Biology subjects

Gao, X.

Publications and source records attributed to Gao, X..

At least 19 recordsLinked to original sources

bioRxiv2026

Reinforcement learning with evolutionary priors unlocks multi-site enzyme optimization from sparse experimental feedback

Industrial enzymes underpin sustainable manufacturing, yet natural proteins rarely combine the activity, robustness and application compatibility required for practical use. Engineering such properties requires identifying beneficial combinations of interacting mutations within vast sequence spaces that are too large to search exhaustively and too sparsely sampled to model reliably. Here we introduce AlphaVariant, a reinforcement-learning (RL) framework that couples a generative evolutionary prior with iteratively updated low-N fitness models to enable feedback-driven search of large combinatorial protein sequence spaces. Across six experimentally measured protein fitness landscapes, AlphaVariant ranked first on the three large, non-enumerable multi-site landscapes. In two successive rounds of savinase activity optimization, AlphaVariant-designed libraries yielded 75.4% and 83.8% activity-improved variants among experimentally characterized candidates even as the accessible mutation space expanded from 25 to 67 positions. The search identified beneficial mutation combinations involving up to 17 structurally dispersed sites and produced a variant with 7.1-fold higher activity than wild type after only two design-test cycles. We then extended AlphaVariant to jointly optimize activity, thermostability and a patent-informed constraint for intellectual-property freedom. This multi-objective search yielded variants that outperformed the commercial protease Uno 100 L in both biochemical assays and application-relevant detergent washing tests. These results demonstrate that generative RL can efficiently navigate vast protein fitness landscapes, progressively incorporate multiple engineering objectives and convert sparse experimental feedback into enzyme variants with industrially relevant performance, effectively resolving the limits of biocatalyst engineering.

bioinformatics↗

bioRxiv2026

Reconstruction of a rare {triangleup}10 desaturation pathway enables production of yellow peach moth pheromone precursors in Camelina sativa

Sex pheromones are environmentally friendly alternatives to traditional insecticides for pest control. The high synthesis cost, however, has restricted large-scale application. Plant-based biofactories provide a promising, sustainable route for cost-effective production that could facilitate the broader use of pheromone-based pest control. Mating disruption of the yellow peach moth, Conogethes punctiferalis--a serious pest of fruit trees and field crops in East Asia--can be achieved using (E)-10-hexadecenal and (Z)-10-hexadecenal, two components of its sex pheromone. To enable rational reconstruction of the biosynthetic pathway towards these unusual fatty acid derivatives in heterologous production systems, we first elucidated their biosynthesis by combining in vivo labeling experiments, transcriptomics and functional characterization of candidate desaturase and reductase genes using Nicotiana benthamiana and yeast expression systems. We identified a rare {Delta}10 fatty acyl desaturase that catalyzes the production of both (E)-10- and (Z)-10-hexadecenoic acids, along with a downstream fatty acyl reductase with higher activity on the E isomer. We further introduced this {Delta}10 desaturase together with an acyl-ACP thioesterase-1 into Camelina sativa to redirect seed lipid metabolism towards the accumulation of the pheromone precursor fatty acids. After four generations of selection, we obtained stably inherited Camelina lines accumulating up to 4.7% of target fatty acids in total seed lipids, with individual seeds containing up to 8.3 g, without significant effects on plant growth or seed viability. These findings demonstrate how elucidating moth pheromone biosynthetic pathways can support the development of sustainable biotechnological production systems for pheromone-based pest management.

synthetic biology↗

bioRxiv2026

BioMINTREL expands biological problem solving through cross-disciplinary representations

Biological research agents often rely on familiar analytical methods, limiting the questions explored in existing data. We introduce BioMINTREL, an AI agent that finds methods from other disciplines and connects them to biological measurements. It can draw on relationships among datasets and complementary ideas from several methods to develop new questions and analyses. We evaluate it on BioMDC-Bench, which assesses 228 possible applications of computational methods to biological tasks. Compared with GPT-5.4 mini and Qwen3.6-35B, BioMINTREL improves F1 agreement with human experts by up to 45.3 percentage points and increases coverage of applications supported by studies published in subsequent years by up to 58.8 percentage points. Four case studies on real data go beyond previous analyses to reveal dose-specific variant effects, constraints on spatial cell communication and routes through perturbation-response states, and identify the sampling needed to test whether rare high-state cells predict later lineage size. These results highlight BioMINTREL's potential to expand questions in existing data and suggest directions for subsequent research.

bioinformatics↗

bioRxiv2026

5-HT1A Antagonist WAY-100635 Prevents Chronic Neurodegeneration and Immune Activation After TBI

Traumatic brain injury (TBI) initiates secondary neurodegeneration that compromises interconnected neural circuits and contributes to persistent visual and neurobehavioral impairment. Pharmacological strategies that preserve vulnerable neurons and synapses while limiting chronic neuroinflammation remain an unmet therapeutic need. We previously identified the serotonin 1A receptor (5-HT1A) antagonist WAY-100635 through screening human stem cell-derived retinal ganglion cells (RGCs) and demonstrated that it promotes neuroprotection through cyclic adenosine monophosphate-dependent mitochondrial biogenesis. Here, we investigated whether post-injury treatment extends this neuroprotective efficacy to cerebral and visual circuitry following TBI. Male C57BL/6J mice aged 2-3 months underwent unilateral controlled cortical impact (CCI) involving the left visual cortex. Mice were randomly assigned to sham, untreated injury or treated injury groups, with 3-10 sham controls and 5-8 mice per injury group across experiments. WAY-100635 was administered intraperitoneally at 5 mg/kg daily, beginning 1 hour after injury and continuing for 6 weeks. Therapeutic efficacy was assessed using motor and psychological-health-relevant risk-taking/behavioral disinhibition by elevated-plus-maze assessment, visual electrophysiology, optokinetic responses, diffusion magnetic resonance imaging, and retinal and cortical histopathology. Unilateral cortical injury produced bilateral RGC dysfunction and loss, with greater degeneration in the eye contralateral to the injured hemisphere, consistent with predominantly crossed murine visual projections. WAY-100635 significantly preserved RGC survival and function, improved visual acuity and contrast sensitivity, and preserved visually evoked cortical responses to contralateral eye stimulation. Diffusion MRI revealed bilateral visual pathway abnormalities, with treatment selectively normalizing fractional anisotropy and radial kurtosis in the optic tract ipsilateral to cortical injury. In perilesional cortex, treatment preserved neuronal cell bodies and synapses detected by synaptophysin immunoreactivity while markedly reducing microglial/macrophage activation and reactive astrogliosis. These effects were accompanied by improved early motor performance and attenuation of injury-associated increases in open-arm exploration in the elevated plus maze. Cortical cavity volume was not significantly reduced, supporting preferential protection of vulnerable circuitry surrounding and connected to the primary lesion. These findings identify post-injury WAY-100635 treatment as a novel therapeutic intervention that combines neuronal and synaptic preservation with reduced chronic glial activation and improved functional outcomes. They support further translational development of serotonin 1A receptor antagonism to limit secondary neurodegeneration and the visual and neurobehavioral consequences of TBI.

neuroscience↗

bioRxiv2026

Multi-omic profiling of Plasmodium knowlesi blood stages identifies host adaptation barriers

The zoonotic parasite Plasmodium knowlesi has emerged as a widespread cause of human malaria in Southeast Asia. We assembled a high-resolution transcriptome of zoonotic P. knowlesi across its intraerythrocytic developmental cycle and combined transcriptomics, quantitative proteomics and thermal stability measurements to compare parasites cultured in macaque and human red blood cells. We found that P. knowlesi accelerates through the ring stage to settle at the same developmental pace as P. falciparum during the trophozoite and schizont stages. Subsequent multi-omic analysis revealed broad transcriptional and proteomic adaptations of P. knowlesi to the human host cell environment. Parasites cultured in human cells expressed glutathione synthetase earlier and displayed changes in upstream processes that include NADPH production by the pentose phosphate pathway and glutamine synthesis. These changes were accompanied by further differences in the expression of ribosomal subunits and mitochondrial respiratory complexes. Together, our findings offer a systems-level view of non-genetic adaptations of P. knowlesi to the human red blood cell environment.

microbiology↗

bioRxiv2026

AbRefine: Framework-Conditioned Refinement for Structure-Guided Antibody CDR Design

Structure-guided antibody CDR design constrains residue selection using backbone geometry, but structural evidence alone may leave multiple amino acids plausible. We investigate whether the fixed antibody framework provides complementary sequence information for resolving this ambiguity. We introduce AbRefine, an information-aware fusion framework that combines backbone-conditioned structural predictions with framework-conditioned sequence distributions derived from a pretrained antibody language model with all designed CDRs masked. Its Information-Aware Fusion (IAF) module learns residue-specific fusion weights from the uncertainty, disagreement, and information content of the two predictive distributions, while keeping both pretrained experts frozen. Controlled interventions show that the signal is target specific: matched frameworks outperform shuffled frameworks by 7.37 percentage points, and the largest gain occurs at positions with high structural uncertainty and informative framework evidence (+5.77$ points). Compared with static fusion, IAF further improves AbMPNN recovery by +1.53 points (95\% CI +0.93$ to +2.14$) and significantly reduces native negative log-likelihood for both AbMPNN and AntiFold, without detectable degradation in fold-back structural compatibility. These results show that framework context provides information beyond backbone geometry and that learning its residue-specific utility enables effective adaptive fusion for antibody CDR design.

bioinformatics↗

bioRxiv2026

Stem cell factor acts locally in the bone marrow to promote stem and progenitor cell maintenance

Stem cell factor (SCF) from Leptin Receptor-expressing (LepR+) mesenchymal stromal cells and endothelial cells is required for hematopoietic stem and progenitor cell maintenance in bone marrow. Consistent with this, Scf is highly expressed by LepR+ cells and at lower levels by endothelial cells in bone marrow. However, endothelial cells outside the bone marrow produce circulating SCF, raising the question of whether the SCF that promotes stem and progenitor cell maintenance is produced inside or outside bone marrow. We found that Scf deletion in spleen and liver endothelial cells reduces SCF levels in the blood without affecting stem or progenitor numbers in the bone marrow. Moreover, deletion of Scf from mesenchymal cells in the long bones, but not vertebrae, of Prx1Cre; Scf fl/fl mice depletes stem and progenitor cells in long bones but not vertebrae. SCF thus acts locally within the bones in which it is made to promote stem and progenitor cell maintenance.

cell biology↗

bioRxiv2026

Ultra-fast and scalable high-resolution full-length single-cell RNA sequencing using CHART-seq

Plate-based full-length single-cell RNA sequencing resolves transcript structure details but remains difficult to scale because each cell usually requires a separate library. Here we developed CHART-seq (Combinatorial Heteroduplex Assay via Recombinant Tn5), which uses orthogonally indexed Tn5 complexes to tagment RNA/cDNA heteroduplexes and permits early sample pooling. The workflow processed up to 96 cells per library, was compatible with 384-well expansion, and completed library preparation within 3 h at a reagent cost below US$1 per cell. At matched sequencing depth, CHART-seq detected more genes and annotated isoforms than Smart-seq2, Smart-seq3, Flash-seq and SHERRY2, while retaining broad genebody coverage and reproducible expression estimates. In the CHART-seq results of vascular smooth muscle cells, TGF-{beta}1 pretreatment before PDGF-BB exposure partly restored contractile features, suppressed a PDGF-associated inflammatory programme, and induced a distinct metabolic-matrix response with coordinated transcript-usage changes. These biological findings remain exploratory because independent biological replicates were unavailable. CHART-seq provides a rapid, scalable route to full-length single-cell transcript profiling with gene-programme and candidate isoform resolution.

genomics↗

bioRxiv2026

Speciesformer learns conserved cellular states for cross-species generative virtual cell modeling

Cells across species are governed by evolutionarily conserved biological programs, yet their molecular states and responses are reshaped by species, tissue and cellular context. A central challenge for virtual cell modeling is therefore to learn cellular states and state transitions that separate transferable biological principles from context-specific variation. Existing single-cell foundation models have advanced cellular representation learning, but most remain focused on single-species analysis, discriminative tasks or specialized forms of generation. Here we present Speciesformer, a cross-species generative single-cell foundation model that integrates evolutionary representation learning with virtual cell-state generation. Speciesformer is pretrained on SpeciesCorpus, comprising 131 million cells from 11 species, 154 tissues, and more than 923 cell types, and maps species-specific genes into a shared evolution-informed gene space. Its encoder learns transferable cell and gene representations that support biological representation probing and cross-species knowledge transfer, providing a common state space for resolving conserved and context-dependent cellular programs. Building on this shared representation, Speciesformer uses a unified generative architecture to model cellular states and state transitions under semantic and interventional conditions, enabling bidirectional generation between transcriptomic states and biological text descriptions, as well as prediction of post-perturbation transcriptomes from initial cell states and intervention descriptions, including in previously unobserved cellular contexts. By unifying evolutionary variation, biological semantics and conditional state transitions, Speciesformer extends cross-species foundation modeling toward a generative virtual cell framework for representing, describing and predicting cellular states across biological contexts.

systems biology↗

bioRxiv2026

Adaptive introgression in Tibetans spans focal and distributed architectures

The Denisovan-derived EPAS1 haplotype is a paradigmatic example of adaptive introgression, but whether archaic ancestry contributed beyond such large-effect loci remains unclear. We analyzed genomes from 1,039 Tibetans, Han Chinese, archaic hominins and ancient Tibetan Plateau individuals. Denisovan-derived adaptive variation was highly localized around EPAS1, whereas Neanderthal-derived variation was distributed across the genome; temporal and elevational patterns reinforced this contrast. Despite individual-level heterogeneity, Neanderthal-introgressed gene repertoires repeatedly converged on immune, metabolic and signal-transduction pathways. A Tibetan-enriched Neanderthal PPARGC1A haplotype was associated with lower fasting glucose and showed concordant support from an independent East Asian type 2 diabetes GWAS. These findings reveal that adaptive introgression can contribute to complex human adaptation through functionally convergent combinations of archaic variants, rather than only through shared large-effect haplotypes.

evolutionary biology↗

bioRxiv2026

Decoupling Transcription from Cell Fate Preserves Germ Cell Totipotency

Only germ cells can give rise to a totipotent embryo, yet a conserved transcriptional program for germ cell identity remains elusive. Profiling Drosophila primordial germ cells (PGCs) across embryogenesis, we find that nascent PGCs are not silent, as previously thought, but instead transcribe the same totipotency program as the early embryo, reflecting an open chromatin state. This chromatin state, established by the pioneer factor Zelda, persists in PGCs throughout development. The associated transcriptional program normally drives somatic differentiation, but in PGCs the conserved RNA-regulator Nanos represses its translation, leaving germline chromatin uncommitted. Cells that fail to establish the totipotent state die before reaching the gonad. Thus, germ cells preserve developmental potential not by restricting transcription, but by preventing transcripts from determining cell fate.

developmental biology↗

bioRxiv2026

Large Language Models Predict Human Social Behavior via Interpretable Mechanisms

The development of large language models (LLMs) offers promising opportunities for predicting human behavior across diverse contexts. However, most prior work has emphasized behavioral imitation, with limited attention to transparent or interpretable models of the cognitive mechanisms underlying human decisions. In this study, we introduce MindEvolve, an autonomous workflow designed to predict behavior in social interactions by generating interpretable symbolic models of cognition. We systematically evaluate the modeling capabilities of multiple LLMs across a battery of socioeconomic games covering four core domains of social cognition: economic preferences, social preferences, social reasoning (theory of mind), and recursive planning. The symbolic models proposed by the LLMs are assessed by expert human evaluators for interpretability and theoretical coherence. Our results show that while most LLMs robustly capture economic and social preferences in relatively simple strategic settings, and in some cases generate novel models that integrate broader knowledge than those proposed by human experts, their capacity to model more complex psychological processes remains limited. However, a subset of state-of-the-art models demonstrates promising performance in capturing higher-order reasoning processes such as theory of mind and recursive planning. Together, these findings highlight the emerging potential of LLMs not only to imitate behavior, but to generate interpretable, mechanistic accounts of human social decision-making. Our results provide a roadmap for advancing LLM-based cognitive modeling toward human-expert-level theory construction.

neuroscience↗

bioRxiv2026

Single-atom inhibition of oncogenic drivers through cysteine coordination

Small-molecule inhibitors rely on molecular recognition within suitable binding pockets, leaving many disease-associated proteins difficult to target. Here, we introduce the concept of a single-atom inhibitor in which gold (Au) engages critical cysteine residues of oncogenic drivers to suppress their activity. We used an AI-assisted few-shot learning approach to identify EGFR-targeting peptides for in vivo Au delivery and showed that the lead candidate, 10714, promoted Au accumulation in EGFR-expressing cells and tumors. In vivo, Au exploited its intrinsic affinity for cysteine to inhibit two structurally distinct oncogenic proteins, engaging Cys797 in EGFR T790M and the mutation-derived Cys12 in KRAS G12C adjacent to their respective nucleotide-binding pockets. Structural and computational analyses supported stabilization of inactive nucleotide-bound states, while mutation of these cysteine residues abrogated Au-mediated inhibition. 10714-Au consequently suppressed oncogenic signaling, reduced non-small-cell lung cancer cell viability, and inhibited tumor growth in EGFR- and KRAS-mutant xenograft models and patient-derived organoids. These findings establish proof of principle for single-atom inhibition across structurally distinct oncogenic drivers and suggest that localized atomic coordination could provide an alternative mode of target engagement to conventional pocket-dependent inhibition.

molecular biology↗

bioRxiv2026

MolJam: A Multidimensional Framework for Assessing Molecular Dataset Quality and Its Impact on Machine Learning

High-quality molecular datasets are essential for reliable machine learning in cheminformatics and bioinformatics, yet dataset quality is rarely assessed systematically and its relationship with downstream model performance remains poorly understood. Here, we present MolJam, an open-source framework for quantitative assessment of molecular dataset quality across five dimensions-structural integrity, data quality, experimental information quality, chemical space coverage, and data distribution-using 12 standardized metrics. Application of MolJam to 11 MoleculeNet and eight ChEMBL-derived datasets revealed widespread and heterogeneous quality issues, including undefined stereochemistry in up to 70.72% of molecules, inconsistent molecular representations, and contradictory labels. We next asked whether improving these quality metrics necessarily improves machine learning performance. Refinement of the ESOL and Lipophilicity datasets increased their MolJam quality scores but produced mixed effects on predictive performance, suggesting a competing influence of reduced dataset size. Controlled ablation experiments further demonstrated that both dataset quality and data quantity contribute to model performance and, notably, that retaining molecules with incomplete stereochemical information can outperform their removal when the resulting gain in data quantity offsets the quality penalty. Thus, molecular dataset curation cannot be reduced to maximizing data cleanliness alone but requires balancing multiple dimensions of data quality against information loss. MolJam provides a standardized framework for diagnosing molecular dataset limitations, comparing benchmark quality, and quantitatively evaluating how data curation decisions influence downstream machine learning.

bioinformatics↗

bioRxiv2026

Within-Subject Optogenetic Model Reveals Spatiotemporal Cortical Reorganization in Artificial Vision

Visual prostheses can elicit simple percepts such as letter forms through dynamically patterned stimulation, yet naturalistic dynamic vision remains out of reach despite hardware fully capable of high-resolution temporal control. This persistent gap suggests the bottleneck lies in how artificial input is represented by the cortex itself. A fundamental question remains unresolved: when artificial vision bypasses natural retinal encoding, what is lost in cortical representation? We address this question in this paper using a retinal optogenetic mouse model that isolates the consequence of bypassing retinal encoding while preserving downstream pathways. Using within-subject V1 cortex electrophysiology and CNN-based decoding under identical dynamic stimuli, we reveal dissociable, dimension-specific gaps. Temporally, artificial vision bypassed the frequency-dependent compression imposed by natural vision, extending stimulus-locked entrainment bandwidth. Spatially, direction-selective representations underwent systematic remapping, with disrupted phase organization and reduced encoding regularity. These findings show that bypassing natural retinal encoding reorganizes cortical representation along dissociable spatiotemporal dimensions, establishing a within-subject framework that renders such structural divergences easy to locate and quantifiable.

bioengineering↗

bioRxiv2026

CryoForge: A Self-Correcting Agent for Cryo-EM Model Building That Learns When to Act and When to Stop

Automated atomic model building has accelerated cryo-EM structure determination, but different builders leave distinct residual error profiles requiring expert inspection. The post-building challenge is to decide which local interpretations are sufficiently supported by experimental evidence to be retained, corrected or rejected. Here we introduce CryoForge, an evidence-gated post-builder agent that separates repair proposal from repair acceptance. Rule and learning-based components identify candidate regions and prioritize legal actions, whereas an independent evidence gate evaluates each edit using map and half-map support, stereochemistry, connectivity and local structural context. Supported edits are retained; unsupported or conflicting modifications are rejected, rolled back, stopped or escalated for expert review. Across a resolution-stratified benchmark, 84.9% of 26,153 released trajectories yielded standard validated improvements and 3.7% yielded low-confidence partial improvements, with no quality-degrading edit retained in the final promoted models. Relative to rule-only control, learned prioritization reduced non-improving candidates and harmful actions while preserving global structural stability. External evaluations using an alternative initializer, same-team automated/manual-assisted challenge submissions and three recently released complex assemblies showed that CryoForge adapts to distinct residual error phenotypes and performs bounded, evidence-supported correction without uncontrolled remodeling. CryoForge provides a builder-independent, scalable and auditable correction layer between automated model generation and expert structural interpretation.

bioinformatics↗

bioRxiv2026

DQHTFI: Dynamic-Query Hypergraph Transformer for Fine-Grained Drug-Target Interaction and Affinity Prediction

Drug-target interaction prediction and binding affinity prediction are two key tasks in drug discovery and drug repurposing. Although deep learning methods have made significant progress, existing models typically rely on global representations of drugs and proteins, making it difficult to adequately model fine-grained interactions between their local units. Fixed multimodal fusion strategies also struggle to dynamically adjust the contributions of different modalities for different drug-target combinations. To address these issues, we propose DQHTFI, a fine-grained interaction prediction framework for drug-target interaction classification and binding affinity regression. DQHTFI employs BRICS fragments and Pfam functional domains as the basic interaction units and jointly learns semantic and structural representations. We design a dynamic-query hypergraph Transformer framework in which hyperedges are constructed among the multimodal features of fragment-domain pairs. Dynamic queries are generated from the cross-conditioned features of fragment-domain pairs to adaptively adjust the contribution of each modality, thereby modeling higher-order interactions between local units. Our proposed model achieves competitive results on multiple benchmark datasets.

bioinformatics↗

bioRxiv2026

siDiff: De Novo siRNA Design via Efficacy-Guided Discrete Masked Diffusion

Target-conditioned de novo siRNA design requires capturing both target mRNA binding context and internal siRNA sequence-activity rules. Traditional computational pipelines predominantly rely on discriminative prediction models that rank pre-filtered candidate pools. However, these approaches struggle with generalization on novel target genes and often overlook potent candidates due to dataset size constraints and motif overfitting. To reconcile generation precision with sequence diversity, we propose siDiff, an efficacy-guided discrete diffusion framework for target-conditioned de novo siRNA design. siDiff pairs a discrete-masked diffusion transformer--which models the underlying sequence distribution over functional duplexes--with a mask-robust efficacy guidance model. During inference, we introduce a biology-aware, three-stage sampling mechanism that performs structural candidate filtering, dynamic unmasking guidance, and cluster-aware redundancy mitigation. This dual mechanism enables the diffusion process to explore broad sequence spaces while the efficacy model prevents distributional shift toward non-functional candidates. Extensive experiments across four datasets, including the public Takayuki benchmark and three curated patent datasets, demonstrate that siDiff significantly outperforms state-of-the-art discriminative baselines and discrete diffusion models, achieving relative hit-rate improvements of over 30% and demonstrating superior generalization on out-of-distribution gene targets. The source code and related materials are available at https://github.com/cybericha/siDiff.

bioinformatics↗