🧬 Computational Biology¶
🧠 NeurIPS2026 · 66 paper notes
📌 Same area in other venues: 🎞️ ECCV2026 (7) · 📷 CVPR2026 (21) · 🔬 ICLR2026 (155) · 💬 ACL2026 (5) · 🧪 ICML2026 (52) · 🤖 AAAI2026 (20)
🔥 Top topics: Biomolecules ×27 · Diffusion Models ×7 · Reasoning ×2
- At FullTilt: Real-Time Open-Set 3D Macromolecule Detection Directly from Tilted 2D Projections
-
FullTilt feeds aligned 2D tilt-series directly into a visually prompted multiclass 3D detector, replacing volumetric sliding-window detection with cross-tilt row attention, near-zero-tilt query initialization, and training-time geometric augmentation to achieve subsecond zero-shot detection on three real-world cryo-ET datasets, while still requiring simulated-data pretraining and input alignment.
- Atom-level Protein Representation Learning Improves Protein Structure Prediction
-
TriProRep introduces a structure-aware pretraining framework that couples amino-acid identity, backbone geometry, and intra-residue full-atom geometry with generator-corrupted corrective pretraining, delivering consistent gains in homodimer flexible docking, residue-level binding property prediction, and monomer structure prediction distillation on the proposed RepSP benchmark.
- Atomic Trajectory Modeling with State Space Models for Biomolecular Dynamics
-
ATMOS is the first all-atom biomolecular dynamics generation framework powered by State Space Models (SSMs), coupling a bidirectional Pairformer latent propagator with conditional diffusion decoding to achieve linear computational complexity and superior fidelity across both protein monomers and protein-ligand complexes.
- Benchmarking Compositional Generalisation for Machine Learning Interatomic Potentials
-
The GMD benchmark systematically evaluates the compositional generalisation of ML interatomic potentials across five controlled tasks—chain extension, fusion, duplication, combination, and fission—revealing that state-of-the-art models and fine-tuned foundation models degrade by 1 to 3 orders of magnitude OOD, with failures driven predominantly by force magnitude extrapolation rather than directional orientation.
- Beyond Drug Discovery: The Nanotechnology Molecular Optimization (NMO) Benchmark
-
NMO tests molecular optimization beyond pharmaceutical priors through three nanophysics tasks constrained by electrode binding, and combines explicit-anchor GGS representations, random-molecule pretraining, and genetic-guided GFNs to discover scientifically promising candidates, although the full model does not maximize AUC on every task.
BioBlobs: Unsupervised Discovery of Functional Substructures for Protein Function Prediction
- Branching Flows: Discrete, Continuous, and Manifold Flow Matching with Splits and Deletions
-
Branching Flows introduces a generative framework based on Generator Matching that transports probability paths over forests of binary trees with learned splitting and deletion rates, enabling variable-length discrete, continuous, and manifold generation while solving the unknown-length infix sampling challenge in antibody and protein design.
- Budgeted Quotient-Residual Guidance for Frozen Pocket-Conditioned Molecular Diffusion
-
Addressing dormant quotient gradients in pocket-conditioned molecular diffusion, Budgeted Quotient-Residual Guidance (QRG) lifts quotient covectors to metric-horizontal ambient directions and caps the correction by the frozen sampler's own step norm, delivering effective inference-time lead-optimization guidance without retraining.
- C3P: Contrastive promoter-protein pretraining yields representations capturing bacterial gene regulation
-
Inspired by multimodal contrastive learning, C3P aligns bacterial promoter DNA with representations from a frozen pretrained protein language model (ESM2), enabling a compact 100M promoter encoder to substantially outperform 7B generative genome language models in regulatory inference and zero-shot co-regulated gene retrieval.
- CausalBind: Causal Modeling and Learning for Protein-Molecule Virtual Screening
-
Addressing the issue that dual-tower virtual screening easily overfits to global spurious correlations, CausalBind models binder-only observations as a Heckman-style collider V-structure, proves that unconstrained dense interactions are non-identifiable while sparse antichain and low-rank interactions are identifiable, and achieves superior early-enrichment retrieval performance on DUD-E and debiased LIT-PCBA benchmarks via sparse concept-pair bottlenecks.
- CellMSA: Context Modeling for Single-Cell Representation Learning
-
CellMSA aligns cross-batch and related-type cells by gene identity, extracts gene-pair dependencies from low-dimensional context, and uses them to guide target-cell encoding, achieving strong results in label-informed integration, classification without test-label retrieval, and perturbation prediction combined with STATE-ST.
- Chem-PerturBridge: a harmonized compendium of small molecule perturbation transcriptomic effects
-
Constructs a harmonized perturbation compendium across 8 assay types encompassing over 37k compounds and 1.25M transcriptomic samples, systematically uncovering that cross-dataset logFC magnitude agreement is weak whereas regulation direction is highly stable, and demonstrating that pooled pretraining substantially benefits out-of-sample perturbation prediction.
- ColdDDI: Evaluating Knowledge Utilization in Cold-Start Drug-Drug Interaction Prediction
-
ColdDDI establishes a reconstructible diagnostic benchmark from DrugBank 5.1.13 by orthogonally decoupling cold-start splits (\(S_0/S_1/S_2\)) from pharmacological mechanisms (PK/PD) and knowledge graph mediator availability (Type A/B), revealing that inductive cold-start failure is driven by the absence of explicit KG bridges rather than drug novelty alone, and demonstrating through controlled perturbations and attention interventions that knowledge access does not imply knowledge utilization across 8 baselines and 13 LLMs.
- ConnectomeBench2: A Unified Benchmark for Automated Connectomic Proofreading
-
ConnectomeBench2 establishes a unified multi-species, multi-task benchmark of 716k expert edits and >4.5M images across four major open connectomes, demonstrating that a single Vision Transformer with shared encoders reaches human-level accuracy and well-calibrated triage in automated connectomic proofreading.
- CORAL: A Benchmark for Structure-aware and Brain-wide Neuron Reconstruction in Light Microscopy
-
Addressing the reliance of existing neuron reconstruction benchmarks on small local crops and geometric distance metrics, this paper presents CORAL, the first structure-aware, brain-wide benchmark featuring a root-to-leaf fiber metric and an iterative whole-brain tracing framework that exposes the dramatic degradation of current algorithms during long-range axonal tracking.
- CORAL: Learning Amyloid Fibril Ligand Docking with Cooperative Binding Rewards
-
CORAL is the first generative docking framework tailored to amyloid fibril–ligand complexes, combining explicit inter-chain periodicity priors with cooperative ligand–ligand stacking rewards in online reinforcement learning to overcome the severe scarcity of experimental co-crystal structures.
- CORTEG: Foundation Models Enable Cross-Modality Representation Transfer from Scalp to Intracranial Brain Recordings
-
CORTEG enables cross-modality neural transfer from large pretrained scalp-EEG foundation models to intracranial electrocorticography (ECoG) continuous decoding, achieving state-of-the-art regression performance and calibrating to held-out patients within 10–30 minutes on a single GPU.
- Data-Driven Soft Labeling Scales DNA Read Classification to Whole-Body Cell-Type Deconvolution
-
Syto replaces single-origin hard labels with the cell-type distribution associated with methylation patterns in reference data, then separately models read classification, sample deconvolution, and proportion calibration; its best reported configuration reduces pseudobulk MSE across 39 cell types from CelFiE's \(3.28\times10^{-4}\) to \(0.88\times10^{-4}\), without establishing clinical cancer-detection performance.
- Deep Generative Models for Phylogenetic Inference with Complex Evolutionary Processes
-
This paper introduces a family of conditional deep generative models over phylogenetic trees across three construction paradigms (bottom-up, middle-out, and top-down), combining multi-step iterative generation, Rao-Blackwellized closed-form discrete diffusion ELBO training, and a graph-conditioned covariance-pooling architecture to achieve accurate likelihood-free Bayesian posterior inference on up to 65 taxa and \(>10^5\) sequence tokens, substantially outperforming Phyloformer 2 on tractable models and surpassing IQ-Tree on realistic evolutionary processes with intractable likelihoods such as insertions and deletions (InDel).
- DEFT: Disentanglement-Enhanced Fine-Tuning for EEG Foundation Models
-
Addressing the challenge that EEG foundation models (EFMs) amplify subject-specific variations that interfere with downstream cross-subject classification, DEFT introduces a plug-and-play post-backbone adapter that factorizes mixed tokens into task-content and subject-style branches using recursive Gaussian inference and temporal-variance constraints, yielding consistent performance gains across four foundation backbones and multiple EEG tasks.
- DegradeQuery: Counterfactual Tuple Pretraining for Context-Aware PROTAC Degradation Prediction
-
DegradeQuery addresses the widespread scarcity of experimental degradation labels in public PROTAC databases by introducing counterfactual tuple pretraining, which learns rich contextual representations by contrasting observed molecule–target–E3 tuples with sampled biological replacements without assigning pseudo-labels, achieving state-of-the-art performance on the PROTAC-8K benchmark.
- Differentiable Retrieval-Augmented Generation for Predicting Cellular Responses to Gene Perturbation
-
PT-RAG introduces the first end-to-end differentiable, cell-type-aware retrieval-augmented generation framework for single-cell gene perturbation response, dynamically injecting relevant perturbation context to significantly boost population distributional fidelity and manifold geometry.
- Discovering Mechanistic Models of Neural Activity: System Identification in an in Silico Zebrafish
-
By benchmarking neural system identification in a transparent neuromechanical zebrafish simulation, this paper demonstrates that unconstrained autoregressive forecasting exploits statistical shortcuts, while integrating binary connectomic structural priors with LLM-guided tree search enables robust out-of-distribution generalization and faithful recovery of neural mechanisms.
- Drift-React: One-step Generation of Reaction Pathways via SE(3) Drifting Fields
-
Presents Drift-React, the first one-step generative framework that predicts complete, physically consistent minimum energy reaction pathways (MEP) and transition states (TS) in a single forward pass directly from reactant and product geometries via an SE(3)-equivariant Sinkhorn drifting field.
- Evolutionary foraging in grids: Intermittent search dynamics emerge in finite, depletable landscapes
-
On finite two-dimensional grids with non-renewable resources, the empirical distributions of step length, velocity, and turning angle evolve through a genetic algorithm rather than a prescribed power law; the resulting second and fourth displacement moments favor an intermittent-search description, without proving explicit state switching or globally optimal foraging performance.
- Fast Organic Crystal Structure Prediction with Unit Cell Flow Matching
-
Clari proposes a flow matching model that jointly models atomic Cartesian coordinates and lattice vectors directly on a single unit cell, discarding expensive bulk supercell crops and triangle-update layers to achieve a 15–30× sampling speedup over OXtal while enabling seconds-scale organic crystal structure prediction and practical high-throughput solid-state screening.
- Few-Step Cofolding with All-Atom Flow Maps
-
This paper introduces the Denoiser Cofolding All-atom Flowmap (DeCAF) framework, which distills pretrained diffusion models into two-time denoiser flow maps directly in \(\sigma\) noise space, and pairs them with lookahead-guided search (DeCAF-SEARCH) to generate physically valid, high-accuracy protein-ligand complexes in only a few inference steps.
- FlashMol: High-Quality Molecule Generation in as Few as Four Steps
-
Addressing the severe sampling bottleneck of 3D molecular diffusion models (requiring up to 1000 steps) and the stability collapse of existing few-step alternatives, FlashMol integrates geometric distribution matching distillation (DMD), EDM-inspired noise schedule respacing, and Jensen-Shannon divergence regularization to generate chemically stable 3D molecules in as few as 4 to 8 steps, achieving up to 250× acceleration over its 1000-step teacher.
- Flexible Flows for Biological Sequence Design
-
FlexFlow introduces a Discrete Flow Matching framework tailored for biological sequence design by combining domain-informed structured coupling, latent edit-based transition rates with test-time Dirichlet temperature scaling, and continuous latent classifier-free guidance, achieving state-of-the-art results across DNA enhancer/promoter and MHC-II peptide generation.
- FlyAOC: Evaluating Agentic Ontology Curation of Drosophila Scientific Knowledge Bases
-
FlyAOC evaluates fruit-fly knowledge base curation end to end, from large-scale full-text retrieval to ontology-grounded candidate outputs, showing that multi-agent context partitioning improves recall of known recoverable annotations while semantic scores do not establish exact biological validity.
- From Retrieval to Reasoning: Agentic Mechanism Prediction from Cell Painting Profiles
-
PhenoAIR reformulates Cell Painting-based drug mechanism of action (MOA) prediction from brittle representation matching into calibrated evidence reasoning, leveraging offline reliability calibration, role-separated agents, and a rule-based controller to achieve superior accuracy in both closed-set classification and open-set novel MOA discovery.
- Generative Molecular Morphing for Flexible-Size Design via Unbalanced Optimal Transport
-
The paper introduces Morph, a framework that formulates 3D molecular generation as a trans-dimensional jump-flow process coupled via unbalanced optimal transport, unlocking dynamic atom insertion/deletion and unconstrained flexible-size molecular design on 3D geometric graphs.
- gfnx: Fast and Scalable Library for Generative Flow Networks in JAX
-
gfnx is the first fully JAX-native open-source benchmark library for Generative Flow Networks (GFlowNets), removing CPU-GPU data transfer bottlenecks via end-to-end JIT compilation to achieve up to 80x wall-clock speedups across biological and discrete sampling domains.
- GLACIER: Rethinking Mass Spectrum Prediction as an Object Detection Problem
-
GLACIER reformulates tandem mass spectrum (MS/MS) prediction as a single-stage object detection task on molecular graphs, replacing two-stage heuristic fragment enumeration with end-to-end differentiable breakpoint detection and dynamic subgraph pooling to achieve an 8× inference speedup and record retrieval accuracy on NIST'20 and MassSpecGym.
- GOLD: Geometric Optimized Latent Diffusion for Structure-Aware RNA Inverse Folding
-
Addressing the atomic steric clashes caused by conventional RNA inverse folding models ignoring side-chain spatial orientations and topological constraints, GOLD projects discrete sequences and continuous glycosidic torsion angles into a unified SO(2)-equivariant latent manifold, combined with a backbone-conditioned DiT diffusion model and training-free analytical physical energy gradient guidance, achieving comprehensive state-of-the-art performance across sequence recovery and 3D structural foldability.
- HELICS: Biobank-scale Conditional Synthetic Genome Generation via Latent Flow Matching
-
Addressing the ultra-high dimensionality and phasing bottlenecks of modeling hundreds of thousands of discrete SNPs, HELICS introduces a phasing-free latent flow matching framework combining chromosome-specific convolutional autoencoders and a U-shaped interpolant transformer to achieve biobank-scale conditional genome synthesis.
- How Post-Training Shapes Biological Reasoning Models
-
This paper investigates how continued pre-training, supervised fine-tuning, and reinforcement learning shape the generalization of biological reasoning models across DNA, RNA, and protein domains, revealing that SFT causes out-of-domain collapse while RL restores robustness under an asymmetric capacity budget.
- How to make the most of your masked language model for protein engineering
-
This paper introduces a sequence-centric stochastic beam search for masked protein language models that leverages fast masked-marginal approximations to evaluate 1-edit sequence neighborhoods, enabling derivative-free multi-objective guidance and achieving a 100% in vitro therapeutic antibody campaign success rate.
- Interpreting Latent Protein Language Model Features with Geometric Annotations
-
This paper introduces an automated, scalable pipeline that interprets ESM-2 sparse autoencoder (SAE) latent representations at residue-level resolution using \(C_\alpha\) backbone geometric descriptors, resolving sub-domain motifs and transferring annotations to unannotated metagenomic sequences.
- Learning Implicit Bias in Generative Spaces for Accelerating Protein Dynamics Emulation
-
By augmenting a frozen SE(3) protein dynamics diffusion emulator with a distance-weighted, history-dependent implicit bias and environment-support regularization during reverse sampling, paired with test-time diffusion refinement, this work accelerates exploration of rare, low-energy metastable states by up to 37x while strictly preserving physical structural validity.
- LEMON-ZEST: Evolution-Informed Tokenization for Efficient Protein Language Modeling
-
LEMON-ZEST clusters evolutionarily conserved fragments into shared tokens and combines stochastic segmentation with a dual-head encoder for global retrieval and residue information, allowing a 200M-parameter model to outperform large baselines on several fold-level retrieval metrics without being best at every classification level or downstream task.
- Lifting Biomolecular Data Acquisition
-
LIFT lifts compressed sensing from finite-dimensional vectors to an infinite-dimensional neural function space, co-designing wet-lab multi-variant delivery (MOI > 1) with functional \(L^1\) regularized deconvolution during training, which provably scales effective dataset size by orders of magnitude and enhances wet-lab TCR-mimic antibody hit rates by 50%–100%.
- Looking Under the Streetlight: Evaluation in Generative Molecular Dynamics
-
Addressing the "streetlight effect" where generative molecular dynamics (GenMD) drifts toward convenient yet uninformative proxy metrics due to costly ground truth, this paper systematically analyzes six recurring misaligned evaluation patterns and establishes a tiered checklist alongside physical benchmark design principles.
- MolHIT: Advancing Molecular-Graph Generation with Hierarchical Discrete Diffusion Models
-
MolHIT introduces a hierarchical discrete diffusion framework with chemical priors and decoupled atom encoding, achieving near-perfect molecular validity (>99%) and establishing a new state-of-the-art across graph and sequence baselines.
- NEAT: Neighborhood-Guided, Efficient, Autoregressive Set Transformer for 3D Molecular Generation
-
NEAT introduces a novel neighborhood-guided training strategy that couples a bidirectional set transformer with a conditional flow matching head, achieving rigorous atom-level permutation invariance in autoregressive 3D molecular generation while delivering 5x to 34x inference speedups over leading diffusion and flow-matching baselines.
- On the Design Space of Discrete Diffusion Online Adaptation for Molecular Optimization
-
This paper systematically explores the design space of discrete diffusion online adaptation under limited oracle budgets, proposing an integrated recipe combining Thompson sampling, CVaR reward shaping, single-pass density entropy regularization, priority replay, and molecular validity penalties.
- On the Selectivity of Generative Models in Structure-Based Drug Design
-
Addressing the critical blind spot where generative structure-based drug design models focus solely on single-target affinity while ignoring off-target toxicity, this paper presents SelectBench—a multi-scenario benchmark with calibrated Z-score delta metrics showing that current generative architectures and inference-time guidance achieve only modest selectivity against close homologs and safety panels.
- PACE: Geometry-Aware Bridge Transport for Single-Cell Trajectory Inference
-
To tackle the fundamentally ill-posed trajectory inference problem from destructive single-cell time courses without cell correspondences, PACE constructs a state- and time-dependent anisotropic Riemannian metric penalizing normal velocity, alternates neural bridge fitting with optimal transport couplings under path-action cost, and distills the dynamics into a global continuous velocity field.
- pCoMole: Pareto-Constrained Molecule Editing with Discrete Flows
-
Addressing the challenge of satisfying hard biochemical and manufacturability constraints during variable-length sequence shrinkage and multi-objective optimization, this paper proposes pCoMole, a discrete flow matching framework with Doob's \(h\)-transform guidance that guarantees terminal feasibility while navigating Pareto trade-offs across GFP, Cas9, and peptidomimetic compression, verified by wet-lab retention of green fluorescence in 10-residue-shortened eGFP designs.
- PepSpecBench: A Unified Evaluation Benchmark for Peptide Tandem Mass Spectrometry Prediction
-
PepSpecBench establishes the first standardized, leakage-controlled benchmark for peptide tandem mass spectrometry prediction, aligning heterogeneous model outputs into a shared 234-dimensional canonical ion space, eliminating sequence memorization via backbone-disjoint splitting, and diagnosing model robustness through multi-species evaluations and physical parameter perturbation probes.
- PHOEBI: An Open-World Benchmark for Multi-Label Bacterial Identification in Phase-Contrast Microscopy
-
PHOEBI tests multi-label identification of known species in unseen mixtures using 40 combinations of six bacteria and approximately 120,000 phase-contrast images, revealing severe failures of per-image trained classifiers and greater stability of geometric prototype decoders on frozen features, although their gains must be interpreted against the all-present baseline and combination-level confidence intervals.
- PocketVE: Stable and Property-Guided Structure-Based Drug Design with Variance-Exploding Diffusion
-
PocketVE combines VE/EDM denoising that preserves the shared protein–ligand coordinate scale, training-time pocket perturbation, and inference-time multi-property CFG, raising 3D validity from 58.6% to 80.6% and reducing strain energy from 457.4 to 127.9 relative to guided TAGMol on CrossDocked2020, without attributing these gains to VE alone or claiming superiority on every docking metric.
- Preserving DEG Rankings for Gene Discovery in Histology-Based Spatial Gene Expression Prediction
-
The paper shifts histology-based spatial expression prediction from matching each gene's spatial profile to preserving which genes should be prioritized for a tissue contrast, using morphology-derived proxy groups, differentiable U statistics, and an across-gene correlation loss to improve DEG ranking and pathway overlap without guaranteeing better conventional spatial PCC.
- ProtGlycanDock: Towards Accurate Protein-Glycan Docking with Tailored Dataset, Benchmark and Model
-
Addressing the critical absence of specialized benchmarks and glycan conformational priors in molecular docking and all-atom structure prediction, this paper curates ProtGlycanDock—the first standard noncovalent benchmark with dual protein- and glycan-holdout evaluation splits—and introduces two complementary finetuning techniques (glycan topology feature injection and a differentiable Carbohydrate Intrinsic energy loss) that enable the open-source Protenix-constraint model to decisively outperform AlphaFold 3 in blind docking.
- Pushing Biomolecular Utility-Diversity Frontiers with Supergroup Relative Policy Optimization
-
Addressing the mode collapse induced by single-objective reward optimization in biomolecular RL, Supergroup Relative Policy Optimization (SGRPO) introduces same-condition supergroup sampling and leave-one-out marginal diversity redistribution, transforming set-level diversity into actionable rollout-level advantages to expand the utility-diversity Pareto frontier across small-molecule and protein design.
- RipplePLM: Structural and Property Decoupling for Protein Mutation Effect Generation
-
Addressing the issues of undifferentiated structural representations and disconnected biochemical property signals in protein point mutation effect generation, RipplePLM introduces Direct-Distal Cross-Attention (DDCA) and the Property Latent Chain (PLChain) to decouple structural context and biochemical properties at the representation level, substantially outperforming prior mutation-specific baselines on MutaDescribe.
- Robust Inference-Time Steering of Protein Diffusion Models via Embedding Optimization
-
Addressing the instability and geometric distortions of coordinate-space posterior sampling in low-density prior regimes, EmbedOpt directly optimizes conditional sequence and pair embeddings at inference time, delivering hyperparameter robustness across two orders of magnitude and a 4x reduction in diffusion steps.
- scTrilemma: Balancing Identity, Invariance, and Fidelity in Single-Cell Representation Learning
-
To reconcile the representation trilemma among biological identity, context invariance, and expression fidelity in label-free single-cell learning, scTrilemma introduces a latent-bottleneck VAE that routes expression-derived signals through expression-gated token encoding (E-Gate), decoder-query cell representation modulation (C-Route), and pseudo-bulk-conditioned priors (PB-Cond), outperforming foundation models across all three demands under a unified reconstruction objective.
- Self-Improvement Imitation with Biologically Guided Search for Protein Design Under Oracle Budgets
-
For protein sequence optimization under strict oracle evaluation budgets, SILO introduces a hierarchical edit policy, incremental stochastic beam search without replacement combined with alanine-scan fitness scoring for active candidate selection, and updates the policy via supervised next-action imitation on oracle-labeled top performers, eliminating value-function estimation under noisy proxies.
- Spectral Feedback for Test-Time Alignment of Protein Diffusion Models
-
Addressing the irreversible, unidirectional nature of discrete diffusion sampling in inverse protein folding, Spectral Feedback leverages the biological epistasis-induced Fourier sparsity of edit-set value functions to iteratively re-mask and resample optimal residues, achieving significant gains in protein stability.
- STRAND: Sequence-Conditioned Transport for Single-Cell Perturbations
-
STRAND introduces a generative perturbation modeling framework conditioning on regulatory DNA sequence representations and latent Schrödinger bridge diffusion, overcoming the resolution gap and closed-vocabulary limitations of gene-level models to achieve nucleotide-resolved, genome-wide zero-shot response prediction.
- Toward Better Geometric Representations for Molecule Generative Models
-
Addressing the non-smooth representation space and underutilized multi-layer semantics in geometric representation-conditioned 3D molecule generation, LENSEs introduces a learnable layer-pooled variational representation head, a molecule perceptual loss, and a node-level REPA loss to achieve a 4.6× smoother latent space and state-of-the-art generation validity.
- Towards Scalable Context-Aware Single-Cell Spatial Transcriptomics Prediction from Histology Images
-
CELLO shares one pathology foundation model forward pass per histology tile and predicts cell-level gene expression through continuous location querying and distance-decay cross-attention, improving average PCC on 52 paired H&E–Xenium samples while achieving a mean 14.0× speed-up over DeepSpot2Cell in timing that excludes upstream segmentation.
- Trustworthy Retrosynthesis: Eliminating Hallucinations with a Diverse Ensemble of Reaction Scorers
-
To eliminate dangerous hallucinations in generative retrosynthesis, RetroTrim pairs a root-aligned BART single-step generator with MetaScorer—an ensemble integrating a reaction prior (RP), a graph-attention plausibility classifier (RGP), and a reaction retrieval scorer (RRS)—completely eradicating hallucinated reactions and delivering the highest fraction of flawless synthetic pathways across 32 novel drug-like targets.
- URSA: Chemistry-Aware Benchmark for Utilitarian Retrosynthesis Assessment
-
Addressing the long-standing evaluation bottleneck between trivial stock termination and overly restrictive exact route reproduction, this paper proposes URSA, a standardized benchmarking suite powered by ChemCensor precedent checking that systematically assesses CASP algorithms and LLMs at the Solv-2 chemical plausibility boundary.
- Variable-Length Generative Protein Design via Generalized Poisson Flow
-
This paper introduces Generalized Poisson Flow (GPFlow), a variable-length generative framework that couples inhomogeneous generalized Poisson processes with within-length flow dynamics across continuous coordinates, discrete sequences, and Riemannian geometries, enabling simulation-free training and significantly outperforming fixed-length baselines across unconditional generation, motif scaffolding, and peptide co-design.