Biology subjects

Lemmin, T.

Publications and source records attributed to Lemmin, T..

12 recordsLinked to original sources

bioRxiv2026

Structure-derived synthetic sequences guide a protein language model toward metalloproteins

MotivationProtein language models (pLMs) capture evolutionary sequence constraints but are limited in modeling underrepresented functional classes due to training data imbalance. Metalloproteins constitute a fundamental but sparsely represented class in sequence databases. We therefore assess whether structure-conditioned synthetic sequences can be used to specialize pLMs toward metal-binding functionality. ResultsWe fine-tuned the generalist model ProtGPT2 on synthetic sequences generated by the inverse-folding model ProteinMPNN, constructing training sets with controlled variation in size and diversity. Fine-tuning increased recovery of canonical metal-binding motifs from 43% in the baseline model to 91% in the fine-tuned models. Generated sequences retained high predicted structural confidence and structural similarity to known folds, despite low sequence identity. Analysis of latent representations from ProtGPT2 indicated that fine-tuned models occupy distinct regions of embedding space relative to both the baseline model and structure-conditioned sequences, consistent with partial incorporation of structural constraints while preserving sequence diversity. A multi-step filtering pipeline applied to sequences lacking canonical motifs identified candidate metal-binding sites in four-helical bundle topologies not detected in a non-redundant subset of Protein Data Bank structures or in AlphaFold-predicted proteomes. Availability and implementationCode, trained models, and datasets are available at: https://doi.org/10.5281/zenodo.18672158 and https://huggingface.co/gsgueglia.

bioinformatics↗

bioRxiv2025

H3BERTa: A CDR-H3 specific language model for antibody repertoire analysis

AbstractAntibodies are central to immune defense and therapeutic design, yet predicting which sequences confer functional activity remains challenging. Deep learning models trained on full variable regions often struggle due to sparse experimental data, signal dilution from conserved framework residues, and the extreme diversity of hypervariable loops. The heavy-chain complementarity-determining region 3 (CDR-H3) is the most variable segment shaping antigen specificity and driving immune diversity. Here, we present H3BERTa, a transformer-based language model trained solely on CDR-H3 sequences, to test whether this short region alone encodes enough biologically meaningful information. H3BERTa embeddings recapitulate biologically relevant sequence features, including J-gene usage and inferred B-cell maturation state. We further show that pseudo-perplexity profiles can be used to analyze repertoires, distinguishing healthy from HIV-1-derived sequences and suggesting measurable immune response signatures. Finally, these embeddings can support classifiers for broadly neutralizing antibodies (bnAbs) using limited labeled sequences, demonstrating their potential for accelerating antibody discovery. Together, our results indicate that the CDR-H3 region alone encodes a rich immunological signature, which H3BERTa robustly captures, providing a focused computational tool for analyzing repertoire diversity and informing antibody engineering.

immunology↗

bioRxiv2025

Deep generative modeling reveals maturation-linked pairing signatures in human antibodies

Understanding how antibody heavy and light chains pair is critical for decoding immune repertoire architecture and designing therapeutic antibodies. However, most antibody sequence databases lack paired chain information. To address this gap, we developed a two-stage deep learning framework. First, we pre-trained separate transformer-based language models on large corpora of unpaired heavy and light chain sequences to capture patterns of gene usage and somatic hypermutation. These models were then integrated via lightweight adapters into a sequence-to-sequence model trained in a machine translation setting, enabling light chain generation conditioned on heavy chain input. Although native light chain recovery was moderate, the model consistently captured functionally meaningful constraints: generated sequences exhibited high germline identity, improved structural quality of predicted folds, and broader coverage of framework and CDR regions. Immunologically, heavy chains from memory B cells preferentially generated light chains with more restricted V gene usage, reflecting maturation-dependent selection. Additionally, generated{kappa} light chains displayed a trimodal similarity distribution, suggesting distinct functional pairing modes ranging from promiscuous to highly specific. This work shows that sequence-to-sequence modeling can uncover inter-chain dependencies and generate structurally and immunologically plausible antibody pairs, providing a foundation for computational repertoire analysis and therapeutic design. HighlightsO_LIA deep generative modeling framework enables conditional generation of light chains from heavy chains, leveraging unpaired data. C_LIO_LIConditioning enhances the structural quality and germline coherence of predicted antibodies. C_LIO_LIMemory B cell-derived heavy chains preferentially generate light chains with restricted V gene usage, consistent with maturation-dependent selection. C_LIO_LIGenerated{kappa} light chains show a trimodal similarity distribution, suggesting discrete pairing modes ranging from promiscuous to highly specific. C_LI

bioinformatics↗

bioRxiv2025

Beyond performance: How design choices shape chemical language models

Chemical language models (CLMs) have shown strong performance in molecular property prediction and generation tasks. However, the impact of design choices, such as molecular representation format, tokenization strategy, and model architecture, on both performance and chemical interpretability remains underexplored. In this study, we systematically evaluate how these factors influence CLM performance and chemical understanding. We evaluated models through finetuning on downstream tasks and probing the structure of their latent spaces using simple classifiers and dimensionality reduction techniques. Despite similar performance on downstream tasks across model configurations, we observed substantial differences in the structure and interpretability of their internal representations. SMILES molecular representation format with atomwise tokenization strategy consistently produced more chemically meaningful embeddings, while models based on BART and RoBERTa architectures yielded comparably interpretable representations. These findings highlight that design choices meaningfully shape how chemical information is represented, even when external metrics appear unchanged. This insight can inform future model development, encouraging more chemically grounded and interpretable CLMs. Scientific ContributionThis study systematically evaluates how core design choices influence chemical language models. Although the performances on downstream tasks were often similar across configurations, we observed substantial differences in internal representations with atomwise tokenized SMILES representations producing more chemically structured latent spaces than representations based on SELFIES. By clarifying the effects of molecular representation format and tokenization strategy, our findings provide actionable guidance for the more informed and interpretable design of future CLMs. Graphical Abstract O_FIG O_LINKSMALLFIG WIDTH=200 HEIGHT=80 SRC="FIGDIR/small/655735v1_ufig1.gif" ALT="Figure 1"> View larger version (23K): org.highwire.dtl.DTLVardef@11c387org.highwire.dtl.DTLVardef@3be797org.highwire.dtl.DTLVardef@e197c2org.highwire.dtl.DTLVardef@b264e2_HPS_FORMAT_FIGEXP M_FIG C_FIG

bioinformatics↗

bioRxiv2024

Large transient assemblies of Apaf1 constitute the apoptosome in cells

Upon cell death signals, the apoptotic protease-activating factor Apaf1 and cytochrome c interact to form the apoptosome complex. The apoptosome is crucial for mitochondrial apoptosis, as it activates caspases that dismantle the cell. However, the assembly mechanism and appearance of the apoptosome in vivo remain unclear. We show that upon onset of apoptosis, Apaf1 molecules accumulate into multiple foci per cell. Disassembly of the foci is linked to survival of the cell. Structurally, Apaf1 foci resemble organelle-sized, cloud-like assemblies. Foci form upon specific molecular interactions with cytochrome c and depending on procaspase-9. We propose that Apaf1 foci correspond to the apoptosome in cells. Transientness and ultrastructure of Apaf1 foci suggest that the dynamic spatiotemporal organisation of apoptosome components regulates progression of apoptosis.

cell biology↗

bioRxiv2024

Atomic insights into the signaling landscape of E. coli PhoQ Histidine Kinase from Molecular Dynamics simulations

Bacteria rely on two-component systems to sense environmental cues and regulate gene expression for adaptation. The PhoQ/PhoP system exemplifies this crucial role, playing a key part in sensing magnesium (Mg2+) levels, antimicrobial peptides, mild acidic pH, osmotic upshift, and long-chain unsaturated fatty acids, promoting virulence in certain bacterial species. However, the precise details of PhoQ activation remain elusive. To elucidate PhoQs signaling mechanism at atomic resolution, we combined AlphaFold2 predictions with molecular modeling and carried out extensive Molecular Dynamics (MD) simulations. Our MD simulations revealed three distinct PhoQ conformations that were validated by experimental data. Notably, one conformation was characterized by Mg2+ bridging the acidic patch in the sensor domain to the membrane, potentially representing a repressed state. Furthermore, the high hydration observed in a putative intermediate state lends support to the hypothesis of water-mediated conformational changes during PhoQ signaling. Our findings not only revealed specific conformations within the PhoQ signaling pathway, but also hold significant promise for understanding the broader histidine kinase family due to their shared structural features. Our approach paves the way for a more comprehensive understanding of histidine kinase signaling mechanisms across various bacterial species and opens the door for developing novel therapeutics that target PhoQ modulation.

molecular biology↗

bioRxiv2024

TemBERTure: Advancing protein thermostability prediction with Deep Learning and attention mechanisms

Understanding protein thermostability is essential for various biotechnological and biological applications. However, traditional experimental methods for assessing this property are time-consuming, expensive, and error-prone. Recently, the application of Deep Learning techniques from Natural Language Processing (NLP) was extended to the field of biology, with an emphasis on protein modeling. From a linguistic perspective, the primary sequence of proteins can be viewed as a string of amino acids that follow a physicochemical grammar. This study explores the potential of Deep Learning models trained on protein sequences to predict protein thermostability which provide improvements with respect to current approaches. We implemented TemBERTure, a Deep Learning framework to classify the thermal class (non-thermophilic or thermophilic) and predict and melting temperature of a protein, based on its primary sequence. Our findings highlight the critical role that data diversity plays on training robust models. Models trained on datasets with a wider range of sequences from various organisms exhibited superior performance compared to those with limited diversity. This emphasizes the need for a comprehensive data curation strategy that ensures a balanced representation of diverse species in the training data, to avoid the risk that the model focuses on recognizing the evolutionary lineage of the sequence rather than the intrinsic thermostability features. In order to gain more nuanced insights into protein thermostability, we propose leveraging attention scores within Deep Learning models to gain more nuanced insights into protein thermostability. We show that analyzing these scores alongside the 3D protein structure could offer a better understanding of the complex interplay between amino acid properties, their positioning, and the surrounding microenvironment, all crucial factors influencing protein thermostability. This work sheds light on the limitations of current protein thermostability prediction methods and introduces new avenues for exploration. By emphasizing data diversity and utilizing refined attention scores, future research can pave the way for more accurate and informative methods for predicting protein thermostability. Availability and ImplementationTemBERTure model and the data are available at https://github.com/ibmm-unibe-ch/TemBERTure

bioinformatics↗

bioRxiv2024

Dissecting AlphaFolds Capabilities with Limited Sequence Information

Protein structure prediction, a fundamental challenge in computational biology, aims to predict a proteins 3D structure from its amino acid sequence. This structure is pivotal for elucidating protein functions, interactions, and driving innovations in drug discovery and enzyme engineering. AlphaFold2, a powerful deep learning model, has revolutionized this field by leveraging phylogenetic information from multiple sequence alignments (MSAs) to achieve remarkable accuracy in protein structure prediction. However, a key question remains: how well does AlphaFold2 understand protein structures? This study investigates AlphaFold2s capabilities when relying primarily on high-quality template structures, without the additional information provided by MSAs. By designing experiments that probe local and global structural understanding, we aimed to dissect its dependence on specific features and its ability to handle missing information. Our findings revealed AlphaFold2s reliance on sterically valid C-{beta} atoms for correctly interpreting structural templates. Additionally, we observed its remarkable ability to recover 3D structures from certain perturbations and the negligible impact of the previous structure in recycling. Collectively, these results support the hypothesis that AlphaFold2 has learned an accurate local biophysical energy function. However, this function seems most effective for local interactions. Our work significantly advances understanding of how deep learning models predict protein structures and provides valuable guidance for researchers aiming to overcome limitations in these models. protein folding, alphafold, side-chain, interpretability

bioinformatics↗

bioRxiv2022

The inhibitory mechanism of a small protein reveals its role in antimicrobial peptide sensing

A large number of small membrane proteins have been discovered in bacteria, but their mechanism of action has remained mostly elusive. Here, we investigate the mechanism of a physiologically important small protein, MgrB, which represses the activity of the sensor kinase PhoQ and is widely distributed among enterobacteria. The PhoQ/PhoP two-component system is a master regulator of the bacterial virulence program and interacts with MgrB to modulate bacterial virulence, fitness, and drug resistance. A combination of crosslinking approaches with functional assays and protein dynamic simulations revealed structural rearrangements due to interactions between MgrB and PhoQ near the membrane/periplasm interface and along the transmembrane helices. These interactions induce the movement of the PhoQ catalytic domain and the repression of its activity. Without MgrB, PhoQ appears to be much less sensitive to antimicrobial peptides, including the commonly used C18G. In the presence of MgrB, C18G promotes MgrB to dissociate from PhoQ, thus activating PhoQ via derepression. Our findings reveal the inhibitory mechanism of the small protein MgrB and uncover its importance in antimicrobial peptide sensing. Significance StatementSmall proteins have high prevalence, vast diversity, and primarily regulatory functions in biological processes across all domains of life. However, their mechanisms of action remain largely elusive. In this study, we investigate the mechanism of the small protein, MgrB. It interacts with the sensor kinase PhoQ, rearranges its conformation, represses its kinase activity, and regulates bacterial response to environmental changes. In particular for antimicrobial peptides, MgrB is required for bacteria to have a selective response to this host-exclusive stimulus. Our findings underline the importance of a small protein in bacterial fitness and drug resistance and provide a molecular basis for engineering novel peptide-based regulators.

microbiology↗

bioRxiv2021

Allosteric mechanism of signal transduction in the two-component system histidine kinase PhoQ

Transmembrane signaling proteins couple extracytosolic sensors to cytosolic effectors. Here, we examine how binding of Mg2+ to the sensor domain of an E. coli two component histidine kinase (HK), PhoQ, modulates its cytoplasmic kinase domain. We use cysteine-crosslinking and reporter-gene assays to simultaneously and independently probe the signaling state of PhoQs sensor and autokinase domains in a set of over 30 mutants. Strikingly, conservative single-site mutants distant from the sensor or catalytic site strongly influence PhoQs ligand-sensitivity as well as the magnitude and direction of the signal, endowing diverse signaling characteristics without need for epistasis. Data from 35 mutants are explained by a semi-empirical 3-domain model in which the sensor, intervening HAMP, and catalytic domains can adopt kinase-promoting or inhibiting conformations, that are in allosteric communication. The catalytic and sensor domains intrinsically favor a constitutively kinase-on conformation, while the HAMP favors the off state; when coupled, they create a bistable system responsive to physiological [Mg2+]. Mutants alter signaling by locally modulating these intrinsic equilibrium constants and couplings. Our model suggests signals transmit via interdomain allostery rather than propagation of a single concerted conformational change, explaining the diversity of signaling structural transitions observed in individual HK domains.

biochemistry↗

bioRxiv2020

RosENet: Improving binding affinity prediction by leveraging molecular mechanics energies with a 3D Convolutional Neural Network

The worldwide increase and proliferation of drug resistant microbes, coupled with the lag in new drug development represents a major threat to human health. In order to reduce the time and cost for exploring the chemical search space, drug discovery increasingly relies on computational biology approaches. One key step in these approaches is the need for the rapid and accurate prediction of the binding affinity for potential leads. Here, we present RosENet (Rosetta Energy Neural Network), a three-dimensional (3D) Convolutional Neural Network (CNN), which combines voxelized molecular mechanics energies and molecular descriptors for predicting the absolute binding affinity of protein - ligand complexes. By leveraging the physico-chemical properties captured by the molecular force field, our model achieved a Root Mean Square Error (RMSE) of 1.26 on the PDBBind v2016 core set. We also explored some limitations and the robustness of the PDBBind dataset and our approach, on nearly 500 structures, including structures determined by Nuclear Magnetic Resonance and virtual screening experiments. Our study demonstrated that molecular mechanics energies can be voxelized and used to help improve the predictive power of the CNNs. In the future, our framework can be extended to features extracted from other biophysical and biochemical models, such as molecular dynamics simulations. Availabilityhttps://github.com/DS3Lab/RosENet

biochemistry↗

bioRxiv2018

Epitope-based vaccine design yields fusion peptide-directed antibodies that neutralize diverse strains of HIV-1

A central goal of HIV-1-vaccine research is the elicitation of antibodies capable of neutralizing diverse primary isolates of HIV-1. Here we show that focusing the immune response to exposed N-terminal residues of the fusion peptide, a critical component of the viral entry machinery and the epitope of antibodies elicited by HIV-1 infection, through immunization with fusion peptide-coupled carriers and prefusion-stabilized envelope trimers, induces cross-clade neutralizing responses. In mice, these immunogens elicited monoclonal antibodies capable of neutralizing up to 31% of a cross-clade panel of 208 HIV-1 strains. Crystal and cryo-electron microscopy structures of these antibodies revealed fusion peptide-conformational diversity as a molecular explanation for the cross-clade neutralization. Immunization of guinea pigs and rhesus macaques induced similarly broad fusion peptide-directed neutralizing responses suggesting translatability. The N terminus of the HIV-1-fusion peptide is thus a promising target of vaccine efforts aimed at eliciting broadly neutralizing antibodies.

immunology↗