides as polypeptides with fewer than 50 amino acids. These databases provide valuable starting points for studying peptide recognition by protein partners and are widely used for machine learning applications, docking, and scoring functions. However, under such a length definition for peptides, we have very different cases that are likely to confound the analysis of peptide binding, including miniproteins, peptide–peptide complexes, intramolecular peptide disulfide bonds, intermolecular disulfide bridges, and proteins undergoing internal cleavage, such as serpins, as well as non‐natural amino acids and covalently bound cofactors. Here, we present a rigorous classification of peptide–protein complexes to generate datasets suitable for comparative energetic analysis with a focus on peptides that are unstructured in the absence of their target protein. The analysis of this dataset shows that peptide binding is typically driven by a small number of hotspot residues mainly enriched in aromatic and bulky hydrophobic side chains. Their number of hotspots and their spatial organization depend on peptide length, secondary structure, and covalent constraints. Short peptides rely on central anchor regions, whereas longer peptides distribute hotspots more broadly, with helices showing periodic spacing and β‐strands relying more on backbone‐mediated stabilization. Disulfide bonds further decrease the number of hotspots per peptide length by either pre‐organizing the peptide or acting as covalent anchors. This work provides a curated resource and general principles for peptide recognition. It highlights the importance of structurally classifying peptide–protein complexes to avoid bias in downstream computational and machine‐learning applications.
- Open Access
Mechanistic pathway modeling reveals how IL-10 generates pleiotropic immune responses
IL-10 is a key anti-inflammatory cytokine whose activity is impaired in autoimmune diseases. However, IL-10 also promotes inflammation under certain conditions, limiting the efficacy of IL-10-based therapies. Because the principles underlying these opposing effects remain unclear, we developed a mathematical model of the IL-10 signaling pathway to understand how such pleiotropic responses arise. Considering that STAT3 signaling is buffered against changes in IL-10RB receptor affinity, we provide a minimal mechanistic explanation for IL-10 variants with anti-inflammatory or pro-inflammatory biased responses produced by an altered receptor affinity. By linking model-predicted pSTAT1 and pSTAT3 abundances with transcriptomic changes, we identified IL-10-responsive genes that are regulated at different pSTAT thresholds. This could explain how IL-10 elicits distinct downstream responses at different signaling strengths, and suggests that, depending on the required response, the affinity of IL10 for its receptors should not always be enhanced. Overall, our study highlights how modeling can help disentangle IL-10 pleiotropy to support the rational development of more effective IL-10-based therapies.
Open AccessMechanistic pathway modeling reveals how IL-10 generates pleiotropic immune responses
IL-10 is a key anti-inflammatory cytokine whose activity is impaired in autoimmune diseases. However, IL-10 also promotes inflammation under certain conditions, limiting the efficacy of IL-10-based therapies. Because the principles underlying these opposing effects remain unclear, we developed a mathematical model of the IL-10 signaling pathway to understand how such pleiotropic responses arise. Considering that STAT3 signaling is buffered against changes in IL-10RB receptor affinity, we provide a minimal mechanistic explanation for IL-10 variants with anti-inflammatory or pro-inflammatory biased responses produced by an altered receptor affinity. By linking model-predicted pSTAT1 and pSTAT3 abundances with transcriptomic changes, we identified IL-10-responsive genes that are regulated at different pSTAT thresholds. This could explain how IL-10 elicits distinct downstream responses at different signaling strengths, and suggests that, depending on the required response, the affinity of IL10 for its receptors should not always be enhanced. Overall, our study highlights how modeling can help disentangle IL-10 pleiotropy to support the rational development of more effective IL-10-based therapies.
- Open Access
- 12 March 2026
- Published by Springer Nature in Nature Communications
- Vol. 17 (1) , 1994
Exon inclusion signatures enable accurate estimation of splicing factor activity
Splicing factors control exon inclusion in messenger RNAs, shaping transcriptome and proteome diversity. Their catalytic activity is regulated by multiple layers, making single-omic measurements on their own fall short in identifying which splicing factors underlie a phenotype. Here, we posit that splicing factor activity, defined as a splicing factor’s ability to modulate exon inclusion, can be estimated from changes in exon inclusion signatures. To test this hypothesis, we benchmark methods for constructing splicing factor→exon networks and estimating splicing factor activity. We find that combining RNA-seq perturbation-based networks with VIPER (Virtual Inference of Protein Activity by Enriched Regulon analysis) accurately captures splicing factor activity as modulated by multiple regulatory layers. This approach integrates splicing factor regulation into a single score derived solely from exon inclusion signatures, allowing functional interpretation of heterogeneous conditions. As a proof of concept, we identify recurrent cancer splicing programs, revealing associations with oncogenic- and tumor suppressor-like splicing factors missed by conventional methods. These programs correlate with patient survival and key cancer hallmarks: initiation, proliferation, and immune evasion. Altogether, we show splicing factor activity can be accurately estimated from exon inclusion changes, enabling comprehensive analyses of splicing regulation with minimal data requirements.
- 12 March 2026
- Published by Springer Nature in Nature Communications
- Vol. 17 (1) , 1994
Open AccessExon inclusion signatures enable accurate estimation of splicing factor activity
Splicing factors control exon inclusion in messenger RNAs, shaping transcriptome and proteome diversity. Their catalytic activity is regulated by multiple layers, making single-omic measurements on their own fall short in identifying which splicing factors underlie a phenotype. Here, we posit that splicing factor activity, defined as a splicing factor’s ability to modulate exon inclusion, can be estimated from changes in exon inclusion signatures. To test this hypothesis, we benchmark methods for constructing splicing factor→exon networks and estimating splicing factor activity. We find that combining RNA-seq perturbation-based networks with VIPER (Virtual Inference of Protein Activity by Enriched Regulon analysis) accurately captures splicing factor activity as modulated by multiple regulatory layers. This approach integrates splicing factor regulation into a single score derived solely from exon inclusion signatures, allowing functional interpretation of heterogeneous conditions. As a proof of concept, we identify recurrent cancer splicing programs, revealing associations with oncogenic- and tumor suppressor-like splicing factors missed by conventional methods. These programs correlate with patient survival and key cancer hallmarks: initiation, proliferation, and immune evasion. Altogether, we show splicing factor activity can be accurately estimated from exon inclusion changes, enabling comprehensive analyses of splicing regulation with minimal data requirements.
- Open Access
- 16 December 2025
- Published by Springer Nature in Nature Communications
- Vol. 16 (1) , 11159
Sources of essential lipids for Mycoplasma pneumoniae via P116 to target liver and atherosclerotic lesions
Mycoplasma pneumoniae (MPN) is a bacterial pathogen that primarily causes atypical pneumonia. It cannot synthesize certain essential lipids and therefore relies on the host for their acquisition to survive. MPN has been detected in increased amounts within ruptured atherosclerotic plaques. In this work, we show that the protein P116 facilitates cholesterol acquisition from LDL, HDL and various cell types. Targeting P116’s C-terminal domain with a monoclonal antibody inhibits cholesterol acquisition and bacterial growth in vitro. Phase contrast epifluorescence microscopy of human arteries reveals that this antibody blocks MPN binding to atherosclerotic lesions ex vivo. Additionally, an MPN chassis injected into hyperlipidemic female mice localizes to the liver and atherosclerotic plaques. Here, we report that P116 plays a role in extracting essential lipids from lipoproteins and host cells and regulates MPN localization to atheromatous plaques. The study highlights MPN’s potential as a tool for targeting atherosclerotic lesions and fatty liver.
- 16 December 2025
- Published by Springer Nature in Nature Communications
- Vol. 16 (1) , 11159
Open AccessSources of essential lipids for Mycoplasma pneumoniae via P116 to target liver and atherosclerotic lesions
Mycoplasma pneumoniae (MPN) is a bacterial pathogen that primarily causes atypical pneumonia. It cannot synthesize certain essential lipids and therefore relies on the host for their acquisition to survive. MPN has been detected in increased amounts within ruptured atherosclerotic plaques. In this work, we show that the protein P116 facilitates cholesterol acquisition from LDL, HDL and various cell types. Targeting P116’s C-terminal domain with a monoclonal antibody inhibits cholesterol acquisition and bacterial growth in vitro. Phase contrast epifluorescence microscopy of human arteries reveals that this antibody blocks MPN binding to atherosclerotic lesions ex vivo. Additionally, an MPN chassis injected into hyperlipidemic female mice localizes to the liver and atherosclerotic plaques. Here, we report that P116 plays a role in extracting essential lipids from lipoproteins and host cells and regulates MPN localization to atheromatous plaques. The study highlights MPN’s potential as a tool for targeting atherosclerotic lesions and fatty liver.
- Open Access
- 14 October 2025
- Published by Oxford University Press (OUP) in Nucleic Acids Research
- Vol. 53 (19)
Using single-cell perturbation screens to decode the regulatory architecture of splicing factor programs
Splicing factors shape the isoform pool of most transcribed genes, playing a critical role in cellular physiology. Their dysregulation is a hallmark of diseases like cancer, where aberrant splicing contributes to progression. While exon inclusion signatures accurately assess changes in splicing factor activity, systematically mapping disease-driver regulatory interactions at large scale remains challenging. Perturb-seq, which combines CRISPR-based perturbations with single-cell RNA sequencing, enables high-throughput measurement of perturbed gene expression signatures but lacks exon-level resolution, limiting its application for splicing factor activity analysis. Here, we show that shallow artificial neural networks can estimate splicing factor activity from gene expression signatures, bypassing the need for exon-level data. As a case study, we map the genetic interactions regulating splicing factors during carcinogenesis, using the shift in splicing program activity—where oncogenic-like splicing factors become more active than tumor suppressor-like factors—as a molecular reporter of a Perturb-seq screen. Our analysis reveals a cross-regulatory network among splicing factors, involving protein–protein and splicing-mediated interactions, with MYC and additional candidate pathways linking cancer driver mutations to splicing regulation. This regulation recapitulates splicing program dynamics during development. Altogether, we establish a versatile framework for studying splicing factor regulation and demonstrate its utility for uncovering disease mechanisms.
- 14 October 2025
- Published by Oxford University Press (OUP) in Nucleic Acids Research
- Vol. 53 (19)
Open AccessUsing single-cell perturbation screens to decode the regulatory architecture of splicing factor programs
Splicing factors shape the isoform pool of most transcribed genes, playing a critical role in cellular physiology. Their dysregulation is a hallmark of diseases like cancer, where aberrant splicing contributes to progression. While exon inclusion signatures accurately assess changes in splicing factor activity, systematically mapping disease-driver regulatory interactions at large scale remains challenging. Perturb-seq, which combines CRISPR-based perturbations with single-cell RNA sequencing, enables high-throughput measurement of perturbed gene expression signatures but lacks exon-level resolution, limiting its application for splicing factor activity analysis. Here, we show that shallow artificial neural networks can estimate splicing factor activity from gene expression signatures, bypassing the need for exon-level data. As a case study, we map the genetic interactions regulating splicing factors during carcinogenesis, using the shift in splicing program activity—where oncogenic-like splicing factors become more active than tumor suppressor-like factors—as a molecular reporter of a Perturb-seq screen. Our analysis reveals a cross-regulatory network among splicing factors, involving protein–protein and splicing-mediated interactions, with MYC and additional candidate pathways linking cancer driver mutations to splicing regulation. This regulation recapitulates splicing program dynamics during development. Altogether, we establish a versatile framework for studying splicing factor regulation and demonstrate its utility for uncovering disease mechanisms.
- Open Access
- 13 August 2025
- Published by Springer Nature in Molecular Systems Biology
- Vol. 21 (10) , 1388-1416
Quantitative essentiality in a reduced genome: a functional, regulatory and structural fitness map
Essentiality studies have traditionally focused on coding regions, often overlooking other small genetic regulatory elements. To address this, we combined transposon libraries containing promoter or terminator sequences to obtain a high-resolution essentiality map of a genome-reduced bacterium, at near-single-nucleotide precision when considering non-essential genes. By integrating temporal transposon-sequencing data by k-means unsupervised clustering, we present a novel essentiality assessment approach, providing dynamic and quantitative information on the fitness contribution of different genomic regions. We compared the insertion tolerance and persistence of the two engineered libraries, assessing the local impact of transcription and termination on cell fitness. Essentiality assessment at the local base-level revealed essential protein domains and small genomic regions that are either essential or inaccessible to transposon insertion. We also identified structural regions within essential genes that tolerate transposon disruptions, resulting in functionally split proteins. Overall, this study presents a nuanced view of gene essentiality, shifting from static and binary models to a more accurate perspective. Additionally, it provides valuable insights for genome engineering and enhances our understanding of the biology of genome-reduced cells.
- 13 August 2025
- Published by Springer Nature in Molecular Systems Biology
- Vol. 21 (10) , 1388-1416
Open AccessQuantitative essentiality in a reduced genome: a functional, regulatory and structural fitness map
Essentiality studies have traditionally focused on coding regions, often overlooking other small genetic regulatory elements. To address this, we combined transposon libraries containing promoter or terminator sequences to obtain a high-resolution essentiality map of a genome-reduced bacterium, at near-single-nucleotide precision when considering non-essential genes. By integrating temporal transposon-sequencing data by k-means unsupervised clustering, we present a novel essentiality assessment approach, providing dynamic and quantitative information on the fitness contribution of different genomic regions. We compared the insertion tolerance and persistence of the two engineered libraries, assessing the local impact of transcription and termination on cell fitness. Essentiality assessment at the local base-level revealed essential protein domains and small genomic regions that are either essential or inaccessible to transposon insertion. We also identified structural regions within essential genes that tolerate transposon disruptions, resulting in functionally split proteins. Overall, this study presents a nuanced view of gene essentiality, shifting from static and binary models to a more accurate perspective. Additionally, it provides valuable insights for genome engineering and enhances our understanding of the biology of genome-reduced cells.
- Open Access
- 17 July 2025
- Published by Wiley in Protein Science
- Vol. 34 (8) , e70220
Predicting interacting hotspots for nanobodies' binding using triplets of residues
Protein–protein interactions (PPI) are fundamental to cellular signaling, forming robust networks that govern critical biological processes such as immune response, cell growth, and signal transduction. Nanobody‐based therapies have emerged as a key strategy for modulating PPIs, offering exceptional potential due to their high specificity, stability, and ability to access challenging epitopes on PPI interfaces inside cells. The rational design of nanobodies relies mainly on understanding and predicting their binding regions, particularly the residues that contribute the most to the binding energy (binding hotspots). Existing computational methods do not fully provide a scalable solution for hotspot identification in nanobody design, leaving a critical gap in the rational design of these therapeutics. Here, we present a scalable and structure‐aware algorithm for hotspot prediction in nanobody design. The algorithm queries a curated database of triplets of interacting residues obtained from ~20,000 non‐redundant PDB structures. We showed that these triplets contain structural and energetic information, being able to assess the stability effect of residue variations in protein structures, Pearson R = 0.63 (MSE = 1.58 kcal/mol). More important than effects on stability is the ability of the algorithm to predict binding hotspots of protein–protein generic complexes and more specifically in complexes containing nanobodies. HotspotPred reached an accuracy of 0.73 for hotspot residue identification in a protein interaction dataset of 1160 Alanine mutants and correctly identified in 63.4% of the cases we predicted at least 2 residues on the binding surface.
- 17 July 2025
- Published by Wiley in Protein Science
- Vol. 34 (8) , e70220
Open AccessPredicting interacting hotspots for nanobodies' binding using triplets of residues
Protein–protein interactions (PPI) are fundamental to cellular signaling, forming robust networks that govern critical biological processes such as immune response, cell growth, and signal transduction. Nanobody‐based therapies have emerged as a key strategy for modulating PPIs, offering exceptional potential due to their high specificity, stability, and ability to access challenging epitopes on PPI interfaces inside cells. The rational design of nanobodies relies mainly on understanding and predicting their binding regions, particularly the residues that contribute the most to the binding energy (binding hotspots). Existing computational methods do not fully provide a scalable solution for hotspot identification in nanobody design, leaving a critical gap in the rational design of these therapeutics. Here, we present a scalable and structure‐aware algorithm for hotspot prediction in nanobody design. The algorithm queries a curated database of triplets of interacting residues obtained from ~20,000 non‐redundant PDB structures. We showed that these triplets contain structural and energetic information, being able to assess the stability effect of residue variations in protein structures, Pearson R = 0.63 (MSE = 1.58 kcal/mol). More important than effects on stability is the ability of the algorithm to predict binding hotspots of protein–protein generic complexes and more specifically in complexes containing nanobodies. HotspotPred reached an accuracy of 0.73 for hotspot residue identification in a protein interaction dataset of 1160 Alanine mutants and correctly identified in 63.4% of the cases we predicted at least 2 residues on the binding surface.
- Open Access
- 16 July 2025
- Published by Wiley in Protein Science
- Vol. 34 (8) , e70210
Artificial intelligence and first‐principle methods in protein redesign: A marriage of convenience?
Since AlphaFold2's rise, many deep learning methods for protein design have emerged. Here, we validate widely used and recognized tools, compare them with first‐principle methods, and explore their combinations, focusing on their effectiveness in protein redesign and potential for therapeutic repurposing. We address two challenges: evaluating tools and combinations ability to detect the effects of multiple concurrent mutations in protein variants, and leveraging large‐scale datasets to compare modeling‐free methods, namely force fields, which handle point mutations well with limited backbone rearrangement, and inverse folding tools, which excel at native sequence recovery but may struggle with non‐natural proteins. Debuting TriCombine, a tool that identifies residue triangles in input structures, matches them to a structural database, and scores mutants based on substitution frequencies, we shortlisted candidates, modeled them with FoldX, and generated 16 SH3 mutants carrying up to 9 concurrent substitutions. The dataset was expanded to include 36 mutants and 11 crystal structures (7 newly solved), along with a parallel set of multiple non‐concurrent mutants from three additional proteins. For broader validation, we analyzed 160,000 four‐site GB1 mutants and 163,555 (single and double) variants across 179 natural and de novo domains. We show that combining AI‐based modeling tools with force field scoring functions yields the most reliable results. Inverse folding tools perform very well but lose accuracy on less‐represented proteins. First‐principle force fields like FoldX remain the most accurate for point mutations. All methods perform worse when applied to unsolved de novo models, underscoring the need for hybrid strategies in robust protein design.
- 16 July 2025
- Published by Wiley in Protein Science
- Vol. 34 (8) , e70210
Open AccessArtificial intelligence and first‐principle methods in protein redesign: A marriage of convenience?
Since AlphaFold2's rise, many deep learning methods for protein design have emerged. Here, we validate widely used and recognized tools, compare them with first‐principle methods, and explore their combinations, focusing on their effectiveness in protein redesign and potential for therapeutic repurposing. We address two challenges: evaluating tools and combinations ability to detect the effects of multiple concurrent mutations in protein variants, and leveraging large‐scale datasets to compare modeling‐free methods, namely force fields, which handle point mutations well with limited backbone rearrangement, and inverse folding tools, which excel at native sequence recovery but may struggle with non‐natural proteins. Debuting TriCombine, a tool that identifies residue triangles in input structures, matches them to a structural database, and scores mutants based on substitution frequencies, we shortlisted candidates, modeled them with FoldX, and generated 16 SH3 mutants carrying up to 9 concurrent substitutions. The dataset was expanded to include 36 mutants and 11 crystal structures (7 newly solved), along with a parallel set of multiple non‐concurrent mutants from three additional proteins. For broader validation, we analyzed 160,000 four‐site GB1 mutants and 163,555 (single and double) variants across 179 natural and de novo domains. We show that combining AI‐based modeling tools with force field scoring functions yields the most reliable results. Inverse folding tools perform very well but lose accuracy on less‐represented proteins. First‐principle force fields like FoldX remain the most accurate for point mutations. All methods perform worse when applied to unsolved de novo models, underscoring the need for hybrid strategies in robust protein design.
- Open Access
Artificial Intelligence And First Principle Methods In Protein Redesign: A Marriage Of Convenience?
Since AlphaFold2's rise, many deep learning methods for protein design have emerged. Here, we validate widely used and recognized tools, compare them with first-principle methods, and explore their combinations, focusing on their effectiveness in protein redesign and potential for therapeutic repurposing. We address two challenges: evaluating tools and combinations ability to detect the effects of multiple concurrent mutations in protein variants, and leveraging large-scale datasets to compare modeling-free methods, namely force fields, which handle point mutations well with limited backbone rearrangement, and inverse folding tools, which excel at native sequence recovery but may struggle with non-natural proteins. Debuting TriCombine, a tool that identifies residue triangles in input structures, matches them to a structural database, and scores mutants based on substitution frequencies, we shortlisted candidates, modeled them with FoldX, and generated 16 SH3 mutants carrying up to 9 concurrent substitutions. The dataset was expanded to include 36 mutants and 11 crystal structures (7 newly solved), along with a parallel set of multiple non-concurrent mutants from three additional proteins. For broader validation, we analyzed 160,000 four-site GB1 mutants and 163,555 (single and double) variants across 179 natural and de novo domains. We show that combining AI-based modeling tools with force field scoring functions yields the most reliable results. Inverse folding tools perform very well but lose accuracy on less-represented proteins. First-principle force fields like FoldX remain the most accurate for point mutations. All methods perform worse when applied to unsolved de novo models, underscoring the need for hybrid strategies in robust protein design.
Open AccessArtificial Intelligence And First Principle Methods In Protein Redesign: A Marriage Of Convenience?
Since AlphaFold2's rise, many deep learning methods for protein design have emerged. Here, we validate widely used and recognized tools, compare them with first-principle methods, and explore their combinations, focusing on their effectiveness in protein redesign and potential for therapeutic repurposing. We address two challenges: evaluating tools and combinations ability to detect the effects of multiple concurrent mutations in protein variants, and leveraging large-scale datasets to compare modeling-free methods, namely force fields, which handle point mutations well with limited backbone rearrangement, and inverse folding tools, which excel at native sequence recovery but may struggle with non-natural proteins. Debuting TriCombine, a tool that identifies residue triangles in input structures, matches them to a structural database, and scores mutants based on substitution frequencies, we shortlisted candidates, modeled them with FoldX, and generated 16 SH3 mutants carrying up to 9 concurrent substitutions. The dataset was expanded to include 36 mutants and 11 crystal structures (7 newly solved), along with a parallel set of multiple non-concurrent mutants from three additional proteins. For broader validation, we analyzed 160,000 four-site GB1 mutants and 163,555 (single and double) variants across 179 natural and de novo domains. We show that combining AI-based modeling tools with force field scoring functions yields the most reliable results. Inverse folding tools perform very well but lose accuracy on less-represented proteins. First-principle force fields like FoldX remain the most accurate for point mutations. All methods perform worse when applied to unsolved de novo models, underscoring the need for hybrid strategies in robust protein design.