Hybrid blood biomarker intervals outperform pure personalization
Pure personalization overflags up to 68 percent of results without outcome links, but conditioning on group normals boosts precision for
q-bio
Quantitative Biology
sort pith recommended most recent
Five-phase JEPA co-training makes EHR rollouts converge
One backbone forecasts patient trajectories and beats multi-task risk baselines without fine-tuning
Benchmark trains models on noisy DEL screens and tests them on real Ki affinities
CA-DEL pairs screens for three carbonic anhydrase isoforms with ChEMBL validation data to check generalization from indirect signals.
· “CA-DEL: An Open Multi-Target, Multi-Modal Benchmark for Learning from DNA-Encoded Library Screens”
New edge-based μ-representation yields distance for orchard networks
It uniquely identifies binary semidirected networks and enables polynomial-time comparisons of evolutionary histories with reticulations.
· “A μ-distance for semidirected orchard phylogenetic networks”
Causal GRN methods beat correlations only in clean data
Isolated tests show dropout and confounders erase causal advantages over simple correlations in single-cell data
Atomic diffusion unifies protein sequence and structure design
One-stage model on atom coordinates and types outperforms staged methods and supports non-canonical amino acids.
· “A-CODE: Fully Atomic Protein Co-Design with Unified Multimodal Diffusion”
MCMC steering improves single-cell perturbation predictions
A Metropolis-Hastings sampler using masked conditionals avoids out-of-distribution artifacts when forecasting gene-knockout responses.
Lab incubation times flatten dengue epidemic peaks
Models using experimental mosquito EIP distributions produce lower but longer outbreaks than exponential assumptions, with little change in
Shot noise limits single ion channel voltage accuracy to 10 mV
Thermal fluctuations from discrete ions set a hard bound on 10 microsecond gating timescales close to observed performance.
· “Thermal fluctuations set fundamental limits on ion channel function”
Fine-tuned LLMs beat frontier models on cell perturbation prediction
Synthetic reasoning traces deliver state-of-the-art PerturbQA results, 87% accuracy on new cell types, and strong gains from 2% of the data.
900 worms of C
Resampled calcium traces and consensus connectomes let labs compare activity across animals and studies.
· “Homogenized textit{C. elegans} Neural Activity and Connectivity Data”
Higher bite rate can wipe out a zombie outbreak
On bounded-degree graphs, the outbreak probability can be continuous yet non-monotone in the infection rate.
Lorentzian family is unique invariant under Riccati transport
Reformulating dynamics on the circle shows the Cauchy law is the sole rotation-invariant measure, unifying exact mean-field reductions.
· “Geometric Origin of Exact Mean-Field Reductions: M{\"o}bius Symmetry and the Lorentzian Ansatz”
Degree and distance contact rules slow epidemic growth
Even mild dependencies shift spread from explosive to polynomial rates on networks with geometry.
Compartment-stratified features classify IBD at AUROC 0.96 under donor-aware splits
Benchmark on two cohorts shows CFN edges linear CLR in colon while compartment labels remove spurious dependency instability.
500,000 fully labeled outbreak simulations
Synthetic epidemics with true case counts and R_t give modelers what real surveillance cannot.
· “Reservoir: A Large-Scale Simulated Dataset for Training and Evaluating Epidemiological Models”
Fusion gates spot lost sensors
In an eight-strategy ECG-PPG fusion comparison, gates reweight sharply at dropout but barely track graded corruption.
Critical genome length for speciation now has a closed form
New formula tracks the genome size at which mating-threshold populations fragment, verified against full simulations.
· “What sets the critical genome length for sympatric speciation? A closed form and asymptotic theory”
Zebrafish schooling is driven by spinal bend sensors
A curvature-sensing spinal loop times tail beats to a neighbor's wake; removing it dissolves the school.
Frequent AI use does not mean AI literacy
On a new 18-item GenAI exam, teens who used chatbots most often scored lowest, while perceived usefulness tracked nothing.
Pairwise epistasis emerges in tokens
Two protein models can match on test accuracy yet store interaction structure at very different stages.
· “Beyond Tokens: Probing Higher-Order Epistasis in Learned Protein Representations”
Predict mixture toxicity curves from single-compound data alone
Finch reaches R² 0.90–0.92 on mixture toxicity curves using only mole-fraction-weighted embeddings.
· “Finch: Toxicity Dose Response Curve Prediction of Chemical Compounds and Mixtures”
Video AI misses how primate brains separate motion from appearance
Temporal integration alone fails; only predictive world models approach the brain's appearance-invariant motion coding.
· “Primate vision reveals a missing principle for robust dynamic AI”
CNN predicts RNA yield from DNA template at R = 0.94
Pooled screening of 100,000 random templates trains a model to prioritize candidates before any synthesis.
· “Optimizing RNA yield using deep neural networks coupled to massively parallel screening”
One identity splits selection's information at every level
Host, microbiome, and association each get their own testable share of evolutionary change.
Local 6-Å couplings recover a protein ensemble's global statistics
A sparse Potts model orders ten A2A ensembles along activation with no labels, and backmaps its samples to atoms.
· “PHASE: encoding global protein ensembles with local Hamiltonians and all-atom backmapping”
Nonlinear dwell-time curves reveal ion channel gating states
A fast log-linear test of adjacent dwell times tells a two-state channel from a multi-state one in under a second.
Multispectral models beat chlorophyll-only for every pigment
Six wavebands plus attenuation data improve estimates of eight pigments, especially alloxanthin and two fucoxanthin markers.
Duplication episodes now inferable on phylogenetic networks
A polynomial feasibility test plus exact search recovers whole-genome events under reticulation.
Phylogenetic split systems capped at n^{5/2}
Previous quartet-based bound was n^4; a graph-and-charging proof cuts it to n^{5/2}.
· “On the maximum size of 2-weakly compatible split systems”
Dendritic branches are necessary for powerful local learning
Without compartments, synapses lack local error signals for deep learning without separate phases.
Most enzyme combos fail: 4 of 1024 glycolysis subsets work
Maps all enzyme subsets of three central-carbon networks, flagging essential enzymes before wet-lab work.
· “Systematic pathway comparison on the powerset of rule-based biochemical systems”
One model fit uncovers hidden cell states in scRNA-seq
GmGM jointly infers cell and gene networks from PBMCs, finding monocyte subsets and proliferating cells that marker-based labels miss.
· “Uncovering Cellular Resolution in scRNAseq via Unbiased Cell and Gene Network Analysis”
Researchers measured brain waves while people tapped along with a visual metronome
Filling blank intervals in a visual metronome with a colored disk improved tapping stability and was associated with reduced visual…
· “Temporal filling-in reduces attentional fluctuations in sustained visual attention”
The paper simulates Q-learning agents playing prisoner's dilemma on networks and varies…
Q-learning agents in spatial prisoner's dilemma cooperate most when they observe roughly three to four neighbors, a non-monotonic effect…
· “Evolution of cooperation with Q-learning: how much information do we need?”
One coupling now handles all nearby parameter paths at once
Pairwise split-coupling variance at the path-count cost of common reaction paths, applied to three estimation settings.
Origin of life is a protocell–replicator merger
A multilevel-selection model locates the first great transition at a protocell–replicator union.
· “The first major transition: Origin of life from a multilevel selection perspective”
Tuning the Raman pump isolates dark excited states in carotenoids
Stepping the pump wavelength across transient absorptions turns it into a state-selective filter for femtosecond Raman spectra.
· “From Steady-State to Ultrafast: Resonance Raman Approaches for Biological Samples”
Heterogeneous SIR epidemics reduce to one pressure equation
Same individual-level process yields power-law, logarithmic, and multigroup incidence closures.
· “From Individual-Based Stochastic Epidemics to Heterogeneous SIR Equations”
Closed membranes turn isopotential after one ion channel opens
Theory: the compact-cell approximation emerges at late times, not from the start, on a microsecond scale.
· “Voltage dynamics of spherical membranes from single ion channel currents”
MRI entropy flags recurrence risk inside complete response
A baseline scan identifies the 25% of complete responders who still recur, changing how pCR is interpreted.
A 'virtual cell' platform aims to make phenotype prediction routine
DigiPhen couples automated multi-omics data with modular models so engineers can design first, test later.
· “DigiPhen: a new paradigm for building predictive models of biological systems”
One network assigns cryo-EM poses to proteins it has never seen
Trained on 3,330 structures, ARCHER matches per-dataset estimators and keeps conformational signal.
· “ARCHER: Amortized cross-specimen pose estimation for cryo-electron microscopy”
One threshold predicts when similarity-based cooperation succeeds
On finite phenotype spaces, cooperation requires sharp recognition, stable strategies, and an intermediate mutation rate.
· “Phenotypic Assortment and the Evolution of Cooperation in Finite Periodic Phenotype Spaces”
One noisy exposure trains both replay directions
A delay-coupled ring model replays forward and backward at quantized speeds, without symmetric plasticity.
· “Forward and reverse delay-driven hippocampal replay without symmetric plasticity”
Transformer designs m1Ψ mRNA 5' UTRs that hit multiple targets at once
New hybrid generator joins ribosome load, GC content, folding energy, and structure control in one design pass.
Motional energy turns gels into droplet clusters
Adding libration and dihedral terms to network Hamiltonian models flips low-T aggregation to droplets.
· “Motional Degrees of Freedom in Network Hamiltonian Models”
Bedtime glucose above 133 mg/dL flips the nocturnal-low switch
Explainable AI on 502 nights produces concrete thresholds for snacks and insulin adjustments in type 1 diabetes.
· “Explainable AI-based Decision Support for Nocturnal Hypoglycemia Prevention in Type 1 Diabetes”
A single abdominal electrode reads gastric dysfunction as well as an array
Single-electrode shear hit AUC 0.820, matching multielectrode phase arrays.
Predicted cell-state shifts lift drug-response prediction
A frozen single-cell perturbation model adds displacement features that beat patient-drug baselines on TCGA and PDX.
· “PerturbRx: Learning Treatment-Conditioned Latent Transitions for Patient Drug Response Prediction”
Metabolic networks are 54 percent sparser than thought
Random rewiring that preserves degree cannot reproduce metabolism's clustering or community structure.
· “Metabolic Network Properties: Comprehensive Analysis Across Domains”
Genomic data removes the ridge hiding household transmission rates
Adding transmission graphs to household SIR models separates within-household and external spread in simulated outbreaks.
Simulation-free flow matching exactly solves second-order transport
TracingFlow learns acceleration fields from sparse snapshots, so trajectories can cross and match every observed distribution.
· “TracingFlow: A Simulation-Free Trajectory Inference Framework Based on Second-Order Dynamics”
Multiclass gkm-SVM matches LS-GKM bit-for-bit
Train one model for enhancers, promoters and CTCF sites; get calibrated probabilities and 80x faster interpretation.
· “mLS-GKM: Efficient Multi-class Regulatory Sequence Classification with Gapped k-mer SVMs”
Simulation-trained network beats baselines on most epidemic forecasts
GENIE turns costly agent-based simulations into fast, location-aware probabilistic forecasts of hospital burden.
Fitness peak is a weighted mean of survival and reproduction optima
Because the optimum and stabilizing-selection strength are inseparable, new OU models of trait evolution emerge.
One ratio decides if a tissue stops growing or grows forever
Active viscoelastic model unifies determinate, indeterminate and proportionate growth in one framework.
· “Growth phases of an active tissue: determinate, indeterminate, and proportionate”
Criticality hides from pairwise abundance correlations
Even near instability, soft modes stay invisible unless their forcing beats the background; correlations cap near 0.6.
Dengue risk in Europe may rise 300-fold by 2100
Warming and air travel shift Europe from imported outbreaks to environment-driven dengue transmission.
· “Climate change and human mobility will shape dengue emergence risk in Europe”
Reputation alone flips the prisoner's dilemma to full cooperation
A trust-weighted blend of own and neighbors' payoffs creates a discontinuous, bistable cooperative phase in a spatial Q-learning model.
· “Emergence of cooperation: A reputation-modulated reinforcement learning”
One flow-matching model beats vessel-specific 3D segmenters
The same model segments portal, cerebral, and coronary vessels with better overlap and continuity in three steps.
· “Flow Matching Meets 3D Curvilinear Structure Segmentation in Medical Imaging”
Tool-selection accuracy jumps from 23% to 94% with a research harness
Governed agentic neuroimaging: sealed plans, audited claims, and verified evidence grounding up nearly fivefold.
Seasonal timing decides whether an environmental pathogen dies out
A stochastic model of snake fungal disease shows extinction risk swings with introduction route and season.
· “A stochastic dose-response framework for environmentally persistent pathogens”
Imaging score places patients on the Alzheimer's continuum
A Bayesian model turns repeated diffusion MRI into a continuous disease score with uncertainty, beating discrete diagnosis.
Fine-tuned Transformer hits 99.9% HCV subtype accuracy
Balanced, evenly fragmented training data lift HCV genotype precision from 48 percent to near-perfect.
· “Hepatitis C Virus Genotyping with a Transformer Neural Network”
Collagen molecules strain ahead of the fibril's D-band
A two-state transition of 6–10 kT marks local untwisting of the triple helix before rupture.
Hi-C map codec keeps local folds
Adapting a photo codec to contact maps preserves TADs and stripes but not A/B compartments; the cause is fixed-size tiling.
Bayesian Last Layer tops 70% of protein benchmarks
Partial-Bayesian models plus supervised PCA gave strongest accuracy and calibration across 27 protein datasets
· “Multitask Bayesian Neural Networks for Multiparameter Protein Engineering”
Grid cells cut spatial aliasing by up to 99 percent
A periodic internal spatial code disambiguates visually identical locations that boundary cues alone confuse.
· “The Role of Grid Cells in Reducing Spatial Aliasing in Hippocampal Place Representations”
Two prototypes explain all degenerate 2D zero-one networks
A row-pattern check on the stoichiometric matrix decides degeneracy; steady states are binomial.
· “Characterization of Minimal Degenerate Zero-One Reaction Networks”
Q-networks turn first-person experience into category theory
Counting connections and loops in a network of possible experiences yields structural markers for unity, self, and time.
AF detection holds AUC above 0.98 across seven ECG datasets
Training ECG reconstruction and AF detection on one shared codebook keeps accuracy high with as few as one lead.
Adaptive therapy delays progression but raises metastasis risk
Fitted to 73 prostate-cancer patients, the model maps the trade-off and shows why schedules don't transfer.
· “Adaptive therapy under parametric, structural, and measurement uncertainty”
Fine-tuning, not few-shot prompting, makes LLMs find antimalarial hits
Fine-tuned TxGemma-9B reaches 0.731 ROC-AUC and LlaSMol finds 5x actives at top 1%; Gemini and o3 stay near random.
Two entropy rules build working proteins
A layer-restricted Boltzmann machine constructs functional chorismate mutases up to 53% divergent from nature.
Codon graphs identify SARS-CoV-2 variants with 98.75% accuracy
Multi-scale codon graphs and spectral features separate four lineages without alignment, in under a second per genome.
Memory and effector CAR T-cells fail by separate routes
Memory cells watch residual disease; effectors clear niche blasts; initial burden drives CD19-negative relapse.
· “Mathematical modelling of immune persistence and relapse pathways in CAR T-cell therapy for B-ALL”
Cryo-EM images yield protein backbones in one step
A rotation-group gradient flow deforms a template until its projections match noisy images—no density map needed.
Patient-separated EHG model predicts preterm birth at chance
Record-separated evaluation puts AUROC at 0.493; resampling must match the clinical unit of prediction.
Interference between gene-loop paths predicts their abundance
A mutual-information split reproduces why C1 and I1 loops dominate E. coli and yeast networks.
· “An information-theoretic perspective on feed-forward loop abundances in transcriptional networks”
