AI x Bio Discovery

Benchmarking DNA Foundation Models: Biological Blind Spots inEvo2 Variant-Effect Prediction

Figure 1

Figure 2

Figure 4

Benchmarking DNA Foundation Models: Biological Blind Spots inEvo2 Variant-Effect Prediction [new]
...reveal systematic misses in short-range biosignals (e.g., codon usage) & sensitivity to neutral ctx, impacting clinical readiness.

Decoding epitope immunodominance in HIV Env using cryoEM and machine learning

Figure 1

Figure 2

Figure 3

Decoding epitope immunodominance in HIV Env using cryoEM and machine learning [new]
...reveals structural determinants through Env-Ab complex analysis, informing a predictive model for immune bias and guiding vaccine antigen design.

Cross-Species Transfer Learning for Electrophysiology-to-Transcriptomics Mapping in Cortical GABAergic Interneurons

Table 1

Table 3

Figure 1

Cross-Species Transfer Learning for Electrophysiology-to-Transcriptomics Mapping in Cortical GABAergic Interneurons [new]
expands Ephys-to-Tx mapping framework using a seq model with attention to improve cross-species (mouse-human) prediction of subclass ID.

Constrained Diffusion as a Paradigm for Evolution

Figure 1

Figure 2

Figure 3

Constrained Diffusion as a Paradigm for Evolution [new]
...models evolution as a diffusion process on discrete genotype space restricted by viability constraints, invertible to recover the evolving viable manifold.

Continuous Diffusion Transformers for Designing Synthetic Regulatory Elements

Figure 1

Figure 2

Figure 3

Continuous Diffusion Transformers for Designing Synthetic Regulatory Elements [new]
...generate regulatory DNA via a transformer denoiser with a 2D CNN encoder, enhanced by DDPO finetuning for genuine regulatory signal.

Equivariant Asynchronous Diffusion: An Adaptive Denoising Schedule for Accelerated Molecular Conformation Generation

Figure 1

Figure 2

Table None

Equivariant Asynchronous Diffusion: An Adaptive Denoising Schedule for Accelerated Molecular Conformation Generation [new]
...uses an adaptive async denoising schedule to better capture mol hierarchy while maintaining molecule-level horizon for 3D generation.

Omics Data Discovery Agents

Figure 1

Figure 2

Figure 3

Omics Data Discovery Agents [new]
transform unstructured omics literature into searchable research objects via LLM agents, enabling automated data reuse and cross-study comparisons.

Discovery of a Hematopoietic Manifold in scGPT Yields a Method for Extracting Performant Algorithms from Biological Foundation Model Internals

Figure 1

Figure 2

Figure 3

Discovery of a Hematopoietic Manifold in scGPT Yields a Method for Extracting Performant Algorithms from Biological Foundation Model Internals [new]
Extracts compact hematopoietic algorithm from int. geometry, reveals lineage gene programs, and generalizes.

How to make the most of your masked language model for protein engineering

Figure 1

Figure 4

Figure 2

How to make the most of your masked language model for protein engineering [new]
...by using stochastic beam search for effective, flexible sampling of optimized biological properties via efficient 1-edit neighborhood evaluation.

JEDI: Jointly Embedded Inference of Neural Dynamics

Figure 2

Figure 4

Figure 3

JEDI: Jointly Embedded Inference of Neural Dynamics [new]
across tasks by learning a shared embedding space for recurrent weights, unifying complex neural dynamics from recordings.

SNPgen: Phenotype-Supervised Genotype Representation and Synthetic Data Generation via Latent Diffusion

Figure 2

Figure 3

Table 1

SNPgen: Phenotype-Supervised Genotype Representation and Synthetic Data Generation via Latent Diffusion [new]
...generates privacy-pres., pheno-aligned synth. genotypes using GWAS-guid. SNP selection, VAE compression, and latent diffusion with disease labels.

scDEcrypter: Uncertainty-aware differential expression analysis for viral infection in scRNA-seq

Figure 1

Figure 2

Figure 3

scDEcrypter: Uncertainty-aware differential expression analysis for viral infection in scRNA-seq [new]
A penalized 2-way mixture model with partial labels and data-splitting to overcome sparse viral reads and under-labeled cells.

The prevalence of protein misfolding as a mechanism for hereditary deafness

Figure 1

Figure 2

Figure 3

The prevalence of protein misfolding as a mechanism for hereditary deafness [new]
...is quantified via a protein folding-informed Bayesian model prioritizes uncertain missense variants, for biophysical rationale & upgraded diagnoses.

FishMamba-1: A Linear-Complexity Foundation Model for Deciphering Polyploid Cyprinid Genomes

Figure 1

Table 1

Figure 2

FishMamba-1: A Linear-Complexity Foundation Model for Deciphering Polyploid Cyprinid Genomes [new]
Linear SSM architecture analyzes 32kb context windows, capturing long-range deps for gene struct. annotation in complex aqua. genomes.

Figure 1

Figure 2

Figure 3

Figure 4

Cell DiffErential Expression by Pooling (CellDEEP) highlights issues in differential gene expression in scRNA-seq [new]
Improves scRNA-seq DEG by aggregating into metacells, flexibly reduces noise, preserves signal for reliab. anlys.

Pairing Data Independent Acquisition and High-Resolution Full Scan for Fast Urinary Tract Infection Diagnosis

Figure 1

Figure 2

Figure 3

Pairing Data Independent Acquisition and High-Resolution Full Scan for Fast Urinary Tract Infection Diagnosis [new]
DIA establishes pathogen-spec. peptide panels for rapid, cult-free UTI Dx via det. in cost-eff. MS1-only urine spec.

Characterizing Physicochemical Selection in Protein Evolution with Property-Informed Models (PRIME)

Figure 1

Figure 2

Figure 3

Characterizing Physicochemical Selection in Protein Evolution with Property-Informed Models (PRIME) [new]
...by modeling AA exchangeability based on physicochemical props to reveal biophysical selective constraints & adaptive tuning.

Figure 1

Figure 2

Figure 3

Figure 4

Unsupervised explainable AI reveals similar oligonucleotide-usage zones matching the highest-resolution human chromosome bands [new]
...by identifying ~2K regions oligo use, aligned to prophase bands, bridge cytogenetics & genome seq

Figure 1

Figure 2

HAETAE: A highly accurate and efficient epigenome transformer for tissue-specific histone modification prediction [new]
Uses 5mC from long-read seq in a 5-base framework, modeling epi-context to decode tissue-specific regulation.

Automated extraction and optimization of protein purification protocols using multi-agent large language models

Figure 1

Figure 2

Figure 3

Automated extraction and optimization of protein purification protocols using multi-agent large language models [new]
Identifies analogous proteins, extracts successful lit mthds, cross-refs failed prots for optimized recomms.

Decoupling Lineage and Intrinsic Information in Single-Cell Lineage Tracing Data with Deep Disentangled Representation Learning

Figure 1

Figure 2

Figure 3

Decoupling Lineage and Intrinsic Information in Single-Cell Lineage Tracing Data with Deep Disentangled Representation Learning [new]
...disentangles intrinsic cell states from lin. with deep gen. FW via disent. rep. & lin-aware GPs.

screenshot

Engineered TnpB genome editors for plants and human cells identified by ribonucleoprotein mutational scanning [new]
...facilitate highly specific and versatile gene modification in diverse biological systems.

Figure 2

Figure 3

Figure 5

Figure 6

Distinct cellular DNA methylation mechanisms underlie common and rare genetic risk for brain disorders [new]
common variants impact mCG in excitatory neurons for heritability, but rare de novo mut. affect neuronal mCH in autism.

screenshot

AF2BIND: predicting small-molecule binding sites using the pair representation of AlphaFold2 [new]
Predicts protein small molecule binding site residues via AlphaFold2 pair features, uncovers 1000s new ligandable sites in the human proteome.

Integrating morphology and gene expression of neural cells in unpaired single-cell data using GeoAdvAE

Figure 1

Figure 2

Figure 3

Integrating morphology and gene expression of neural cells in unpaired single-cell data using GeoAdvAE [updated]
aligns cell form & function in shared latent space, uncovers transcr. shifts & gene markers corr. w/ morph changes.

PhosSight: a Unified Deep Learning Framework Boosting and Accelerating Phosphoproteome Identification to Enable Biological Discoveries

Figure 1

Figure 2

Figure 3

PhosSight: a Unified Deep Learning Framework Boosting and Accelerating Phosphoproteome Identification to Enable Biological Discoveries [new]

Joint Geometric-Chemical Distance for Protein Surfaces

Figure 1

Figure 2

Table I

Joint Geometric-Chemical Distance for Protein Surfaces [new]
...integrates structural and physicochemical discrepancies via surface alignment of geometry and chemical fields, unifying functional interface comparison.

Reaction-Conditioned Enzyme Discovery with Multimodal Deep Learning

Figure 1

Figure 2

Figure 3

Reaction-Conditioned Enzyme Discovery with Multimodal Deep Learning [new]
Unifies reaction encoding w/ protein language models to enable zero-shot discovery of enzymes for previously unseen chemical reactions, moving beyond homology.

torch-projectors: A High-Performance Differentiable Projection Library for PyTorch

Table 1

Table 2

Table 3

torch-projectors: A High-Performance Differentiable Projection Library for PyTorch [new]
provides differentiable Fourier-space projections for EM (single-particle analysis, electron tomography) with 2D/3D operators and grad. support.

Developing SCL2205 : A Protein Sequence-based Spatial Modelling Dataset for the Protein Language Model Frontier [new]
High-quality dataset for robust DL protein localization modeling, stringent partitioned to minimize leakage.