Publications, newest first.

Sep 2

Mapping the combinatorial coding between olfactory receptors and perception with deep learning

The MolOR team combined molecular graphs with protein-language-model representations to predict which smell receptors an odorant activates, then used those predictions to improve estimates of perceived odor. Connecting chemical structure, receptor biology and perception could guide the search for ligands for poorly understood receptors and the design of new odorants. The predictions do not by themselves establish how a new molecule will smell to people.

Sep 2

Multi-objective optimization in the context of generative chemistry

A Nature Communications perspective examines how generative drug-design systems balance competing goals such as potency, pharmacokinetics and safety. Its practical message is that choosing and evaluating those goals deserves as much attention as generating more molecules, particularly when predicted properties can be misleading. This is a methodological perspective, not a new experimental demonstration of better drug candidates.

• AI x biology
• generative chemistry
• drug discovery
• machine learning

Sep 2

Unsupervised discovery of functional sequence patterns from protein language model with MotifAE

Columbia researchers built MotifAE to extract interpretable sequence motifs and domain features from protein language models, linking some of those features to experimental measurements of folding stability. The method could help protein engineers understand which patterns a model uses when predicting function or the effects of mutations. This is evidence for a research tool, not proof that its predictions generalize to every protein family.

Sep 2

AI models for medication adherence prediction: closing the gap to clinical readiness

A review of 41 studies found that most AI models for predicting medication adherence had serious development-quality concerns, while 80% were at high risk of bias in evaluation. The findings suggest that clearer outcomes, better handling of missing data and stronger validation matter more than simply choosing a more complex algorithm. This is an assessment of existing research, not a validated tool for deciding which patients will miss medication.

Sep 1

From sequences to strategies: Early detection of new SARS-CoV-2 variants via genetic distance to reduce hospitalizations

Researchers combined genetic-distance clustering with a neural classifier to study emerging SARS-CoV-2 variants using sequences from six European countries between 2020 and early 2024. The framework could help connect genomic surveillance with estimates of hospitalization risk and the effects of vaccine matching or containment measures. The reported reductions in hospitalization burden come from modelled scenarios, not observed causal effects of an intervention.

Sep 1

Generative deep learning reconstructs subcortical neural signals from cortical recordings for closed-loop brain stimulation

A generative model reconstructed deep-brain signals from cortical recordings, with testing across 723 hours of data from 49 patients at three centres. The approach could help adaptive brain-stimulation systems keep tracking useful biomarkers when implanted sensing is disrupted or unavailable. Signal reconstruction and state detection are not the same as demonstrating better treatment outcomes in a prospective trial.

Sep 1

Condition controllable generation of 3D molecules using textual prompts by multimodal equivariant diffusion model

Shanghai Jiao Tong researchers introduced TDmol, a diffusion model that links written molecular descriptions with two- and three-dimensional structures to guide the generation of new molecules. Text-based design constraints could give drug-discovery teams a more flexible way to explore chemical structures with desired properties. The abstract reports computational improvements, not experimental confirmation that the generated molecules are effective medicines.

Sep 1

Deep learning from retinal images for prediction of biomarker-informed obesity subtypes associated with cardiometabolic risk

Researchers developed ROSA to infer biomarker-defined obesity subtypes from a retinal photograph, studying 63,693 people across UK and Chinese cohorts. A scalable imaging readout could help characterize metabolic differences that body mass index misses, but external performance was modest: macro-AUROC was 0.66 and sensitivity 0.45. Those limitations make this a research finding rather than a ready-to-use clinical screening test.

Sep 1

A machine learning method for calculating highly localized protein stabilities

PFNet uses machine learning to estimate local protein stability from hydrogen-exchange mass-spectrometry measurements that are normally resolved at the peptide level. Extracting residue-level stability from a familiar assay could make protein dynamics easier to study at scale and give engineering teams a more detailed view of their designs. The accessible abstract does not give error estimates or enough benchmark detail to judge performance across all proteins and complexes.

Sep 1

A perspective on federated foundation models in biomedical sensing and imaging

A perspective from University at Buffalo researchers outlines how federated, multimodal foundation models could learn across distributed biomedical imaging and sensing datasets. It provides a research agenda for collaboration without centralizing all patient data, including proposed benchmarks and unresolved implementation challenges. This is a perspective, not a new experimental demonstration of improved privacy or clinical performance.

Aug 31

mmVelo: A deep generative model for estimating cell state-dependent dynamics across multiple modalities

mmVelo uses a generative model to infer changes in cell state from RNA measurements and project those dynamics into chromatin accessibility, with demonstrations in mouse tissues and human brain data. This could help researchers connect changing gene expression with regulation when not every biological layer has been measured. The inferred trajectories are model estimates, not direct time-lapse observations of each cell.

Aug 31

Artificial intelligence tools in sepsis prediction: a systematic review and meta-analysis

A review of 53 studies found that machine-learning sepsis models generally discriminated better than traditional comparators, but only about a third of positive alerts represented sepsis in the pooled analysis. The contrast matters for deployment because apparently strong accuracy can still produce substantial alarm fatigue. Marked differences between studies and widespread bias leave improvements in patient outcomes unproven.

Aug 27

Finetuning Foundation Models for Temporal Clinical Transcriptomics Data

Mathur and colleagues fine-tuned foundation-model gene embeddings and used temporal graph neural networks to distinguish treatment-response patterns across four inflammatory diseases. The work tests whether biological foundation models can help analyze small, noisy longitudinal clinical datasets rather than only static reference atlases. The publisher abstract does not report cohort sizes or complete benchmark metrics.

• Peer reviewed
• Foundation models
• Data & benchmarks

Aug 26

DeepPathway: Predicting Pathway Expression from Histopathology Images

Ahsan and colleagues trained a contrastive learning model on spatial transcriptomics and tissue images to predict pathway-level expression from routine histology. Pathway-level inference could make spatial biological readouts more accessible where direct spatial transcriptomics remains too expensive. Predicted pathway activity is not a substitute for direct transcriptomic measurement.

• Peer reviewed
• Single-cell & spatial
• Data & benchmarks

Aug 26

Deciphering the comprehensive relationship between 5′ UTR and 3′ UTR sequences with deep learning

Suga and colleagues combined pretrained RNA-language-model representations with contrastive learning to identify coordinated relationships between 5′ and 3′ untranslated regions. Modeling both untranslated regions together could inform future attempts to optimize therapeutic mRNAs. The results are computational associations and do not establish that a specific UTR pair will improve a therapeutic mRNA.

• Peer reviewed
• Foundation models
• Drug discovery

Aug 26

Coala: a standard-based framework for converting CWL-described command-line tools into agentic toolsets

Hu and colleagues built Coala to expose Common Workflow Language tools to language-model agents through the Model Context Protocol. Standards-based tool execution could make scientific agents easier to audit, reproduce, and extend. The accepted-manuscript abstract does not report a comparative user study.

• Peer reviewed
• AI agents
• Data & benchmarks

Aug 25

Predictive design of tissue-specific mammalian enhancers that function in the mouse embryo

Chen and colleagues used deep learning to design 15 synthetic enhancers that activated the intended tissues in mouse embryos. The study moves AI-guided regulatory-DNA design from cell systems into a mammalian in-vivo test. The experiments covered three tissues at one embryonic stage.

• Peer reviewed
• Protein design
• Data & benchmarks

Aug 24

Adaptive model-guided protein evolution with sparse data optimizes compact eukaryotic genome editors

A sparse-data machine-learning loop helped researchers engineer a compact Fanzor genome editor that worked across mammalian-cell targets and in a mouse experiment. The study shows how language-model and structure priors can narrow a large protein search space when experimental data are scarce. The workflow does not explicitly model higher-order interactions among multiple mutations.

• Peer reviewed
• Protein design
• Data & benchmarks

Aug 22

GRASSP: RNA Language Model–Enhanced Graph Attention with Adaptive Gating for RNA–Small Molecule Binding Site Prediction

GRASSP combines pretrained RNA representations with lightweight graph refinement for RNA-small molecule binding-site prediction. RNA-targeted discovery needs models that combine sequence-scale priors with local structural context without a heavy structural pipeline. Results are benchmark-based rather than prospective experimental validation.

• Peer reviewed
• Foundation models
• Drug discovery
• Data & benchmarks

Aug 22

Delays between CE mark and FDA regulatory approval of AI-enabled software for radiology

A cross-jurisdiction analysis finds longer second-authorization delays when radiology AI reaches Europe before the United States. Regulatory sequence and evidence transparency affect when medical-AI products become available across jurisdictions. The observational analysis cannot show why an individual authorization was delayed.

• Peer reviewed
• Regulation
• Data & benchmarks

Aug 19

Prospective evaluation of a large language model clinical decision support system in the emergency department

A four-week emergency-department evaluation found mostly appropriate sampled outputs but falling clinician adoption and no reduction in length of stay. The study adds prospective workflow evidence to a field dominated by retrospective accuracy tests and shows that technically acceptable output does not guarantee sustained clinical use. This was a stage 1 prospective evaluation, not a randomized efficacy trial.

• Peer reviewed
• AI agents
• Regulation
• Data & benchmarks