scLSD (Single-Cell Latent Space Dynamics), a thermodynamics-inspired framework modeling differentiation on a learned Waddington landscape in latent space. It infers a low-dimensional cell state and a differentiable potential function governing developmental flow w/ local entropy.
scLSD uses a neural ordinary differential equation to reconstruct continuous differentiation trajectories from time-ordered single-cell data. It accurately recovers lineage hierarchies, predicts fate commitment for unseen cell types, and enables accurate prediction of trajectory direction and cell fate.
□ LOTUS: lineage-aware optimal transport reconstructs continuous clonal fate dynamics
LOTUS (Lineage-aware optimal transport for unbalanced single-cell dynamics), a framework for reconstructing continuous cell-fate dynamics from time-resolved, lineage-tracing single-cell data.
LOTUS formulates cell-population evolution as a lineage-regularized unbalanced optimal transport problem within a neural probability-flow framework. It jointly learns continuous transcriptional state dynamics, non-conserved cellular mass and lineage-consistent clone propagation.
□ GENESIS: Single-cell perturbation prediction near experimental reproducibility
GENESIS (Generative ENErgy-based Single-cell Intervention Simulator), a generative framework that predicts single-cell response distributions directly from the covariates defining an experimental condition.
GENESIS trains a neural network with the energy score. The attraction term encourages generated cells to resemble the observed cell, while the repulsion term promotes diversity to prevent collapse to a single mean profile and capture cell-to-cell variability.
□ MAP: A knowledge-driven framework for predicting single-cell responses for unprofiled drugs
MAP (Mechanism-Aware perturbation response Predictor) learns mechanism-aware drug and gene representations by aligning molecular structures, protein targets and mechanistic descriptions in a unified embedding space.
MAP conditions a perturbation predictor on these knowledge-informed representations. This design enables extrapolation to compounds with scarce or absent perturbation profiles, providing a mechanistically grounded route towards in silico screening at scale.
□ CellCompass: Building virtual cell dynamics that evolve with biological measurements
CellCompass combines a scientific agent with an extensible dynamical engine built on modular transport, bridge and flow-matching formulations. CellCompass can recover cellular dynamics under controlled ground truth.
CellCompass designed LineageBridge as an unbalanced velocity-growth flow-matching model in which clone barcodes restrict transport between time points, terminal clone composition allocates transported mass across future fate branches and branch mass represents fate probabilities.
□ CausalODE: Inferring gene regulatory networks from single-cell perturbation data by optimizing causal reach of latent mediator dynamics
CausalODE, a differentiable framework for learning cyclic regulatory graphs by nodeling RNA dynamics and delayed regulation through latent mediators while optimizing causal reach.
CausalODE learns a graph that optimizes the causal reach metric, captures transient cellular responses to perturbation, and explicitly models the latent dynamics of protein levels. CausalODE allows nonlinear responses, feedback loops, and time-lagged regulation.
□ Learning Strand-Equivariant Tokenizations of DNA with MirrorDNA
MirrorDNA, a compact DNA language model with a learned tokenizer that guarantees mirror-symmetric token boundaries across reverse-complement sequences, regardless of its learned parameters. This symmetry holds through the whole model.
EMBER, a gene-centric foundation model that learns context-dependent functional representations of genes. EMBER combines interaction-guided contrastive learning w/ masked expression reconstruction during pretraining on a single-cell corpus curated for diversity across cell types.
EMBER decouples gene representation learning from task-specific prediction through a composable embedding framework: the pre-trained encoder remains frozen, while lightweight probes combine contextualized gene embeddings to predict individual and combined perturbation effects.
□ scDisentangle: sequential disentanglement for single-cell perturbation response prediction
scDisentangle, a generative model that decomposes each cell into attribute-related embeddings and a sequence of residual factors.
Each successive factor captures the residual variation left unexplained by attributes and earlier factors, and as a result information shared between the attribute embeddings and the residual factors is suppressed by construction.
□ scCont: Interpretable Contrastive Learning for Functional Characterization of Cellular State Transitions
scCont, a fully unsupervised contrastive learning framework for extracting interpretable functional transition units. scCont uses an unsupervised k-nearest neighbors approach to project local transcriptomic neighborhoods into a low-dimensional embedding.
scCont employs Shapley network attribution to determine gene-latent relationships, mapping latent dimensions directly into functional gene programs.
□ GSET: Gene Signals from Embedding Trajectories select compact contexts for mechanistic GRN inference
GSET (Gene Signals from Embedding Trajectories) is an upstream retrieval framework that extracts dynamic information from temporal gene embeddings using two complementary signals.
An optimal transport residual captures gene-specific deviations from global embedding movement, while a KL divergence detects changes in local co-expression neighborhoods.
AutoDNA, an AI-native autonomous biomolecular engineering laboratory targeting multi-objective experiments across diverse instruments. AutoDNA realizes the co-evolution of AI and the automation system within an end-to-end multi-user framework.
□ scRM: A single-cell foundation model trained on scRNA-seq reads
scRM, a hierarchical transformer inspired by the byte latent transformer. It uses the rich sequenced reads, rather than read alignment or a count matrix, so models can be trained on one scRNA-seq experiment, and any trained model can be applied to arbitrary species.
□ DCAF: Unifying Data Validity, Evaluation Validity and Predictive Modeling for Single-cell Perturbation Simulation
Data-Criterion-model Associated Framework (DCAF), unified framework for single-cell perturbation simulation that jointly realizes data validity, evaluation validity and modeling predictability.
DCAF-model adopts a Joint Embedding Predictive Architecture (JEPA) to enhance the encoder's understanding of pre- and post-perturbation status.
DCAF incorporates the first neural network that explicitly models regulatory dynamics, providing interpretable and reliable prediction at the full transcriptome level.
□ ICL: Efficient biological design via genomic in-context learning
Evo2 learns to generate functional regulatory DNA sequences by simply concatenating in-context examples in the model's prompt. ICL can offer a data- and compute-efficient way for Evo 2 to generate proteins, protein-RNA systems, and even whole genomes.
The resulting inference engine (Evo2-flash) exhibits significant gains in both inference speed and memory efficiency. Evo 2 7B generates sequences with up to 7,000 bases per second on an H100 GPU, accelerating generation from 9.7x to 35x compared to Evo 2 with vortex.
□ SMFNet reveals principles of combinatorial transcription factor assembly
SMFNet, a generative diffusion model that predicts distributions of single-molecule chromatin profiles from DNA sequence, and PRISM, an interpretation method that attributes individual molecular configurations to predictive sequence features.
SMFNet captures chromatin-state frequencies and dependencies between TF footprints and nucleosome occupancy. SMFNet varies TF identity, number, and spacing in silico to determine how individual and combined TFs shape chromatin accessibility and nucleosome positioning.
□ GeoSinkhorn Flow: Geometry-Aware Flow Matching for Perturbation-Conditioned Dynamics in Single cell Data Phenoscapes
Geosinkhorn Flow extends a static Geodesic-Sinkhorn phenoscape into a continuous and predictive dynamical system. It provides geometry-aware supervision for local population transitions, which conditional flow matching integrates into continuous, condition-aware trajectories.
GeoSinkhorn Flow trains a geometric autoencoder whose latent Euclidean distances remain consistent with the heat-diffusion geometry underlying the phenoscape, allowing trajectories learned in this space to follow the same manifold structure used to organize the static map.
□ COMPASS: Systematic mapping of the prokaryotic noncoding sequence space with a genomic language model
COMPASS (COMPrehensive Atlas of Structured Sequences) uses embeddings from Evo-PIN, a fine-tuned genomic language model trained on annotation-free DNA sequences, to build a global map of the prokaryotic intergenic sequence space without alignment requirements or limitations on scale.
Probe-based search in the embedding space deorphanized divergent members of characterized ncRNA classes that fall below the thresholds of their covariance models. Rubric-guided batch investigation nominated structured RNAs with no family assignment using genomic context evidence.
□ NTv3: Nucleotide Transformer: Foundation Models for Genomics:
NTv3 is a genomic foundation model that reads DNA at single-base resolution across up to 1 Mbof context in one pass. One architecture supports three capabilities: functional-track prediction, gene annotation and regulatory sequence design.
NTv3 employs a U-Net-style architecture: a convolutional tower that reads the sequence, a Transformer tower that captures long-range context, and a deconvolutional tower that returns predictions at base resolution.
□ Gene Embedding Hub: a standardized resource for human gene representations
Gene Embedding Hub, a standardized collection of human protein-coding gene embeddings derived from protein, DNA and RNA sequences, biomedical text, bulk expression data, single-cell expression models and multimodal models.
The Hub provides a reusable basis for benchmarking gene embeddings, evaluating alternative sources of biological information in predictive models, and studying different combinations across modalities.
□ REGA-perturb: Resolving how genetic perturbations act through gene regulatory paths
REGA-perturb, a Reference REgulatory Path-Guided Analysis framework for tracing single-cell perturbation responses through gene regulatory paths.
REGA-perturb models the propagation of perturbation signals through a hierarchical regulatory architecture, linking perturbation targets to downstream transcriptional responses via latent regulatory programs, transcription factor motifs, regulatory elements and target genes.
REGA-perturb represents each perturbation target by an activity profile across a shared set of latent regulatory programs. It reconstructs observed transcriptional profiles and provides an interpretable decomposition of perturbation-associated regulatory effects.
□ LEEN: local equivariant energy networks with evolution-guided structural reasoning for protein stability prediction
LEEN (Local Equivariant Energy Network), a lightweight ΔΔ G prediction framework that combines thermodynamically consistent energy modeling with evolutionary-guided geometric reasoning.
LEEN constructs a spatial graph of residues within this radius and processes it with an E(n)-equivariant graph neural network (EGNN) that respects the rotational and translational symmetries of three-dimensional molecular structure.
LEEN employs an evolution-guided message gating, a mechanism by which frozen protein language model embeddings modulate the structural message passing at every layer of the equivariant encoder.
□ TRACER: Twin-contrastive learning for fine-grained and rare spatial domain identification in spatial transcriptomics
TRACER, a twin-contrastive learning framework for fine-grained spatial domain identification, with a focus on rare-domain-related and transitional-layer settings in spatial transcriptomics.
TRACER employs a graph convolutional autoencoder to learn representations of spatial spots from gene expression and their spatial neighborhoods and further incorporates two complementary contrastive objectives.
□ RhizAeon: Clean Breaks: Attention Force Field Deconvolution of Recombination Across Sequence Manifolds
RhizAeon models recombination through continuous trajectory mechanics and zero-dritt force field dynamics in linear time (O(NKL)). RhizAeon transforms recombination detection into a physical trajectory deconvolution problem.
Projecting alignments to canonical metric coordinates and routing lineage-specific rate variation into an unpolarized root sink, the engine iteratively trims mosaic tracts and tests breakpoints via Brownian-bridge statistics, dyadic matched filters, maximum-likelihood polishing.
□ RFhap: accurate alignment-free long-read phasing for improved diploid de novo assembly
RFhap, a trio-based phasing method that combines multi-k-mer information with machine learning to improve long-read haplotype assignment. It uses standard trio data and an alignment-free k-mer engine to look up parent-specific k-mers of multiple lengths in offspring long reads.
biHash, a bisulfite-aware rolling hash algorithm that utilizes context-specific dinucleotide encodings to support matching between selected reference and bisulfite derived k-mers.
□ Mimic: A Unified Neural Network Architecture for Modelling Genomic Sequences and Module-level Comparison
Mimic is the sequence-embedding architecture package in the Poseigen family. It provides unified single- and dual-input model. Mimic was used to model two different types of genomic datasets: protein-DNA interactions and enhancer activity. Along with other architecture components, different sequence embedding strategies were tried such as recurrent layers and convolutional layers.
□ ProtCell: a foundation model for single-cell proteomics
ProtCell combines protein-identity and continuous-abundance embeddings in a transformer encoder and learns cell- and protein-level representations through masked abundance modeling.
ProtCell connects fragmented profiles through transferable annotation and batch correction while preserving biological organization. Contrastive fine-tuning aligns cell embeddings across batches using unbalanced optimal transport, while preserving within-batch structure.
□ clrcycle: Dimensionality Reduction for Cyclic Compositional Data
clrcycle applies a centered log-ratio transformation, learns a cyclic feature ordering by maximizing variance captured by the first Fourier mode, and projects samples onto the corresponding cosine and sine coordinates.
The resulting two-dimensional geometry assigns each sample a phase and a radius measuring the magnitude of its projection onto the learned Fourier circle.
□ Verix: a toolkit for benchmarking and harmonization of complex structural variants
Verix introduces granular diagnostic metrics that enable researchers to thoroughly characterize CSV predictions and stratify performance based on breakpoint sensitivity, precision, and structural integrity.
□ Starlng: STability Analysis of coReguLated "Nests" of Genes
Starlng (STability Analysis of coRreguLated Nests of Genes), a framework designed to evaluate and improve the robustness of graph-based gene clustering, leading to the identification of gene modules that characterise transitions between cell states across pseudotime trajectories.
□ Scooby-2: fine-tuning AlphaGenome for single-cell genomic track prediction
Scooby-2 combines AlphaGenome with a cell-specific decoder to predict single-cell RNA-seq coverage and chromatin-accessibility genomic tracks from DNA sequence.
Scooby-2 improves variant effect prediction, while nonetheless still being prone to the decay in variant effect prediction fidelity reported for sequence-to-function models.
□ GaussF: An Unbiased Gaussian Framework for High-Precision Transcript Quantification with a Programmable Self-Benchmark Model
GaussF, a novel alignment-free methodology specifically designed for precise isoform quantification by directly addressing these biases and the isoform recognition problem without relying on empirical data.
GaussF is predicated on the observation that k-mer counts categorized by GC content across a natural transcript exhibit an intrinsic Gaussian distribution. GaussF characterizes isoforms through identifying minimal-shared regions and restructures the k-mer reference accordingly.
□ CauTrigger: deciphering biological system state transitions by hierarchical causal decomposition
CauTrigger, a hierarchical deep latent variable model that integrates do-calculus to identify potential causal triggers and predict perturbation outcomes across multiple regulatory layers.
By disentangling causal and spurious latent factors, CauTrigger recursively reconstructs the underlying regulatory hierarchy and enables interpretable reasoning from observational data.
□ Kontext: Kernel-based method bridging spatial domain identification and cell-cell communication
Kontext, a kernel-based framework built on a dual characterization of the cells based on both their expression and local environment to identify spatial domains and their underlying mechanistic origin in terms of ligand-receptor interactions.
Kontext relies on the computation of the "local environment embedding" which summarizes each cell local context as a weighted average of neighboring gene expression profiles, with weights jointly capturing spatial proximity and transcriptomic similarity both encoded with kernels.
□ flashscenic: Accelerated GRN Analysis with RegDiffusion for Large-Scale Single-Cell Data
flashscenic, a GPU-accelerated framework that broadens the practical scope of motif-validated GRN analysis for large scale single-cell transcriptomics.
Vectorized GPU Architecture and Algorithmic Equivalence: flashscenic replaces key bottlenecks of the original pipeline with GPU native implementations, incorporating RegDiffusion for network inference together with fully vectorized PyTorch implementations of cisTarget pruning and AUCell scoring.
□ dscNanoATAC/dscNanoCUT: A droplet-based long-read single-cell epigenome sequencing method for the high-throughput profiling of multidimensional regulatory elements within individual cells
dscNanoATAC leverages Nanopore long-read sequencing and involve a commercial microfluidic device for the high-throughput, genome-wide profiling of the epigenome at the single-cell level.
dscNanoCUT is a computational pipeline tailored for long-read droplet sequencing data to investigate allele-specific epigenetics, co-accessibility, and regulation in complex genomic regions.
□ Pycea: a scverse toolkit for single-cell lineage tracing analysis
Pycea, a scverse-compatible toolkit for lineage analysis and visualization built around TreeData, an extension of AnnData that supports the storage of lineage trees alongside other single-cell measurements.
Pycea provides a suite of intuitive and scalable methods for visualizing lineage trees, identilying clades, comparing lineage distances, inferring ancestral states.
□ Local fractal dimension as an indicator of functional reorganization in molecular interaction networks
DIFFRACT (Differential Fractal Analysis of Condition-Specific Topologies) integrates replicated molecular-abundance profiles with interaction scaffolds. It combines quantitative proteomics with PPI networks to identify network reorganization through changes in fractal dimension.
□ scSTEC: Annotating Chromatin Subcompartments from Individual Single-Cell Hi-C Data via Direct Two-Dimensional Embedding and Chromosome-wise Consensus
scSTEC (Single-cell Subcompartment Annotation via Two-dimensional Embedding and Chromosome-wise Consensus), a computational framework for direct subcom-partment annotation from individual scHi-C matrices.
The foundation of scSTEC rests on the observation that the construction of the estimated inter-chromosomal matrix used by scGHOST is intrinsically of rank 2.
This low-rank structure reveals that the geometry underlying the clustering of genomic bins in this high-dimensional representation can be captured in two dimensions, thereby avoiding the need to explicitly construct and cluster the high-dimensional matrix.
□ VIGAT: VLM-guided gene activity quantification from single-cell ATAC-seq
VIGAT, a vision-language model (VLM)-guided tool that automates visual inspection of ATAC coverage plots to adaptively identify informative genomic regions for gene activity quantification.
VIGAT jointly examines accessibility patterns across multiple marker genes, integrating genomic annotations and biological context to select gene-specific regions enriched for cluster-discriminative signals. Selected regions are used to generate a gene-by-cell activity matrix.
□ AdaGeneBudget: Cell-Adaptive Gene-Token Allocation for Efficient Single-Cell Foundation Models
AdaGeneBudget combines reference-derived inverse detection-frequency weights with each cell's expression values and retains the shortest ranked prefix that reaches a target score mass within budget bounds.
Cells with concentrated score mass therefore receive shorter inputs, whereas cells with more diffuse score mass receive longer inputs.
RNA-processing elements were integrated into endogenous transcription units at tens of thousands of mapped positions, with different elements tested at shared positions.
Genomic position dominated RNA production levels but differentially influenced RNA processing events. A cryptic polyadenylation site, used natively only upon U1 inhibition, was rarely used ectopically.
□ BioSimBench: Can language models simulate biology?
A benchmark of real lab experiments. It tests how well models predict how cells respond to drugs, gene knockdowns, and immune signals they have never seen.
The replicate baseline is especially useful: a model should recover the genes that change consistently across repeated experiments. Optimizing against a single measurement risks rewarding experimental noise as biological accuracy.
□ RAPTORGraph: graph-based pathway modeling for causal discovery in single-cell perturbations
RAPTORGraph is a β-variational autoencoder with a GraphPathway encoder that explicitly models complex gene-to-gene interactions within learned pathways.
RAPTORGraph isolates the influence of perturbed genes through preconditioning, yielding clean, single-node latent interventions required for identifiable causal discovery and eliminating spillover.
□ CloneVI: A deep generative model for deciphering the relationships between cell state and mitochondrial genotype
CloneVI, a deep generative model that infers clonal structure from a latent cell state. It estimates the clonal probability of each cell from this cell state, together with the allele frequency of each clone by reconstructing the observed mutation counts.
CloneVI recovers the ground-truth clonal structure in both single-clone and mixed-clone cell popula-tions. In clonal hematopoiesis single-cell RNA-seq data with enriched mitochondrial transcripts, the inferred clones were associated with specific mitochondrial DNA mutations.
□ All-Atom Structure Tokenizer: AtomLM-9b for Biomolecular Structure Prediction
The All-Atom Structure Tokenizer compresses biomolecular geometry into discrete structural tokens and achieves high-accuracy all-atom reconstruction.
A hierarchical geometric encoder and finite scalar quantization map chemical units to a shared vocabulary of 8,192 tokens, while a joint coordinate de-tokenizer reconstructs backbone, side-chain, nucleic-acid, and ligand heavy atoms.
□ MoleMap: fast alignment-free molecule mapping for long-read and linked-read sequencing data
MoleMap, a "molecule mapping" method for long reads and linked-read molecules that pre-filters WGS data for downstream targeted analyses.
MoleMap uses a minimized open-addressing k-mer index of the reference genome and a fast k-mer clustering procedure to assign reads or molecules to genomic regions.
□ Tracel: Unified cell segmentation for spatial transcriptomics via transcript reassignment
Tracel (Transcript to cell mapping) formulates cell segmentation as mapping each transcript to its cell of origin. Tracel proceeds in two optimal transport (OT) steps.
Tracel formulates a fused Gromov-Wasserstein OT problem that combines cross-dataset expression similarity with the similarity structure within each dataset, yielding an expected expression profile for each spatial cell.
□ CRISP: Uncovering cross-signal interactions in single-cell and spatial omics
CRISP, Cross-Signal Representation Interaction via Similarity Projection, a framework that characterizes the interactions between a pair of biological signals through the correlation structure of the expression patterns they induce in single-cell data.
□ brisc: a multithreaded CPU toolkit for single-cell analysis
brisc, the first end-to-end multithreaded CPU toolkit for single-cell analysis. brisc spans preprocessing, dimensionality reduction, integration, label transter, clustering, embedding, differential expression and plotting.
□ scPhyloGRN: gene regulatory network inference guided by lineage topology in single-cell developmental data
scPhyloGRN, a supervised framework that integrates single-cell gene expression with cellular phylogeny to prioritize directed transcription factor (TF)-target candidates within the observed gene set.
scPhyloGRN maps cell ancestry to gene associations that guide graph construction, attention in a dual-stream encoder, and lineage-consistency regularization.
□ Z-value: Theoretical prediction of assembly contiguity
Z-value is a formula that takes as input the read length, coverage, and basic properties of the genome, and returns a prediction of the mean contig length produced by a simple single-k assembler on error-free data.
□ spearfish: ultrafast Δ-mapping of long reads to redundant transcriptomes
spearfish, a long-read mapper built for this redundancy. Its delta mapping (Δ-mapping) computes each target's alignment-score difference from the best candidate. spearfish skips sequence that all candidates share, which cannot change the difference, and aligns where they differ once per distinct sequence rather than once per target.
Astrolabe, a programmable closed-loop framework that supports diverse biomolecular design tasks by composing design primitives.
Astrolabe alternates all-atom structure prediction with sequence design under user-defined objectives. Astrolabe synchronizes updates across residues, chains or complexes, and intervenes when predicted structures miss intended geometries.
□ ASTRA enables zero-shot super-resolution of spatial transcriptomics across capture geometries
ASTRA (Assay-aware Spatial Transcriptomic Reconstruction and Allocation), a cross-platform pretrained model that reconstructs expression from histology and measured RNA by zero-shot inference.
ASTRA uses an arbitrary-parent representation of count-bearing regions and learns from high-definition data aggregated over parents of varied shape and size. ASTRA enables frozen inference across capture geometries.
Carbon-X, an open 14-billion-parameter model for DNA understanding and generation. Carbon-X brings DNA sequences and natural language into one model: researchers can ask questions about a sequence, predict the effects of mutations, and generate new DNA.
□ DynaVelo: Deep dynamical models of single-cell multiomic velocities predict loss-of-function and rescue perturbations in B cells
DynaVelo, a generative neural ordinary differential equation model that learns the joint dynamics of gene expression and transcription factor (TF) motif activity.
DynaVelo can learn dynamic GRNs using two approaches: Jacobian matrices and in silico perturbations. Jacobians measure derivatives of changes, while in silico perturbations measure finite changes.
Single-cell states and velocities learned in the latent space define a vector field over cells, which are mapped back to the observational space by a decoder. DynaVelo also learns latent times for all cells, providing a cell ordering in the joint vector field.
□ CellDiffusion: Enhancing scRNA-seq data integration with accurate inference of cell-cell association
CellDiffusion is a graph-based framework that models intercellular relationships by constructing a cell-cell network and propagating transcription-derived features across cells. The diffusion process gradually homogenizes connected cells, enabling biologically coherent alignment.
CellDiffusion uses an Optimized Mutual Nearest Neighbor (OMNN) algorithm to connect cells across batches, linking putatively matched cell types. CellDiffusion tolerates incorrect graph connections since they are expected to be sparser relative to correct biological linkages.
CellDiffusion leverages diffusion dynamics to intrinsically attenuate weak associations. CellDiffusion uses an attention mechanism to guide and optimize the feature propagation through all diffusion layers in the latent space.
□ scChronos: Hierarchical Temporal Context Learning for Missing Snapshot Prediction in Single-Cell Time Series
scChronos, a hierarchical temporal context learning framework that formulates missing-snapshot prediction as target-conditioned population distribution prediction rather than individual-cell propagation.
scChronos introduces within-snapshot aggregation: it compresses each observed snapshot into compact cluster-representative cell embeddings, filtering out cell-level noise while preserving population information.
scChronos employs a dual-branch temporal context integration module, in which a global branch attends to all reterence snapshots according to their relevance to the target time and a local branch captures immediate transitions from nearby time points.
□ LoCoFlow: Local-Context Flow Matching for Single-Cell Dynamics and Fate Inference
LoCoFLOW constructs a point cloud from its local neighborhood, thereby extending the fundamental unit of dynamic modeling from an isolated expression vector to a local population state.
First, matching point cloud distributions across time points integrates cell states with neighborhood geometry and employs probabilistic pairing to preserve potential multiple successor correspondences.
LoCoFLOW decomposes point cloud evolution into item movements and changes in their internal discrete structure.
□ Katalyst: Knowledge-Guided Semantic Alignment via Contrastive Learning for Enzyme Turnover Prediction
Katalyst, a knowledge-guided semantic alignment framework for kcat prediction based on contrastive learning. It integrates enzyme sequences, protein textual descriptions, substrates and products into a shared latent space using multi-channel encoders and contrastive alignment.
Katalyst incorporates functional descriptions as complementary biological knowledge. Frozen pre-trained encoders equipped with trainable adapters project heterogeneous modalities into a unified latent space while preserving their domain-specific knowledge.
□ PIE: Generalizing perturbation effects across unseen perturbations, contexts and datasets
PIE (Perturbation is Everything) directly predicts differentially expressed genes and changes in gene expression for each (context, perturbation) pair, using external biological knowledge, baseline gene expression, and observed perturbation responses.
PIE decouples the size of the input and the output from the size of the model. Learned latent vectors cross-attend to the Knowledge Source tokens, and the latents then self-attend to one another. Finally, gene embeddings query the latents, producing a per-gene feature vector.
□ DNA-BFN: Bayesian Flow Networks for Generative Modeling of Regulatory DNA Sequences
DNA Bayesian Flow Networks (DNA-BFN), a generative framework tailored for regulatory DNA sequence design. DNA-BFN models sequence generation as an iterative Bayesian updating process over per-position nucleotide distributions, operating directly in probability space.
DNA-BFN realizes a diffusion-like generative process without explicitly perturbing discrete nucleotide sequences, enabling fully differentiable training while preserving discrete sequence structure.
□ LineageFlow: Growth-Corrected State Dynamics for Probabilistic Cell-Fate Prediction
LineageFlow, a model for finite-horizon, clone-disjoint prediction of descendant fate and transcriptional state. LineageFlow propagates expected lineage mass under state transitions and relative net growth, and generates continuous expression profiles.
LineageFlow maintains a distinction between biological time, which indexes the descendant forecast, and the interpolation time used to generate expression.
LineageFlow adapts the continuous-time Markov chain (CTMC)–flow architecture by changing the state and observation space and adding relative-growth propagation.
□ CONCERT: Generative modeling of spatially resolved perturbation responses
CONCERT, a niche-aware Gaussian process variational autoencoder that embeds each perturbation with the niche of the receiving cell and learns a spatial kernel for each per-turbation, whose lengthscale sets how far the effect extends in tissue.
CONCERT consists of a disentanglement module, a perturbation-signal module and a generative module. For counterfactual prediction, CONCERT changes the perturbation label of a cell, or a continuous attribute such as dose or time.
□ DeepD enables efficient genome modeling and controllable long-context sequence generation across domains of life
DeepD, a phylogeny-informed sparse mixture-of-experts model operating at single-nucleotide resolution over a 1-million-token context while reducing active computation.
Pretrained on unselected whole-genome sequences across all domains of life, DeepD learns genic and intergenic sequence features and supports broad sequence-to-function prediction across species.
DeepD was trained at two scales: a 1.8-billion-parameter model on 1 trillion tokens (DeepD-2B) and a 63-billion-parameter model on 5.69 trillion tokens (DeepD-60B). The DeepD-60B pretraining corpus comprised approximately 5.90 Tb of nucleotide sequence.
□ U-Chrom: a unified framework for single-cell and spatial 3D genome data
U-Chrom is built around a single data model, ChromData, which is stored in .chromdata.zarr format and read one cell at a time from disk or over HTTP. This model underpins its analyses, web browser, and public data atlas.
scOmnom automates routine analysis while keeping biological choices inspectable. BISC (biology-informed structural clustering) evaluates candidate Leiden resolutions using structural separation, agreement between adjacent partitions, and safeguards against overfragmentation.
scOmnom connects count-matrix processing to biological interpretation through a sequence of independently executable analysis steps. After quality control and integration, BISC selects a clustering resolution and cells receive reference-based annotations.
□ Bolero: Predicting cell-state-specific gene regulation from DNA across species
Bolero, a model that learns how sequence is read as a function of the cell state's position in an embedding space and predicts chromatin accessibility and transcript abundance from DNA at any point in that space.
□ Optimal transport reveals unresolved cells in single cell multi-omics
CCA-OT relies on agreement between shared features, Diffusion-OT incorporates within-modality graph structure, MaxFuse uses graph smoothing and iterative pivot refinement for weakly linked assays, and Ensemble-OT combines all three.
Moscot follows a related principle by using Wasserstein, Gromov-Wasserstein or fused Gromov-Wasserstein transport depending on the available cross-domain information.
DI-ElasticFlow keeps reference cells fixed and moves query cells smoothly toward coupling-derive positions. A single label-free Safe-upgrade rule accepts a correction only when modality mixing improves without weakening local neighbourhood structure or cell-type separation.
□ T2G: fast and accurate transcriptome to genome BAM liftover for short and long read RNA sequencing
LongVerse has three parts: a Nextflow pipeline, an MCP server that exposes it as 17 tools, and the agent that operates them. The pipeline runs from raw ONT or PacBio signal to haplotype-resolved methylation, with one process per tool and every tool in a pinned container.
□ Minerva: Coevolutionary discovery with a genome language model
Minerva, a framework for coevolutionary mining that uses genome language models to predict pairwise interactions directly from sequence. Minerva enables fast, alignment-free prediction of base pairing, monomeric protein contacts, and repetitive sequence motifs.
Given a genomic region, Minerva produces a coevolutionary map whose entries reflect how strongly, and in what way, each pair of positions interacts. It predicts these maps by passing the input sequence through a genome language model and extracting coevolutionary interactions.
DGP-LA, a dynamic graph-prior latent adaptation framework that refines biological priors and injects them as gated latent corrections within a cell-state-conditioned transition model.
DGP-LA couples a transition backbone with perturbation-conditioned sparse graph refinement. It propagates gene-aligned target support over the refined graph and selectively injects the resulting topology-aware signal into the hidden representation.
ProxiNet, a spatially supervised framework that learns transcriptomic signatures of pairwise cellular proximity from spatial reference datasets and transfers these relationships to dissociated single-cell transcriptomes.
ProxiNet predicts cellular proximity across brain regions and spatial technologies, incl. zero-shot cross-technology transfer, and gradient-based attribution identified genes and broader transcriptional programs associated with proximity predictions.
□ Quantum Hamiltonian-Based Generative Modeling of Single-Cell Transcriptomics for Gene Regulatory Network Inference
QHGM (quantum Hamiltonian-based gene expression model), a generative frame-work for modeling pseudotime-ordered single-cell gene expression data.
In QHGM, gene interactions are encoded via a parameterized Hamiltonian, and the outcomes of quantum measurements provide a discrete representation of the gene expression profile.
VQ-Net, a scalable variational quantum algorithm for network inference for learning Hamiltonian parameters from scRNA-seq data. The algorithm uses a quantum circuit to perform Hamiltonian evolution.
□ bi[o]hub: International, cross-sector collaboration commits nearly $2 billion to build foundational data for AI models to predict and treat disease
› DOE is investing $500M+ across measurement, modeling & compute
› NIH is bringing $500M+ of federally funded datasets & repositories
› Meta, Google DeepMind & @IsomorphicLabs are jointly investing another $300M
This consortium could make lab-to-lab variation something models must explain, rather than something we preprocess away. That would give “universal” real biological substance.
SORT, a matrix factorization framework that jointly recovers multilevel GEPs and their spatial activities, using orthogonality to reduce program redundancy and graph total variation to preserve spatial boundaries.
By factorizing multiple sections into a common transcriptional profile space, SORT resolves programs from sample-specific and group-shared patterns to region-associated signatures, with graph total variation to maintain spatial coherence and orthogonality to reduce redundancy.
SORT systematically distinguishes programs by their level of sharing and spatial deployment, enabling each level of program to be interpreted in its own biological context.
□ BioMod: Biological Module Tokenization for Single-Cell Foundation Modeling
BioMod, a 50-million-parameter single-cell foundation model that introduces biological modules as representation units between genes and cells.
BioMod constructs an overlapping vocabulary of 660 biological modules derived from curated pathways and data-driven gene programs, replacing thousands of gene-level inputs with compact module representations.
□ LoGicAl: Local ancestry and genotype calling uncertainty-aware ancestry-specific allele frequency estimation from admixed samples
LoGicAl, a novel likelihood-based method for estimating ancestry-specific allele frequencies from admixed samples, simultaneously accommodating uncertainty from ancestry calling, genotyping, and statistical phasing.
□ GLARE: Genomic Language model for Automated Repetitive Element classification
GLARE (Genomic Language model for Automated Repetitive Element classification) fine-tunes the NTv3 genomic language model for TE classification. It predicts TE superfamilies and derives orders from them, so labels at the two levels are always consistent.
□ METANet: a supervised ensemble learning framework for reconstructing direct and functional tissue-specific transcription factor networks
METANet (Motif Expression TF Association Networks), a supervised ensemble learning framework that combines TF motifs, cis-regulatory element activity, and linear and non-linear expression-derived features to predict TF binding.
METANet combines five distinct sets of regulatory edge features using XGBoost: a tissue-specific motif-based feature (Marbach) and four expression-based features derived from LASSO and BART regression on both tissue-specific and tissue-aggregate contexts.
□ scStable: assessing scRNA-seq analysis stability under bulk-derived between-sample variation
scStable, a framework for assessing scRNA-seq analysis stability by generating synthetic samples that preserve cell-level heterogeneity from a single-sample reference while injecting bulk-derived between-sample variation.
scStable better recapitulates observed between-sample variation than ZINB-WaVE, subsampling, and Gaussian noise injection. It enables stability-aware analysis, incl. prioritization of stable differentially expressed genes and selection of clustering methods and hyperparameters.
□ STAVelo: deciphering spatiotemporal cellular dynamics via spatial RNA velocity inference from spatial transcriptomics data
STAVelo, a graph attention encoder-decoder framework that models spatial information in spatial transcriptomics data and infers RNA velocity from neural representations of transcriptional kinetic parameters.
STAVelo employs the above transcription kinetic model to infer RNA velocity from ST data. The projection of RNA velocity onto tissue space, termed spatial velocity, exhibits distinct biological significance compared to cellular velocity.
□ gRely: Relyability for genome-trained sequence-to-function model predictions
gRely, a meta-model that estimates a score that an individual VEP is correct, from 1,161 features derived from the target variant, gene, tissue, and model.
gRely remains discriminative among low-magnitude variants, which effect-size filtering discards; SHAP attribution shows that reliability in the low-effect regime is driven by chromatin context and model cross-replicate agreement rather than effect size.
Furthermore, the Borzoi trained meta-model generalizes across other architectures, such as AlphaGenome and Enformer.
□ GenoCLAMP: consolidation of sequence motifs from genomic-scale data
Genomic CLustered Alignment of Motif Profiles (GenoCLAMP), an algorithm that unifies motif-clustering and alignment to collapse redundant motifs into high-quality merged matrix representations.
GenoCLAMP uses a greedy clustering algorithm that intertwines the alignment process with every step of the clustering process. This circumvents the need for post-clustering alignment, so the BLLR of each cluster directly corresponds to its final alignment.
Dipler, a two-step strategy that combines the identification of group-specific k-mers from whole-genome sequencing data with de novo assembly of PacBio reads enriched for those k-mers.
By removing randomly shared k-mers arising from insufficient phenotype representation and filtering out reads carrying only a small number of phenotype-specific k-mers, they recovered the full locus directly from sequencing data.
□ SVC: A Vision Transformer-based Spatially and Virtually Embedded Cell Model for Deciphering Subcellular Spatial Transcriptomic Heterogeneity
Spatially and Virtually Embedded Cell (SVC), a Vision Transformer-based spatially and virtually embedded cell model that learns subcellular transcript organization within native tissue context dırectly from subcellular-resolution spatial transcriptomics.
SVC represents each gene's subcellular localization as an image patch and uses other genes within the same cell as context, enabling subcellular contextual learning of spatial dependencies that govern intracellular organization.
□ HitSV: Structural variant detection across sequencing technologies
HitSV is an alignment-based SV detection approach supporting long-read, short-read and hybrid sequencing data. It addresses a long-standing limitation-alignment signals are systematically distorted in repetitive and complex regions, leading to missed, fragmented variants.
HitSV redesigns the core stages -SV site recognition, SV signature clustering, and SV reconstruction and genotyping-by explicitly modeling the interaction between genomic repetitiveness and signature intensity.
□ VINTER: a generative vision-language model for annotating and interrogating spatial tissues
VINTER (Vision-language Interrogation of Neighborhood Transcriptomic Evidence and Reasoning), a generative vision language model that reads histology, measured gene expression and spatial neighbourhood as one sequence.
VINTER connects the pathology foundation model UNI to a 7-billion-parameter language decoder aligned on 1.45 million histology-caption pairs from 546 spatial transcriptomic samples.
□ HyperCom: a hypergraph-based method to infer cell-resolved cell-cell communication
HyperCom constructs a hypergraph from known ligand-receptor interactions and the expression of those ligands and receptors in individual cells and performs graph diffusion on the hypergraph to score cells on a continuous spectrum from sender to receiver for each LR pair.
HyperCom generates a cell-type agnostic ranking of LR pairs by communication across the dataset or by condition. Validation of HyperCom on physical cell-to-cell contact, spatial, and multicondition datasets highlights its ability to capture nuanced per-cell communication signals.
□ PORTAL: Rethinking Projector Training in Multimodal LLMs
PORTAL (Projector via Optimal-transpoRT ALignment), a training-free approach that constructs the projector from the marginal statistics of the two modalities.
PORTAL is the Bures-Wasserstein optimal transport map between Gaussian approximations of the vision feature distribution and the language-model input distribution, computed in a shared top-k principal subspace.
□ SHERLOCK: Structured representation learning and causal inference of downstream perturbation effects
SHERLOCK (Single-cell Hierarchical Embedding of peRturbations and Learning Of Causal Knowledge), an interpretable deep generative framework for single-cell perturbation analysis that represents perturbation effects as structured interventions on a latent baseline cellular state.
SHERLOCK learns correlated and sparse perturbation representations that organize genetic and pharmacological perturbations according to shared transcriptional responses.
SHERLOCK formulates perturbations as interventions within a structural causal model, enabling counterfactual estimation of their downstream transcriptional effects under explicit identifiability assumptions.
□ RAGCell: Retrieval-Augmented Generation as Supervision for Versatile Single-cell Analysis
RAGCell leverages LLMs to construct cell-level and feature-level knowledge databases, which serve as supervision signals for training the cell model and significantly reduce the training cost.
Aligning cell representations with text embeddings from the bi-level knowledge databases, enabling knowledge transfer from textual spaces to cellular spaces and effectively mitigating the heterogeneous gap.
□ SCITRAM: a Single-cell Integrated transcription regulation modeler for predicting master transcription factors
SCITRAM (Single-Cell Integrated TRAnscription regulation Modeler), a user-friendly stand-alone computational solution for predicting master transcription factors from single-cell transcriptomics data.
SCITRAM reconstructs a primary gene regulatory network and subsequently uses this network as template to model transcriptional cascades driven by the in-silico activation of transcription factors identified within the system.
□ scXMatch: Quantifying distribution shifts in single-cell data
scXMatch (single-cell cross match) is a Python package that implements Rosenbaum's cross-match test using distance-based matching to assess statistical dependence between two groups of high-dimensional data.
scXMatch scales to realistically sized single-cell datasets and that it can reliably distinguish signal from noise, using experimental perturbation of varying strength as ground truth.
□ RASTRO: Branching Point Decision-Making in Unitig Graphs
RASTRO, a reference-free framework for scoring candidate paths at unitig branching points using nucleotide sequence information. RASTRO-S employs a sample-specific n-gram model to estimate extension plausibility via bits-per-base surprisal.
RASTRO-L uses the Evo2 pretrained DNA language model for autoregressive path scoring. Both operate directly on unitig graphs without reference genomes or assembled contigs.
□ QuImputer: an LD-informed QUBO/Ising formulation of genotype imputation with quantum-hardware execution
Qulmputer to encode allele-frequency priors and signed linkage disequilibrium (LD) in a reduced Ising/QUBO objective, with only missing genotypes represented as optimization variables.
The same objective can be passed to exhaustive classical search, quantum approximate optimization algorithm simulation, or quantum hardware.
□ CRAIC: interactive interrogation of multiple sequence alignment uncertainty
CRAIC (Conserved-Region Alignment by Iterative Convergence) is a free desktop workbench for building, viewing and, above all, interrogating multiple sequence alignments.
□ GeneICL: A Tabular Foundation Model for Bulk Transcriptomics
GeneICL is a 4.2M-parameter tabular foundation model combining a semi-synthetic pretraining prior built from measured bulk expression profiles with a parameter-efficient recurrent architecture.
GeneICL operates in principal component (PC) space and uses a recurrent, weight-tied in-context learner to predict query samples from labeled support examples. A single pretrained model supports both regression and classification.
□ DECONVersation: Single Cell Foundation Model-Derived Embeddings for Robust Cell Type Deconvolution of Bulk RNA-seq
DECONVersation leverages single-cell foundation models (scFMs) trained on large-scale single-cell datasets to generate embeddings, high dimensional vectors that capture the context of an entire cell/sample by aggregating information across thousands of gene tokens.
DECONVersation incorporates cell embedding extraction methods from four scFMs: Geneformer, Cell2Sentence, scGPT, and CellHermes. Foundation model embeddings are robust to batch effects and implicitly encode complex gene relationships, sidestepping explicit marker gene selection.
□ scGSCA: gene-set differential co-expression analysis for single-cell RNA sequencing data
scGSCA (single-cell gene-set co-expression analysis) fits gene-wise zero-inflated negative binomial mixed models with subject-level random effects and computes gene-set co-expression statistics from quantile residual correlations.
scGSCA provides an overall differential gene-set co-expression analysis and an interaction test to assess whether the co-expression pattern between two cell types varies by sample phenotype.
□ CLARA: Compound Library Active Learning with Reasoning Agents
A language-model agent can successfully allocate exploit, uncertainty, and diversity queues during batch ligand discovery. It consistently improves over random sampling and can match the strongest fixed policy on some targets, but it is not universally superior.
□ Efficient approximate Bayesian inference for large, heterogeneous biological datasets
An adaptive minibatch Sequential Monte Carlo Approximate Bayesian Computation algorithm that changes the batch size as inference proceeds.
Intuitively, coarse selection is sufficient during early exploration, but refining the posterior requires comparisons that represent more of the dataset. It derives a minibatch size update rule that keeps the extra noise from minibatching in balance with the ABC tolerance.
□ AnnZarro: scalable, interactive exploration of cell-by-cell and gene-by-gene relationships in single-cell data
AnnZarro is a point-and-click browser viewer for exploring large matrices of per-cell, per-gene and pairwise quantities linked to an AnnData object’s cells × genes data, allowing users to browse results too large to plot at once one slice at a time.
AnnZarro draws each of the 95.6 million cells of Tahoe-100M as a point. CELLxGENE and Cirrocumulus show datasets of one million or more cells in their documentation and Vitessce millions of points.
□ ALBIS: simulation of multi-resolution and multi-dimensional spatial transcriptomics data
ALBIS (A muLti-resolution Biological In-silico Simulator for spatial transcriptomics), a 3D simulation framework that generates multiple spatial measurement resolutions from a shared transcript-level molecular ground truth.
ALBIS constructs a 3D tissue with domain-dependent cell type composition, assigns explicit spatial coordinates to individual RNA molecules, and generates observations at cellular, subcellular, and multicellular resolutions.
□ DNA-LLM2Vec: Architecture-Aware Task-Conditioned Adaptation of Decoder DNA Foundation Models
DNA-LLM2Vec, a staged decoder-to-encoder framework that opens bidirectional context, repairs the resulting mismatch with masked next-token prediction, and applies contrastive views encoding strand and positional invariance.
DNA-LLM2Vec turns pretrained causal DNA decoders into frozen bidirectional encoders whose value is task- and architecture-conditioned rather than universal. DNA-LLM2Vec employs two architecturally distinct backbones: DNAGPT and HyenaDNA.
BioMINTREL (Biomedical Method Inference and Transfer from Research Evidence and Literature), an AI agent that expands the problem-solving space available to biomedical agents: the questions and methods they can consider for available data.
BioMINTREL draws on established methods from other disciplines and can start from biological data without a predefined question. It can formulate a joint task from relationships among several datasets, or develop a model by relating the computations and ideas of several methods.