Biology subjects

Kim, H.

Publications and source records attributed to Kim, H..

At least 19 recordsLinked to original sources

bioRxiv2026

Imbalance between RNA-binding proteins and the WAVE regulatory complex underlies the toxic gain-of-function effects of the CYFIP2 Arg87Cys variant in neurodevelopmental disorders

Pathogenic variants in cytoplasmic FMR1-interacting protein 2 (CYFIP2) cause neurodevelopmental disorders, with hotspot substitutions at arginine 87 (Arg87) associated with particularly severe symptoms. CYFIP2 functions both as a component of the WAVE (Wiskott-Aldrich syndrome protein family verprolin-homologous protein) regulatory complex (WRC), regulating the actin cytoskeleton, and as an interactor of RNA-binding proteins (RBPs) involved in mRNA processing/translation. Although Arg87 variants reduce CYFIP2 protein stability, heterozygous knock-in mice carrying Arg87Cys (Cyfip2+/R87C) show markedly more severe phenotypes than heterozygous knockout (Cyfip2+/-) mice, supporting a toxic gain-of-function mechanism. However, how the dual roles of CYFIP2 contribute to this mechanism remains unclear. To address this, we examined Cyfip2+/R87C mice at postnatal week 7, before overt neuronal loss and gliosis, focusing on the progressive dispersion of hippocampal cornu ammonis 1 (CA1) pyramidal neurons, a phenotype absent in Cyfip2+/- mice. Single-nucleus RNA sequencing suggested candidate alterations related to translation and cell adhesion/migration in CA1 pyramidal neurons. Complementarily, human embryonic kidney 293T cells expressing CYFIP2 Arg87Cys showed defects in proliferation and morphology, accompanied by altered expression of cell adhesion-related genes. Quantitative proteomic analysis of the CA1 region uncovered an imbalance between CYFIP2-associated protein networks, characterized by increased RBPs and reduced WRC components. Phosphoproteomic analysis further revealed altered phosphorylation of synaptic and translational regulators. Consistently, whole-cell recordings and biocytin-based morphological analyses showed increased frequency of miniature excitatory postsynaptic currents and altered dendritic morphology in Cyfip2+/R87C CA1 pyramidal neurons, most prominently in the more severely dispersed deep-layer neurons. Moreover, the cell adhesion molecule Eph receptor A3 was reduced, possibly as a result of RBP dysregulation. Together, these findings suggest that an imbalance between RBP- and WRC-associated pathways underlies the gain-of-function effects of CYFIP2 Arg87Cys and contributes to early neuronal phenotypes, most notably CA1 layer dispersion. This mechanism, distinct from CYFIP2 haploinsufficiency, supports allele-specific therapeutic strategies for Arg87 variants.

neuroscience↗

bioRxiv2026

Live Spatial Transcriptomics Enables Longitudinal Tissue Mapping

Spatial transcriptomics has become an important technology in biomedical research by preserving the spatial context of gene expression in tissues and enabling the study of cellular states and interactions in situ. However, most existing methods provide only static snapshots of individual tissue specimens, requiring inference of the molecular dynamics underlying development, disease progression, and therapeutic response from different specimens rather than direct tracking in the same living tissue. Here, we introduce Live Spatial Transcriptomics (Live-ST), a technology that enables longitudinal, transcriptome-wide profiling of living tissue. Live-ST combines electroporation-based intracellular sampling and spatial barcoding to enable repeated, spatially resolved mRNA measurements in intact living tissue through next-generation sequencing. We demonstrate its capabilities by (i) benchmarking spatial mRNA capture in patterned two-dimensional cell systems, (ii) spatially profiling transcriptomes in intact, patient-derived living melanoma tissues while preserving tissue viability, and (iii) longitudinally tracking drug-response dynamics within the tumor microenvironment. Together, Live-ST enables repeated, non-destructive measurements of spatial gene expression, providing a new framework for studying dynamic tissue biology in native physiological and pathological contexts.

bioengineering↗

bioRxiv2026

Vaccination targeting the novel endogenous tumor-associated antigen Gpr149 induces protective immunity against checkpoint inhibitor-refractory SB28 glioblastoma

Background: Glioblastoma is a lethal brain tumor notable for its limited spontaneous induction of CD8+ T cell responses. Vaccination may overcome this limitation by generating tumor-specific CD8+ T cells capable of mediating anti-tumor immunity in the brain. However, preclinical vaccine development is constrained by a lack of immunologically stringent glioblastoma models with defined endogenous antigens. We therefore sought to identify endogenous, vaccine-targetable antigens expressed by the immune checkpoint-refractory SB28 syngeneic mouse glioblastoma model and to evaluate their targetability using a heterologous prime-boost regimen with the chimpanzee adenovirus ChAdOx1 and the modified vaccinia Ankara (MVA) viral vector-based vaccines. Methods: RNA sequencing and whole-exome sequencing were used to identify candidate tumor-associated antigens (TAAs) and neoantigens, and tumor-specific expression was validated by RT-qPCR. ChAdOx1 and MVA vectors encoding candidate antigens were generated and evaluated for immunogenicity and anti-tumor efficacy. Antigen-specific CD8+ T cell responses and minimal epitopes were characterized by ex vivo restimulation, MHC-I binding assays, tetramer staining, and tumor cytotoxicity assays. Results: We identified five candidate TAAs (Gpr149, Cdh6, Fbn2, Igf2bp1, and Shox2) but no suitable neoantigen targets in SB28. Vaccination induced CD8+ T cell responses against the candidate TAAs, and immunogenic minimal epitopes were identified for each. Among the candidates, Gpr149 elicited the strongest vaccine-induced CD8+ T cell responses. In mice challenged intracranially with SB28, Gpr149-targeted ChAdOx1/MVA vaccination combined with immune checkpoint blockade significantly prolonged median survival and resulted in long-term survival in 50% of mice, whereas vaccination targeting the other TAAs did not improve survival. CD8+ T cells recognizing the immunodominant Gpr149128-135 epitope (VCYNFYTM) specifically recognized and lysed wild-type SB28 cells, but not Gpr149-knockout SB28 cells, confirming endogenous processing and presentation of the targeted epitope. Conclusions: Gpr149 is an endogenous SB28 tumor antigen that can be effectively targeted by ChAdOx1/MVA vaccination to generate protective anti-tumor immunity against intracranial glioblastoma. Protection was associated with a dominant CD8+ T cell response to the Gpr149128-135 epitope VCYNFYTM. These findings establish endogenously expressed Gpr149 in SB28 as a physiologically relevant model for evaluating antigen-targeted immunotherapies and provide a systematic approach for identifying endogenous vaccine targets in other preclinical models.

immunology↗

bioRxiv2026

Expanding the therapeutic window of bacteria for cancer therapy

Engineered bacteria enable self-amplifying therapeutic delivery and immune activation, yet these same features drive systemic toxicity and off-target growth. Here, we create a "tumorobe" that overcomes these limitations by integrating reduced immunogenicity, tumour-restricted growth, and enhanced payload release. Specifically, we identify lipopolysaccharide modifications that reduce immunogenicity while preserving fitness, restrict growth to tumours using a hypoxia-lactate "AND" gate, and enhance extracellular payload release via lipoprotein deletion. This strain expands the maximum tolerated dose in a route-dependent manner by up to 1,000-fold. Across multiple murine models, including colorectal cancer and glioblastoma, the strain drives robust antitumour efficacy as both a standalone therapeutic and an immunotherapeutic delivery platform, achieving durable complete responses in up to 75% of colorectal tumours when armed with a cytokine payload. Together, these results establish a broadly applicable platform for safe and effective bacterial cancer therapy.

synthetic biology↗

bioRxiv2026

Engineered probiotics target granulomas to treat mycobacterial infection

Granulomas are a pathological niche for mycobacteria that remain difficult to access with conventional therapies. Here we show that probiotic Escherichia coli Nissle 1917 (EcN) infiltrates and colonizes granulomas and can be engineered as a delivery system that senses the local microenvironment and releases antimycobacterial payloads. Specifically, we leveraged the hypoxic microenvironment of necrotic granulomas to restrict bacterial growth, and developed a dual-mode lysis circuit combining intracellular and extracellular killing for sustained payload release with limited evolutionary escape. We further engineered EcN to express a mycobacteriophage-derived enzyme, generating a strain that suppressed drug-sensitive and drug-resistant mycobacteria in vitro and potentiated conventional antibiotic activity. Engineered probiotics also reduced mycobacterial burden in vivo, and in a murine model of rifampicin-resistant infection, combination treatment with pretomanid achieved greater bacterial reduction than either monotherapy. Together, these results establish a synthetic biology framework for targeting granulomatous infections with programmable living microbes.

synthetic biology↗

bioRxiv2026

Predicting conformational changes induced by biomolecular interactions

Biomolecular interactions trigger structural changes of protein complexes essential for cellular signaling, regulation, and drug action. Modern co-folding models can now generate detailed structures of multi-component complexes directly from sequence. However, recent evaluations reveal that they frequently fail to capture specific conformational states induced by a ligand or partner. Here, we show that explicitly modeling the structural transition from an unbound to a bound state, rather than generating the bound complex alone, better resolves these interaction-induced states. Our generative learning framework, K-Fold, transforms independently predicted unbound component priors into an assembled complex through an ultrafast diffusion bridge, operating entirely without multiple sequence alignments. By eliminating co-evolutionary dependencies, K-Fold accurately recovers non-coevolved interfaces in antibody-antigen and targeted-degrader complexes, as well as ligand-directed regulatory states in GPCRs and kinases. Furthermore, by coupling this diffusion bridge with diverse monomeric equilibrium ensembles, K-Fold provides a computationally tractable foundation for sampling complex multi-state landscapes without retraining. These results demonstrate that formulating molecular assembly as an explicit structural transition moves structure prediction beyond static geometries toward capturing biomolecular dynamics.

molecular biology↗

bioRxiv2026

Transient Splicing-Associated Isoform Remodeling Defines Early MAPKi Resistance Adaptation in Melanoma

A long-standing question in targeted therapy resistance is how tumor cells rapidly enter drug-tolerant states before stable resistance emerges. In BRAF-mutant melanoma, early adaptation to MAPK inhibitor (MAPKi) therapy is accompanied by profound phenotypic plasticity, yet the molecular mechanisms that enable this transition remain poorly understood. Here, we show that MAPKi adaptation is characterized by a transient transcript isoform remodeling that is largely invisible to conventional gene-expression analysis. We integrated and analyzed multiple different datasets, including bulk and single-cell RNA-seq datasets from melanoma cell lines, murine models, patient-derived xenografts and patient tumors, spanning no-treatment, on-treatment/minimal residual disease, and acquired-resistant states. We find that differential transcript usage peaks during the adaptive on-treatment phase and then is markedly attenuated when resistance was established. These dynamics were conserved across experimental systems and supported by independent exon-usage analyses. Notably, most genes undergoing differential transcript usage showed no corresponding change in total gene expression, and genes with isoform-level changes more strongly highlighted MAPKi resistance-associated pathways, including PI3K/AKT/mTOR and receptor tyrosine kinase signaling. Further, classifiers trained on cell-line transcript-usage features identified the adaptive state more accurately than gene-expression-based models in independent PDX and patient datasets. Interestingly, MAPKi-associated isoform changes recapitulated melanoblast-associated developmental splicing programs and were linked to a few key splicing regulators. Together, these findings identify an underappreciated regulatory layer connecting drug tolerance, developmental plasticity, and acquired resistance. Beyond the current context, our results suggest that comprehensively tracking isoform and splicing alterations may potentially advance our understanding of the emergence of treatment resistance in additional indications and treatment regimes.

cancer biology↗

bioRxiv2026

Transcriptome dissection of the Zeb2-associated network reveals a role for NRPS15 in zearalenone biosynthesis in Fusarium graminearum

Fungal secondary metabolites, including mycotoxins, play important roles in fungal biology and represent major concerns for food and feed safety. Zearalenone (ZEA) is an estrogenic polyketide mycotoxin produced by several Fusarium species, yet the genetic network governing its production remains poorly characterized. Here, we investigated genes associated with ZEA production in Fusarium graminearum using two complementary transcriptomic approaches. Time-resolved transcriptome analysis revealed transcriptional changes preceding detectable ZEA accumulation. Temporal clustering further revealed coordinated changes in primary metabolism, transcriptional regulation, carbohydrate utilization, and stress- and transport-associated pathways during ZEA-inducing cultivation. Genotype-based transcriptome analysis of ZEB2 deletion and overexpression strains subsequently identified Zeb2-responsive genes. Functional screening identified NRPS15, encoding a previously uncharacterized nonribosomal peptide synthetase-like protein adjacent to the ZEA biosynthetic cluster, as an important contributor to ZEA production. Deletion of NRPS15 markedly reduced ZEA production and expression of PKS4, PKS13, ZEB1, and ZEB2. NRPS15 expression did not restore ZEA production in the absence of Zeb2. NRPS15 orthologs were conserved in several ZEA-producing Fusarium species, and deletion of NRPS15 substantially reduced ZEA production in Fusarium culmorum. These findings expand the functional genetic context of ZEA production beyond the canonical four-gene biosynthetic cluster.

molecular biology↗

bioRxiv2026

Mitochondrial NAD+ import gates neuronal resilience to active SARM1

SARM1 is an inducible NADase that executes pathological axon degeneration, but how NAD+ destruction causes neuronal failure remains unclear. Because SARM1 is predicted to act outside the mitochondrial matrix, cytosolic NAD+ loss has been viewed as the critical lethal lesion. We show that SARM1 activation rapidly depletes cytosolic and mitochondrial NAD+ and collapses ATP in both compartments, revealing early, coupled metabolic failure. Enhancing mitochondrial NAD+ import through SLC25A51, the mitochondrial NAD+ transporter, preserves mitochondrial NAD+, maintains mitochondrial function, and delays axon degeneration and neuronal death despite ongoing SARM1 activity. Unexpectedly, SLC25A51 restores cytosolic NAD+-dependent metabolism, reversing a glycolytic block and increasing NAD+ synthesis during stress. Glycolytic inhibition attenuates protection, identifying cytosolic metabolic recovery as an important component of neuronal resilience. AAV-SLC25A51 improves function in a chronic SARM1-dependent neuropathy model. Thus, mitochondrial NAD+ import gates SARM1-driven degeneration by limiting coupled mitochondrial-cytosolic metabolic failure, presenting a downstream intervention point for axon preservation.

neuroscience↗

bioRxiv2026

Prime editing enables high-throughput identification and characterization of selective MutSβ inhibitors

DNA mismatch repair (MMR) maintains genome stability but paradoxically drives expansion of pathogenic repeats responsible for trinucleotide repeat disorders. Attenuating repeat instability through targeted inhibition of MMR is therefore a therapeutic strategy of interest; however, progress is constrained by a lack of robust methods for defining and isolating MMR subpathways. We exploited features of prime editing to profile the effects of core MMR genes on thousands of programmed DNA mispairs in cells and resolved distinct MMR processes, including a noncanonical, loop-directed mechanism. Guided by these results, we built a set of scalable reporters that detect loss of separable MMR processes and screened ~72,000 small molecules for activity against MutS{beta}-directed MMR. This effort identified a novel series of MutS{beta} inhibitors that reduced (CAG)n expansion in a cell model. Altogether, this study establishes several methods for systematically interrogating MMR, with demonstrated capacity to identify inhibitors of therapeutic value.

bioengineering↗

bioRxiv2026

Variant-Specific Rewiring of Lysosomal Positioning by STARD9 Produces Opposite Cholesterol Trafficking Outcomes

Residual cardiovascular risk persists despite lipid-lowering therapy, independent of LDL and Lp(a), reflecting cholesterol dysregulation. We identify STARD9, a lysosomal cholesterol-sensing kinesin, as a causal gene for divergent familial dyslipidemias and premature atherosclerosis. Its START domain binds cholesterol to govern lysosomal positioning. Rare variants segregate with autosomal dominant hypercholesterolemia and premature coronary disease, while common variants associate with reduced HDL and elevated triglycerides. Both converge on lysosomal cholesterol sequestration, ER depletion, mTORC1 SREBP2 activation, impaired autophagy, and NF-kB inflammation, yet diverge through opposite lysosomal positioning: the rare variant disperses lysosomes peripherally, phenocopying START domain loss across cholesterol sequestration, lysosomal positioning, and TFEB activation, triggering LDLR degradation and CASM/TFEB-dependent efflux that preserves HDL, whereas the common variant causes perinuclear retention that suppresses TFEB dependent lipid-handling transcripts, yielding hypertriglyceridemia and low HDL. In Stard9-knockout mice, enterocyte cholesterol sequestration drives SREBP2 dependent NPC1L1 upregulation and apical localization, increasing intestinal cholesterol absorption. These findings position lysosomal cholesterol trafficking as a targetable node in statin-refractory cardiovascular disease.

genetics↗

bioRxiv2026

Histotripsy-induced vascular remodeling in the tumor microenvironment enhances local and off-target intratumoral drug delivery

Solid tumors are known to be refractory to treatment due to a poorly perfused microenvironment that blocks immune cell infiltration and effective drug delivery. Histotripsy is a non-invasive focused ultrasound modality that generates cavitation microbubbles to mechanically disrupt solid tumors, resulting in decreased hypoxia and stimulation of tumor-targeted adaptive immune responses. In this study, we aim to investigate whether post-histotripsy hypoxia reduction and cytotoxic immune cell infiltration are associated with normalization of tumor vasculature and enhanced perfusion. Structural analysis of both treated and contralateral off-target tumors following partial histotripsy revealed significant vascular remodeling, enhanced vascular integrity, and decreased vascular leakage indicative of vascular normalization in multiple in vivo models of melanoma, hepatocellular carcinoma, and pancreatic adenocarcinoma. These effects appeared antigen-specific, as no remodeling was observed in contralateral off-target tumors of discordant pathology. Mechanistically, CXCR3+CD8+ T cells played a role in mediating these vascular changes, as genetic knockout of CD8 and pharmacological antagonism of CXCR3 impaired both remodeling and normalization. Additionally, significant changes in endothelial CXCR4 and Angiopoietin-1 and -2 expressions were observed following histotripsy, indicating their potential involvement in promoting vascular normalization and remodeling. Significant vascular remodeling was accompanied by enhanced tumor perfusion and intratumoral delivery of chemotherapeutic agents and therapeutic monoclonal antibodies. These findings suggest that histotripsy not only creates a more favorable microenvironment for immunotherapy, but also optimizes drug delivery into solid tumors. Combinatorial approaches using histotripsy with chemotherapy or immunotherapy may be a novel therapeutic strategy to overcome current limitations of conventional solid tumor therapies.

immunology↗

bioRxiv2026

Signal peptidase complexes set species-specific rules for signal peptide cleavage

Protein secretion is essential for cell function and begins when signal peptides direct nascent proteins into the secretory pathway, where they are cleaved by the signal peptidase complex (SPC). Although this pathway is highly conserved, signal peptides do not always function efficiently across species, and the molecular basis for this incompatibility remains unclear. Here, we show that human signal peptides with longer hydrophobic cores efficiently target and translocate nascent proteins in both yeast and human but are frequently left uncleaved in yeast, identifying signal peptide cleavage as a major barrier to cross-species compatibility. Molecular dynamics (MD) simulations reveal that, despite their highly conserved architectures, yeast and human SPCs generate distinct membrane-thinning profiles near the signal peptide-binding region. Our results support a model in which SPC-lipid interactions tune the local membrane environment to accommodate signal peptides of distinct hydrophobic core length, providing a biophysical mechanism for species-specific signal peptide recognition. These findings offer a mechanistic framework for understanding and potentially engineering protein secretion across diverse expression hosts.

cell biology↗

bioRxiv2026

Blockade of Kv3.1 by MK-801, a PCP-Derivative NMDA Receptor Inhibitor: Implications for Models of Schizophrenia

MK-801, a phencyclidine (PCP) derivative also known as dizocilpine, is a potent noncompetitive antagonist of the N-Methyl-D-aspartate receptor (NMDAr). The NMDAr plays a critical role in mediating excitatory synaptic transmission in the central nervous system (CNS) and is important in regulating synaptic plasticity, learning, and memory. MK-801 is known to induce schizophrenia-like phenotypes in animal models by blocking the NMDAr. Several studies have reported that the cognitive impairment associated with schizophrenia is linked to functional defects of ion channels in parvalbumin-positive GABAergic interneurons. Kv3.1 is a voltage-gated K+ (Kv) channel involved in the rapid repolarization of the action potential in neurons; it is richly expressed in parvalbumin-positive GABAergic interneurons and is associated with fast, repetitive spike generation. A decrease of Kv3.1 in the CNS has been reported to be associated with schizophrenia. In the present study, the effect of MK-801 on Kv3.1 was investigated using the whole-cell patch-clamp technique in Chinese hamster ovary (CHO) cells stably expressing Kv3.1. MK-801 caused a concentration-dependent inhibition of Kv3.1, with an IC50 of 10.81 M and a Hill coefficient of 0.89. The blocking potency was stronger at depolarized potentials, indicating a voltage-dependent block. MK-801 also produced a use-dependent block, inducing progressive inhibition with repeated stimulation at increased frequencies (1 Hz and 2 Hz), consistent with a delay in recovery from inactivation of Kv3.1 in the presence of MK-801. In addition, MK-801 induced a hyperpolarizing shift in the voltage dependence of the steady-state inactivation curve of Kv3.1. Taken together, these results indicate that MK-801 blocks Kv3.1 expressed in CHO cells in a concentration-, voltage-, and state-dependent manner. Given the importance of Kv3.1 in parvalbumin-positive, fast-spiking GABAergic interneurons, these findings suggest that MK-801 may alter the firing patterns of these inhibitory neurons, contributing to the onset and symptoms of schizophrenia.

pharmacology and toxicology↗

bioRxiv2026

Cyanophage CP12 Rewires Host Carbon Regulation through Interface Remodeling and Redox Buffering

Picocyanobacteria drive ocean carbon fixation, and cyanophages reshape host metabolism during infection. In cyanobacteria, the intrinsically disorder Calvin cycle protein 12 (CP12) assembles glyceraldehyde-3-phosphate dehydrogenase (GAP2) and phosphoribulokinase (PRK) into the inhibitory dark complex, yet how phage CP12 homologs modulate redox-sensitive partners remains unclear. Here, we examined viral CP12 across sequence, structure, and post-translational modification (PTM) complexities to resolve the mechanistic remodeling of dark-complex assembly and regulation. Protein-family analysis of CP12 homologs across diverse lineages showed that viral CP12 preserves interface-dominated positions while shifting partner-facing chemistry toward charged and geometry-modulating features. Matched molecular dynamics simulations (MD) of Prochlorococcus MED4 and cyanophage P-HM2 showed that phage CP12 preserves assembly while strengthening PRK-facing contacts and reducing GAP2-facing interface burial. MED4 redox proteomics identified coordinated cysteine oxidation across CP12 and GAP2 under light disturbance, guiding MD to explore thiol-PTM states. Conformational divergence increased with PTM load and localized mainly to GAP2 modifications. Thiol PTM at GAP2 imposed the largest CP12 binding-energy cost, which P-HM2 CP12 buffered, yielding smaller comparative binding-energy penalties than host CP12. These findings link sequence-driven chemistry to interface dynamics and redox PTM responsiveness to light, defining phage CP12 as a regulatory mimetic that may retune host carbon regulation during infection.

microbiology↗

bioRxiv2026

Microfluidic Devices for Imaging and Biochemistry Analysis of Microbes Under Mechanical Pressure

Growth-induced pressure arises when proliferating cell populations are confined within rigid microenvironments and is increasingly recognized as a determinant of microbial physiology in soils, biofilms, and host tissues. We present two complementary microfluidic devices that confine microbes within polydimethylsiloxane (PDMS) chambers and apply defined, optically read-out growth-induced pressure of up to 1.5 MPa. The first, the self-closing (SC) chip, confines cells in multiple small chambers accommodating hundreds of cells; it provides excellent nutrient supply and rapid medium exchange, supports single-cell imaging, and allows multiplexing of cellular or chemical conditions. The second, the pressure-recovery (PR) chip, confines cells in a single 5 cm-long chamber accommodating hundreds of thousands of cells; a scalpel-cut step recovers live cells from the channel within minutes for bulk biochemical assays. The two devices share a single two-layer soft-lithography fabrication process and a common brightfield wall-displacement pressure readout. Together, they enable single-cell imaging and bulk biochemical analysis under matched, defined pressure conditions. We illustrate the protocol with two representative validations: rapid {beta}-estradiol-induced transcription in the SC chip, in which nascent transcription foci appear within 5 min independently of the applied pressure, and the PR chip, which produces uniform growth-induced pressure along the entire 5 cm channel and recovers between 0.1x10^6 and 1x10^6 cells per device in a pressure-tunable manner.

biophysics↗

bioRxiv2026

How do Co-Folding Models Organize Structural Information?

Co-folding models accurately predict biomolecular complexes, but how their internal representations support joint structure prediction remains unclear. We analyze the Boltz-1 trunk by decomposing its pair representation into intra-chain and inter-chain blocks and applying ablation, layer-wise representation geometry, and activation patching. Our results show that structural information is distributed across three streams: the single representation carries token-internal atomic detail, the intra-chain pair representation encodes chain-level geometry, and the inter-chain pair representation controls relative molecular arrangement. Across the trunk depth, representations follow a Mix-Compress-Refine trajectory, while linear probes reveal that intra-chain distance content is largely supplied by MSA preconditioning and inter-chain distance content is constructed progressively across Pairformer layers at each recycle. Apo-holo patching further shows that ligand-induced backbone changes are not fully resolved before diffusion; instead, the diffusion module reconciles partially inconsistent intra- and inter-chain constraints. These findings provide a mechanistic view of how co-folding trunks organize structural information and suggest more partition-aware designs for future biomolecular structure prediction models.

bioinformatics↗

bioRxiv2026

AtlasFold: Protein structure prediction with metagenomic-scale language models

Protein language models (PLMs) trained on evolutionary sequences learn representations that encode protein structure, enabling direct structure prediction without multiple-sequence alignments (MSAs). Here we present the Atlas model family, an open and trainable system spanning protein language modeling, monomer folding, and protein-complex prediction. AtlasLM-3B is a 3B-scale language model trained with masked language modeling on approximately 1.56 billion sequences, including metagenomic data, and outperforms the similarly sized ESM2-3B in unsupervised contact prediction. Building on these representations, AtlasFold predicts all-atom protein structures and achieves state-of-the-art accuracy among PLM-based folding models. Fine-tuning AtlasFold for protein-complex prediction produces AtlasFold-Multimer (AtlasFold-M), whose antibody-antigen prediction performance is comparable to that of AlphaFold3 and ESMFold2. This protein-specific folding architecture enables fast, memory-efficient inference with AtlasFold and AtlasFold-M. By releasing the training code and data, stage checkpoints, and model weights under the MIT License, we provide a foundation for advancing PLM-based protein structure prediction.

bioinformatics↗