Tokenizing loops of antibodies

Single-domain antibodies (sdAbs) are a promising class of biotherapeutics with unique structural traits within their paratope region. The distribution of canonical conformations explored by their complementarity determining region (CDR) loops differs to some extent from conventional two-chain Fv fragments of monoclonal antibodies (mAbs). In this study, we explored in detail the canonical structures of sdAb CDR-H1 and CDR-H2 loops and compared those with mAbs from the IGHV3 and IGHV1 gene families. We surveyed the antibody structures catalogued in SAbDab and clustered the CDR canonical loops in Cartesian space. While most of the sdAb clusters were sub-populations of previously defined canonical Fv conformations of CDR-H1 and CDR-H2, our stricter clustering approach defined narrower clusters in sequence-space. Meticulous visual inspection of sub-populations allowed a clearer understanding of sequence-structure relationships. The packing densities within structural pockets contacted by CDR-H1 and CDR-H2 canonical conformations were analyzed on the premise that these pockets cannot be left vacant as they would leave exposed supportive hydrophobic residues. The fine resolution of the canonical clusters defined here revealed unique signatures within these pockets, including distinct structural complementarities between CDR-H1 and CDR-H2 canonical clusters, which could not be perceived with the previous coarser clusters. We highlight examples where a single residue change in CDR-H1 sequence is sufficient to induce a dramatic population shift in CDR-H2 conformation. This suggests that preferences in combining CDR-H1 and CDR-H2 emerged naturally during antibody evolution, leading to preferred sets of conserved amino acids at key positions in the framework as well as within the CDR loops. We outline a game of musical chairs that is necessary to maintain the integrity of the antibody structures that arose during evolution. Our study also provides refined CDR-H1 and CDR-H2 structural templates for sdAb homology modeling that could be leveraged for improved antibody design.

Remote sensing image scene classification takes image blocks as classification units and predicts their semantic descriptors. Because it is difficult to obtain enough labeled samples for all classes of remote sensing image scenes, zero-shot classification methods which can recognize image scenes that are not seen in the training stage are of great significance. By projecting the image visual features and the class semantic features into the latent space and ensuring their alignment, the variational autoencoder (VAE) generative model has been applied to address remote-sensing image scene classification under a zero-shot setting. However, the VAE model takes the element-wise square error as the reconstruction loss, which may not be suitable for measuring the reconstruction quality of the visual and semantic features. Therefore, this paper proposes to augment the VAE models with the generative adversarial network (GAN) to make use of the GAN’s discriminator in order to learn a suitable reconstruction quality metric for VAE. To promote feature alignment in the latent space, we have also proposed cross-modal feature-matching loss to make sure that the visual features of one class are aligned with the semantic features of the class and not those of other classes. Based on a public dataset, our experiments have shown the effects of the proposed improvements. Moreover, taking the ResNet models of ResNet18, extracting 512-dimensional visual features, and ResNet50 and ResNet101, both extracting 2048-dimensional visual features for testing, the impact of the different visual feature extractors has also been investigated. The experimental results show that better performance is achieved by ResNet18. This indicates that more layers of the extractors and larger dimensions of the extracted features may not contribute to the image scene classification under a zero-shot setting.

The rising prevalence of vision-threatening eye diseases poses a major global health and economic burden, yet timely diagnosis remains limited by workforce shortages, diagnostic delays, and restricted access to specialized care. Artificial intelligence (AI) offers potential solutions. In particular, recent progress in foundation models and large language models—especially multimodal large language models (MLLMs)—has shown promise in medical image interpretation and automated clinical documentation. However, advancing MLLMs for ophthalmology is hindered by the lack of unified, comprehensive benchmark datasets for development and evaluation. Most existing benchmarks were designed for earlier models, which focused on narrow tasks or specific disease conditions. These benchmarks typically provide outputs in the form of disease labels rather than free-text responses. As a result, they are less suitable for assessing emerging generative models. In this work, we present LMOD+, a large-scale multimodal ophthalmology benchmark dataset comprising 32,633 instances with multi-granular annotations across 12 common ophthalmic conditions and 5 imaging modalities. The dataset integrates imaging, anatomical structures, demographics, and free-text annotations. It supports primary ophthalmic applications such as anatomical structure recognition, disease screening, disease staging, and demographic prediction for potential performance bias evaluation. Alongside the dataset, we introduce a systematic and unified data curation pipeline that repurposes existing or new datasets for MLLM development. LMOD+ extends our preliminary LMOD benchmark—the first multimodal ophthalmology benchmark for MLLMs—with three major enhancements. First, we expanded the dataset by nearly 50% (from 21,933 to 32,633 instances). The color fundus photography (CFP) modality, the most accessible imaging modality in ophthalmology, was significantly enlarged to cover a broader range of pathological conditions. Second, we broadened task coverage to include (a) 12 binary disease diagnosis tasks for prevalent conditions such as diabetic retinopathy, age-related macular degeneration, and retinal vein occlusion; (b) multi-class ophthalmic disease diagnosis; (c) disease severity classification, including a diabetic retinopathy staging task, which uses two internationally adopted grading standards: the international clinical diabetic retinopathy classification and the Scottish diabetic retinopathy grading scheme classification; and (d) demographic prediction (age and sex) to assess potential model bias. Third, we systematically evaluated 24 state-of-the-art MLLMs, including recent models from the InternVL, Qwen, and DeepSeek families. Our evaluations highlight both the promise and limitations of current MLLMs in ophthalmology. For example, Qwen-7B and InternVL achieved accuracies of 58.26% and 57.83% in disease screening under a zero-shot setting with a single model—a considerably more challenging paradigm than traditional fine-tuning, where separate models are trained for each specific task. InternVL also demonstrated potential in anatomical recognition. Nonetheless, overall performance remained suboptimal and often close to random baselines for challenging tasks such as disease staging, underscoring the substantial gap between general-domain MLLMs and the specialized requirements of ophthalmology. We publicly release the dataset, curation pipeline, and leaderboard to encourage community-wide development and evaluation of MLLMs, with the goal of advancing ophthalmic applications and ultimately reducing the global burden of vision-threatening diseases through AI. The dataset website, benchmark leaderboard, and download link are available at https://kfzyqin.github.io/lmod_plus .

Activation-induced deaminase (AID) mediates the somatic hypermutation (SHM) of Ig variable (V) regions that is required for the affinity maturation of the antibody response. An intensive analysis of a published database of somatic hypermutations that arose in the IGHV3-23*01 human V region expressed in vivo by human memory B cells revealed that the focus of mutations in complementary determining region (CDR)1 and CDR2 coincided with a combination of overlapping AGCT hotspots, the absence of AID cold spots, and an abundance of polymerase eta hotspots. If the overlapping hotspots in the CDR1 or CDR2 did not undergo mutation, the frequency of mutations throughout the V region was reduced. To model this result, we examined the mutation of the human IGHV3-23*01 biochemically and in the endogenous heavy chain locus of Ramos B cells. Deep sequencing revealed that IGHV3-23*01 in Ramos cells accumulates AID-induced mutations primarily in the AGCT in CDR2, which was also the most frequent site of mutation in vivo. Replacing the overlapping hotspots in CDR1 and CDR2 with neutral or cold motifs resulted in a reduction in mutations within the modified motifs and, to some degree, throughout the V region. In addition, some of the overlapping hotspots in the CDRs were at sites in which replacement mutations could change the structure of the CDR loops. Our analysis suggests that the local sequence environment of the V region, and especially of the CDR1 and CDR2, is highly evolved to recruit mutations to key residues in the CDRs of the IgV region.

A variable domain of heavy chain antibody (VHH) has different binding properties than conventional antibodies. Conventional antibodies prefer binding to the convex portion of the antigen, whereas VHHs prefer epitopes, such as crevices and clefts on the antigen. Therefore, developing candidates with the binding characteristics of camelid VHHs is important. Thus, To this end, a synthetic VHH library that reproduces the structural properties of camelid VHHs was constructed. First, the characteristics of VHHs were classified according to the paratope formation based on crystal structure analyses of the complex structures of VHHs and antigens. Then, we classified 330 complementarity-determining region 3 (CDR3) structures of VHHs from the Protein Data Bank (PDB) into three loop structures: Upright, Half-Roll, and Roll. Moreover, these structures depended on the number of amino acid residues within CDR3. Furthermore, in the Upright loops, several amino acid residues in the FR2 are involved in the paratope formation, along with CDR3, suggesting that the FR2 design in the synthetic library is important. A humanized synthetic VHH library, comprising two sub-libraries, Upright and Roll, was constructed and named PharmaLogical. A validation study confirmed that our PharmaLogical library reproduces VHHs with the characteristics of the paratope formation of the camelid VHHs, and shows good performance in VHH screening.

Improvement of anti- Burkholderia mouse monoclonal antibody from various phage-displayed single-chain antibody libraries

Multi-modal neural models that are able to encode and process both visual and textual data are becoming more and more common in the last few years. Such models enable new ways to learn the interaction between vision and text, and thus can be successfully applied to tasks of varying complexity in the domain of image and text classification. However, such models are traditionally oriented to learn grounded properties of images and of the objects they depict and less suited to solve tasks involving subjective characteristics, such as the emotions they can convey in viewers. In this paper, we provide some insights in the performances of the recently released OpenAI CLIP model for an emotion classification task. We evaluate the model both under zero-shot settings and via fine tuning on an image-emotion dataset. We compare the performances of CLIP both in a zero-shot and fine-tuning setting on (i) a standard benchmark dataset for object recognition (ii) an image-emotion dataset. Moreover, we evaluate to which extent a CLIP model adapted to emotions is able to retain general knowledge and generalization capabilities

A single-domain antibody fragment in complex with RNase A: non-canonical loop structures and nanomolar affinity using two CDR loops.

SDR grafting—a new approach to antibody humanization

The development of in vitro antibody selection technologies has allowed overcoming some limitations inherent to the hybridoma technology. In most cases, large repertoires of antibody genes have been assembled to create highly diversified libraries allowing the isolation of antibodies recognizing virtually any antigen. However, these universal libraries might not allow the isolation of antibodies with specific structural properties or particular amino acid contents that are rarely found in natural repertoires. Purpose-oriented libraries specially designed to incorporate desired characteristics have been successfully used. However, the workload required for library construction has limited the attractiveness of this approach compared to the use of large universal libraries. We have developed an approach to capture synthetic or natural diversity into the complementarity determining regions 3 (CDR3) of human antibody repertoires using Type IIS restriction enzymes. In this way, we generated several libraries either biased in amino acid content or towards long CDRH3 loops. The latter were successfully used to identify antibodies inhibiting the enzymatic activity of horseradish peroxidase, whereas libraries enriched in histidines allowed for the isolation of antibodies binding to human Fc in a pH-dependent manner. These libraries indicate that tailored diversification of CDR3 is sufficient to generate purpose-oriented libraries and isolate antibodies with uncommon properties.

Multimodal content generation, which leverages visual information to enhance the comprehension of cross-modal understanding, plays a critical role in Multimodal Information Retrieval. With the development of large language models (LLMs), recent research has adopted visual instruction tuning to inject the knowledge of LLMs into downstream multimodal tasks. The high complexity and great demand for resources urge researchers to study efficient distillation solutions to transfer the knowledge from pre-trained multimodal models.(teachers) to more compact student models. However, the instruction tuning for knowledge distillation in multimodal LLMs is resource-intensive and capability-restricted. The comprehension of students is highly reliant on the teacher models. To address this issue, we propose a novel Multimodal Distillation Calibration framework (MmDC). The main idea is to generate high-quality training instances that challenge student models to comprehend and prompt the teacher to calibrate the knowledge transferred to students, ultimately cultivating a better student model in downstream tasks. This framework comprises two stages: (1) multimodal alignment and (2) knowledge distillation calibration. In the first stage, parameter-efficient fine-tuning is used to enhance feature alignment between different modalities. In the second stage, we develop a calibration strategy to assess the student model's capability and generate high-quality instances to calibrate knowledge distillation from teacher to student. The experiments on diverse datasets show that our framework efficiently improves the student model's capabilities. Our 7B-size student model, after three iterations of distillation calibration, outperforms the current state-of-the-art LLaVA-13B model on the ScienceQA and LLaVA Test datasets and also exceeds other strong baselines in a zero-shot setting.

Molecular analysis of the heavy chain of antibodies that recognize the capsular polysaccharide of Neisseria meningitidis in hu-PBMC reconstituted SCID mice and in the immunized human donor

Phage display is commonly used to select target-binding antibody fragments from large libraries containing billions of unique antibody clones. In practice, selection outputs are often highly heterogenous, making it desirable to recover sequence information from the selected pool. Next Generation DNA Sequencing (NGS) enables the acquisition of sufficient sequencing reads to cover the pool diversity, however read-lengths are typically too short to capture paired antibody complementarity-determining regions (CDRs), which is needed to reconstruct target-binding antibody fragments. Here, we describe a simple in vitro protocol to bring the DNA encoding the antibody CDRs closer together. The final PCR product referred to as a "CDR strip" is suitable for short read-length NGS. In this method, phagemid ssDNA is recovered from antibody phage display biopanning and used as a template to create a heteroduplex with deletions between CDRs of interest. The shorter strand in the heteroduplex is preferentially PCR amplified to generate a CDR strip that is sequenced using NGS. We have also included a bioinformatics approach to analyze the CDR strip populations so that single antibody clones can be created from paired CDR sequences.

This paper introduces a neural style transfer model to generate a stylized image conditioning on a set of examples describing the desired style. The proposed solution produces high-quality images even in the zero-shot setting and allows for more freedom in changes to the content geometry. This is made possible by introducing a novel Two-Stage Peer-Regularization Layer that recombines style and content in latent space by means of a custom graph convolutional layer. Contrary to the vast majority of existing solutions, our model does not depend on any pre-trained networks for computing perceptual losses and can be trained fully end-to-end thanks to a new set of cyclic losses that operate directly in latent space and not on the RGB images. An extensive ablation study confirms the usefulness of the proposed losses and of the Two-Stage Peer-Regularization Layer, with qualitative results that are competitive with respect to the current state of the art using a single model for all presented styles. This opens the door to more abstract and artistic neural image generation scenarios, along with simpler deployment of the model.

Humanization of murine monoclonal antibodies for human therapy has commonly been achieved by complementarity-determining region (CDR) grafting, in which murine CDR loops are grafted onto human framework regions. Difficulties with that method have revealed the importance of certain framework residues in determining both the 3-D structure of CDR loops and the overall affinity of the molecule for its specific ligand. In the general model of structure-function relationships presented here, each amino acid position in the variable region is classified according to the benefit of achieving a more human-like antibody versus the risk of decreasing or abolishing specific binding affinity. Substitutions of human residues at low-risk positions (exposed to solvent but not contributing to antigen binding or antibody structure) are likely to decrease immunogenicity with little or no effect on binding affinity. Changes at high-risk positions (directly involved in antigen binding, CDR stabilization or internal packing) are avoided to preserve the biological activity of the antibody. Moderate-risk changes are made with caution. This model has been tested experimentally using H65, an anti-CD5 murine monoclonal antibody, whose binding activity had been greatly reduced by two previous attempts at humanization by conventional CDR grafting. The new 'human-engineered' H65 antibody containing 20 low-risk human consensus substitutions (expressed as either IgG or Fab) retains the full binding avidity of parental murine and chimeric H65 antibodies. A human-engineered antibody with an additional 14 moderate-risk substitutions has unexpectedly enhanced avidity (3- to 7-fold). This method is generally applicable to the design of other human-engineered antibodies with therapeutic potential.