Nature Machine Intelligence

Predicting molecular receptor status, including estrogen receptor (ER), progesterone receptor (PR), and human epidermal growth factor receptor 2 (HER2), directly from routine hematoxylin and eosin (H&E)-stained whole-slide images (WSIs) could reduce reliance on costly immunohistochemistry, but image-only models have no access to the transcriptomic programs that define these subtypes. Here we show that coupling a whole-slide vision transformer (GigaPath) with bulk RNA-sequencing profiles through a training-time-only contrastive objective improves receptor-status classification from WSIs alone, without requiring any RNA-seq data at inference. Rather than training a dedicated bulk-transcriptomics encoder, we purpose the gene-sentence encoding and pretrained text encoder from BioBERT1 and apply them to bulk RNA-seq profiles from The Cancer Genome Atlas breast cancer cohort (TCGA-BRCA). Under five-fold cross-validation, this RNA-guided pre-training consistently improved AUROC for ER, PR, and HER2 relative to a contrastive-pre-training-free baseline and reproducibly introduced receptor-relevant structure into the learned slide representation. However, evaluation on independent external cohorts revealed a striking, receptor-dependent divergence: gains for ER and PR, both governed by broad, multi-gene luminal transcriptional programs, generalized robustly, whereas the apparent AUROC improvement for HER2, driven by focal ERBB2 amplification rather than a coordinated transcriptional signature, masked a severe loss of classification sensitivity at the operating threshold. These findings establish training-time transcriptomic supervision as a practical route to enhancing histology-only molecular subtyping, while showing that gene-selection strategy, and not model architecture alone, determines which molecular alterations such supervision can teach a vision model to recognize.

Transcriptomic classification is often hindered by the small number of samples relative to the high dimensionality of gene expression data. We introduce General Gene Expression (GGE), a deep meta-learning framework designed to support robust classification in this limited-sample setting. By training across 5,220 distinct biomedical prediction objectives drawn from 1,779 different human datasets, GGE learns a model initialization that captures biological patterns shared across heterogeneous classification tasks. This learned initialization has two key advantages. First, it enables improved performance to new datasets using only a small number of labeled samples. Second, because it is learned across diverse prediction objectives, it can be applied to a broad range of biomedical problems. We show that GGE outperforms established classifiers in data-limited settings across a wide range of applications, including datasets generated using different RNA-seq platforms and preprocessing pipelines. In addition, attention-based analysis identifies recurrent genes that contribute to performance across multiple biological objectives, providing insight into shared determinants of human biological states. Together, these results establish GGE is a general-purpose framework for human transcriptome-based classification in biomedical settings where labeled data are scarce.

Ribosome profiling is widely used to identify translated small open reading frames (sORFs), but existing prediction models often generalize poorly to newly collected sORF datasets. In addition, many microproteins with strong mass spectrometry (MS) evidence lack support in public Ribo-seq resources, suggesting that Ribo-supported sORF and MS-supported microprotein detection may represent complementary prediction tasks. Here, we developed sORF-Trans2MS, a two-module deep learning framework for predicting Ribo-supported sORFs and MS-supported microproteins. Module I uses RNA-FM representations for upstream, ORF, and downstream regions with a Transformer encoder for Ribo-supported sORF translation prediction. Module II integrates pretrained RNA-FM representations of RNA context with ESM-2 protein representations using multi-scale CNN encoders for predicting translated microproteins supported by MS but not Ribo-seq data. Module I was trained on 9,579 Ribo-supported sORFs and matched background sORFs, while Module II used 14,697 MS-supported (MS++Ribo-) microproteins and matched backgrounds. Module I achieved an AUROC of 0.870 internally and 0.930 on an independent GENCODE dataset. Ablation analysis highlighted upstream RNA context, including a distinct signal around 50 nt upstream of AUG. Module II achieved an AUROC/AUPR of 0.804/0.797 internally and 0.785/0.773 on an independent mass spectrometry dataset. Model interpretation highlighted upstream RNA context and the protein N terminus, while literature-supported microproteins further illustrated the complementary roles of the two modules. Overall, sORF-Trans2MS improves prediction and generalization of Ribo-supported sORFs and extends computational discovery to MS-supported candidates that are not represented in current public Ribo-seq resources.

Although protein toxins represent valuable pharmacological templates, predicting toxicity directly from primary sequences is inherently challenging because of their evolutionary dynamics. Active toxins and their benign homologues frequently share identical structural scaffolds and differ by only a few key residue substitutions, making global sequence similarity an unreliable indicator of function. Conventional predictors compound this challenge through flawed negative-set designs that sample non-toxic controls from arbitrary background proteins and reduce redundancy through identity-based clustering under random partitioning. This practice introduces global confounding shortcuts that inflate benchmark performance while failing to capture true functional boundaries. To address these limitations, we developed NeoToxPred, a sequence-only binary classifier that fine-tunes the pretrained ESM-C 600M protein language model end-to-end with a six-layer feed-forward classification head. Although the neural architecture is intentionally standard to ensure computational scalability, the core novelty of the framework lies in its rigorous training-data curation: (i) an orthologous negative set drawn from the same InterPro families under controlled taxonomic proximity to eliminate phylogenetic shortcuts, (ii) a length-stratified negative control set to eliminate sequence length as a predictive cue, and (iii) a strict family-disjoint splitting strategy to prevent performance inflation caused by pattern memorization. On held-out internal benchmarks, NeoToxPred achieved a Matthews correlation coefficient (MCC) of 0.897 and an F1-score of 0.949 on long sequences and an MCC of 0.869 on short peptides, decisively outperforming five recent predictors while maintaining stable performance across varying sequence lengths and evolutionary distances. In a genome-scale external validation using the previously unseen proteome of the redfin waspfish (Paracentropogon rubripinnis), NeoToxPred successfully recovered 11 of 16 empirically validated toxins without collapsing into a single-class prediction, whereas structure-dependent baselines identified substantially fewer positives. Computational alanine scanning further confirmed that the framework contextually localizes its predictions to functionally active residues. Taken together, these findings suggest that systematic negative-set design should be prioritized alongside algorithmic complexity in biological sequence classification.

Foundation models (FMs) have rapidly become dominant in artificial intelligence and are increasingly being adopted in computational pathology. Numerous pathology-specific FMs have been developed and evaluated for a variety of downstream tasks, most commonly using frozen image embeddings with linear probes. While a pathology FM, at least implicitly suggests broad reusability across tasks, the transferability of these models to neurodegenerative disease-related tasks remains largely unexplored. In this study, we evaluated fourteen frozen feature extractors, including general-vision and pathology FMs, across four pathology image classification datasets spanning two specialized neurodegenerative disease immuno-histochemistry (IHC) tasks (Tau neurofibrillary tangle (NFT) and amyloid-{beta} plaque classification) and two cancer-related hematoxylin and eosin (H&E) tasks (the breast tissue BACH dataset and the multi-tissue TIL dataset). The cancer-related H&E datasets represent tasks more closely aligned with the predominant pretraining domain of many current pathology FMs, whereas the neurodegenerative IHC datasets represent a more specialized domain with limited representation in existing FM pretraining cohorts. Frozen embedding linear probes were compared against a conventional supervised ResNet-50 convolutional neural network (CNN) trained directly on image tiles using identical train, validation, and test splits. Across both neurodegenerative IHC datasets, the supervised CNN substantially outperformed all frozen FM linear probes. In contrast, pathology FMs achieved performance comparable to, and in some cases exceeding, the supervised CNN across the two cancer-related H&E datasets, with UNI2-h achieving the highest performance on BACH and several pathology FMs performing on par with the CNN on the larger TIL dataset. Furthermore, a supervised CNN trained using only 1% of the Tau NFT training data (1,938 tiles) still exceeded the performance of the best frozen FM linear probe trained on the complete dataset. Together, these results suggest the transferability of frozen pathology FMs may depend strongly on how well the downstream task is represented by their pretraining domain. These findings demonstrate frozen pathology FMs transfer effectively to the cancer-related H&E tasks evaluated here but may be less effective for specialized neurodegenerative IHC tasks less well represented in current FM pretraining cohorts. Expanding pathology FM pretraining datasets to include a broader range of disease domains and staining modalities may therefore improve transferability to specialized pathology applications. Our findings also highlight the continued importance of conventional supervised learning and motivate future work investigating nonlinear probes and end-to-end FM fine-tuning.

Molecular structures are abundant, while experimental bioactivity is sparse and distributed across assays. We introduce GRASP, a 93.5M-parameter graph Transformer that learns from these sources in sequence. GRASP first learns molecular structure through replaced-token detection on 1.54 billion ZINC20 molecule presentations, then adapts to sparse bioactivity across 642 ChEMBL assays. Across 23 cluster-held-out OpenADMET endpoints, GRASP achieves a mean MAE of 0.374, compared with 0.383 for CheMeleon, 0.408 for ECFP4-LightGBM, and 0.440 for Chemprop. ChEMBL adaptation reduces mean MAE from 0.407 to 0.374 and improves 21 of 23 endpoints, showing the value of intermediate bioactivity supervision. On 22 TDC ADMET tasks, GRASP gives a better point estimate than CheMeleon on 13 tasks. These results support staged molecular representation learning, where abundant molecular structure and sparse experimental measurements provide complementary supervision at different stages of training.

Longitudinal transcriptomic studies often capture only a few molecular snapshots of each patient over time. This makes it challenging to understand how patients are progressing, compare them when progression rates differ, and capture molecular changes between observed visits. We introduce CohortFM, a flow-matching framework that reconstructs continuous patient-specific molecular trajectories from sparse longitudinal observations and organizes them through an integrated mixture model into a compact set of cohort-level progression patterns. CohortFM separates molecular progression from calendar time, places patients on a shared progression scale, uses observed intermediate states to guide the learned trajectories, and maps the resulting patterns back to real patient samples for biological interpretation. We evaluated CohortFM across COVID-19, sepsis, influenza, and tuberculosis. Despite different sampling patterns, the method generalizes well and consistently produces meaningful trajectories. These trajectories capture biological differences between patients that cannot be explained by the observed samples alone. Compared with state-of-the-art trajectory baselines, CohortFM produces more biologically distinct trajectories and a more complete representation of the patient population. Its finer progression scale also reveals biological programs and temporal dynamics that are not captured by visit-level analysis. We further evaluated interpolation and true forward prediction to test whether CohortFM can accurately recover unobserved states. CohortFM remains competitive with strong interpolation baselines and is among the strongest learned predictors in the harder forecasting setting. Ablation experiments further show that flow matching is the core learning mechanism and that our design choices improve continuous trajectory learning. Our results support CohortFM as a practical framework for turning sparse molecular snapshots into continuous, biologically interpretable trajectories. By going beyond visit labels, CohortFM makes it possible to study disease progression and predict future molecular states even when patients are sampled only a few times.

Single-cell RNA sequencing (scRNA-seq) has transformed our understanding of cellular heterogeneity, yet most datasets are limited to a handful of model organisms, leaving critical gaps in cross-species biology. Existing computational methods for integrating multi-species scRNA-seq data often rely on one-to-one ortholog mapping, which fails to account for gene duplications, losses, or functional divergences. To address these challenges, we introduce CycloCross, an adversarial method based on the CycleGAN architecture, designed to translate scRNA-seq data between species without requiring a priori gene correspondences. Unlike generative models that synthesize data from noise, CycloCross leverages real data from a source species, enabling more realistic and data-efficient translations. We demonstrate its effectiveness in germ cell development, translating data between mouse, macaque, and human with conserved yet divergent transcriptional programs. CycloCross not only preserves temporal dynamics but also enables interpolation and extrapolation of gene expression at missing time points, offering predictions for experimentally inaccessible states. Furthermore, CycloCross generates biologically plausible samples even with limited data, outperforming traditional generative models.

Cardiomyocyte maturation spans embryogenesis through postnatal life and involves coordinated transcriptional transitions that are disrupted in disease and incompletely recapitulated by stem cell-derived cardiomyocytes. However, no unified framework exists for assigning absolute transcriptional maturation age across datasets, platforms, and experimental conditions. Here, we developed CMAC (Cardiomyocyte Maturation Age Clock), integrating 18 murine single-cell RNA-sequencing datasets spanning embryonic and postnatal timepoints, into a batch-invariant probabilistic atlas using single cell variational inference (scVI) with a 100-dimensional latent representation. Analysis of the developmental trajectory revealed highly non-uniform transcriptional change, with 63.9% of the total transcriptional journey completed by birth and the neonatal transition representing the largest single transcriptional step. An XGBoost regressor trained on the CMAC latent space predicted absolute chronological age with a leave-one-dataset-out mean absolute error (MAE) of 4.09 days (median 3.27 days), Pearson r = 0.903, and Spearman {rho} = 0.908 across 18 datasets and six sequencing platforms, outperforming linear regression (MAE = 6.18 days) and PCA-based representations (MAE = 6.35 days). In an independent P14 cohort comprising 98,163 cardiomyocytes from six biological replicates, CMAC achieved a mouse-level MAE of 0.33 days, with all six mice predicted within 1.1 days of chronological age. Application of CMAC to a previously published PGC1/{beta} double-knockout model with established delayed cardiomyocyte maturation independently recapitulated this phenotype, quantifying maturation deficits of 4.77 to 8.60 days across P7 to P28, with the largest deficit at P14. Together, these results establish CMAC as a quantitative, deep learning framework for assigning absolute transcriptional maturation age and quantifying deviations from normal cardiomyocyte development.

Large-scale Neuropixels recordings sample many brain areas, but no single experimental session records every area of interest. NeuroPaint, a transformer-based autoencoder, integrates such recordings by aligning area-specific latent dynamics across sessions and "inpainting" activity in unrecorded areas. Could a linear model, which would be computationally efficient and potentially more interpretable, suffice to achieve this? To address this question, we introduce Linear NeuroPaint, a linear variant of NeuroPaint, and evaluate it on two multi-area Neuropixels datasets. Linear NeuroPaint achieves strong predictive performance on recorded areas, however, it is less accurate than nonlinear models at cross-area inpainting, despite extensive hyperparameter optimization. Moreover, its parameters are not readily interpretable: incomplete cross-session alignment leads parameter-defined neuron clusters to exhibit heterogeneous response profiles. Together, these results show that linear transformations are insufficient for cross-session alignment and cross-area prediction. Nevertheless, Linear NeuroPaint provides a computationally efficient baseline for benchmarking multi-session neural models.

This paper predicts antimicrobial resistance (AMR) from matrix-assisted laser desorption/ionization time-of-flight (MALDI-TOF) mass spectra using DRIAMS, evaluated across 13 species-antibiotic datasets spanning sensitive/resistant imbalance ratios of 2.95:1 to 141.21:1 and raw spectral similarities of 0.9166-0.9946. We propose HetSAGE, a heterogeneous graph neural network (GNN) that combines full-spectrum "raw" edges with consensus "biomarker" edges derived from a 4-method feature-selection vote and applies per-edge-type dropout to regularize the two views separately. HetSAGE outperforms a plain multilayer perceptron (MLP) only when spectral similarity is high, consistent with oversmoothing effects reported in the broader GNN literature; the biomarker view converges on features close to independently established clinical markers: within {+/-}1 Da in Staphylococcus aureus, Oxacillin, and within range in Klebsiella pneumoniae, Meropenem, confirming it is not a black box. At the same time, per-edge-type dropout protects this smaller signal from dilution by regularization tuned for the noisier raw spectrum. Together, these results indicate that heterogeneous graph structure is not a universal win for MALDI-TOF AMR prediction but a conditional one, tied to a measurable property of the data, spectral similarity, rather than to architecture alone. This caveat is largely unaddressed in current DRIAMS-based GNN work. It provides a concrete signal for practitioners deciding when a graph model is worth the added complexity over a plain MLP.

Focal cortical dysplasia (FCD) is among the leading causes of drug-resistant focal epilepsy and can be challenging to detect on routine MRI due to the subtle and heterogeneous appearance of lesions. Motivated by the scarcity of clinically representative cohorts, we investigate the feasibility of simulating FCD-like lesions in healthy MRI using a composition of hand-crafted image transformations. We develop SynthFCD, a phenomenological FCD simulator, and create 2,164 synthetic examples from healthy controls in the heterogeneous FOMO300k dataset. We then pretrain deep learning (DL) segmentation models on synthetic lesions and fine-tune them on real FCD cases from the University Hospital Bonn (UHB) cohort. Comprehensive evaluations show consistent improvements across simulator presets, with detection-rate gains of up to 18.5 percentage points over training from scratch and 7.4 points over self-supervised pretraining, while achieving state-of-the-art segmentation performance (0.366 DSC) on the challenging held-out UHB benchmark. These findings suggest that simulation-based synthetic data generation can bridge existing radiological knowledge and advances in DL to provide meaningful supervisory signals that could ultimately aid the detection of underrepresented and diagnostically challenging cases.

ICD coding is the backbone of medical-record structuring, insurance payment, and clinical research. Today's best automatic coders are deep models, but they assume tens of thousands of labeled records, GPU servers, and machine-learning engineers. Many hospitals, especially small ones, have none of these. We study Chinese electronic medical records (EMRs) under a hard budget: 2,200 labeled records, CPU only, and fully interpretable decisions. Our method is lightweight and follows the way human coders work. First, codebook knowledge injection marks standard disease and operation names in the text as useful clues. Second, multi-view field weighting gives more weight to the clinical sections that matter most. Third, character n-gram TF-IDF features feed linear classifiers that score each candidate code. Fourth, a small set of clinical rules acts as a final safety check. Our main methodological finding is about this last step: naively applied rules do more harm than good (overall F1 drops from 0.8070 to 0.7995, and main-operation accuracy from 0.9727 to 0.9432). We therefore audit each rule on held-out data by counting how often it fixes a wrong prediction or breaks a correct one, and we remove rules that do more harm than good. This turns rule post-processing into a safety check that never hurts. On the CCL2026-Eval Task 6 dataset, our method reaches macro-F1 of 0.899 (main diagnosis) and 0.966 (main operation) in 5-fold cross-validation. A frequency-stratified analysis shows a clear long-tail limit: codes seen fewer than three times score zero F1. In a head-to-head test, a fine-tuned Chinese pretrained encoder scores lower than our method on all four tasks at this training size, while needing a GPU. Ablation tests show that the gains come mainly from field-weighted text. The framework trains in about 40 minutes on an ordinary CPU, and every prediction can be broken down into interpretable feature contributions.

Understanding the effects of genetic, molecular, and experimental perturbations is essential for decoding cellular mechanisms and guiding biomedical interventions. Existing computational approaches are typically designed for individual tasks, such as predicting perturbation responses, characterizing downstream transcriptional effects, or modeling perturbation combinations, and therefore do not provide a unified framework for learning interpretable perturbation representations while enabling causal analysis of downstream effects and characterization of condition-dependent responses. We present SHERLOCK, 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. % By formulating perturbations as interventions within a structural causal model, it enables counterfactual estimation of their downstream transcriptional effects under explicit identifiability assumptions. The same framework quantifies how perturbation responses vary across conditions and compositionally models combinatorial perturbations, enabling prediction of held-out combinations and classification of genetic interactions. Across genome-scale CRISPR, chemical, and spatial perturbation datasets, SHERLOCK recovers perturbation relationships concordant with known biological pathways and pharmacological properties, identifies condition-dependent responses, and predicts combinatorial perturbation effects. Together, SHERLOCK provides a unified framework for interpretable and causal analysis of perturbation effects across diverse single-cell perturbation experiments.

In-context learning (iCL) is an emergent capability of Large Language Models (LLMs), allowing them to perform new tasks at inference time using prompt-injected examples. While extensively studied in LLMs, the boundaries of its capabilities and underlying mechanisms remain poorly characterised in genomic Language Models (gLMs). Here, we map the operating regime of Evo 2, a nucleotide-level foundation gLM, across five binary classification tasks spanning biological and artificial sequences. We observe robust iCL on shorter natural sequences (F1=0.902 for miRNA, 0.785 for Toxins), degrading with sequence length, and collapsing at kilobase scale. Strikingly, we find no benefit in model scaling, as the 7B model systematically outperforms the 40B variant. We further find that perplexity, a widely used gLM performance proxy, poorly predicts accuracy. Mechanistic interpretability reconciles these observations: the logit-lens profiling suggests a prediction-generalisation trade-off, while Jacobian Scope indicates models might track prompts' structure rather than the signal-carrying content.

Single-cell perturbation response prediction seeks to recover transcriptional responses under unseen perturbation conditions. Because RNA sequencing measurements are destructive, control and perturbed cells are typically observed as unpaired populations without cell-level correspondence, shifting the prediction objective from individual cellular outcomes to perturbation-specific population distributions. The Schr"odinger Bridge (SB) provides a principled framework for modeling these population-level transitions as stochastic transport, but learning bridge dynamics from high-dimensional, sparse gene-expression profiles remains challenging. We propose PerturbBridge, a conditional latent Schr\"odinger Bridge framework that reformulates stochastic population transport over high-dimensional, sparse gene-expression profiles as bridge learning in a compact cell latent space. This formulation enables an efficient approximation of SB-based stochastic population transport in single-cell perturbation prediction. PerturbBridge first learns a compact latent representation following the \beta-TCVAE paradigm \citep{chen2018isolating}, alleviating the challenges caused by high dimensionality and sparsity during bridge learning. An interpolation-consistency regularizer further encourages agreement between decoded latent interpolations and the corresponding expression-space interpolations. PerturbBridge then learns a perturbation-conditioned latent bridge using a tractable stochastic SB approximation between latent control and target populations, with endpoint MMD regularization promoting alignment between generated and observed target distributions. Experiments on the Norman CRISPR and Sci-Plex3 chemical perturbation benchmarks demonstrate competitive performance across all evaluation metrics. Notably, PerturbBridge achieves state-of-the-art performance in differential-expression recovery on both benchmarks, highlighting the effectiveness of latent stochastic transport modeling for population-level single-cell perturbation prediction.

Membrane permeability remains a major challenge in cyclic-peptide drug discovery, where optimization often depends on estimating changes relative to measured reference compounds. Existing predictors mainly model absolute permeability, while heterogeneity across literature sources complicates pooled training and comparative evaluation. We present CycDelta, a multimodal residue-graph model trained directly on within-source permeability differences. CycDelta combines directed message passing with pretrained Uni-Mol monomer embeddings, physicochemical descriptors and assay information in a shared encoder for reference and query peptides. To evaluate this comparative formulation under controlled conditions, we establish CycDeltaBench, comprising in-distribution, scaffold-based out-of-distribution, external and zero-shot unseen-assay tests. On the out-of-distribution task, CycDelta achieved Pearson 0.69 versus 0.46 for the strongest baseline. It also led all evaluated methods on two external data sets and achieved zero-shot Spearman 0.70 versus 0.32 for the strongest baseline on an unseen assay. Residue-level attribution relates predictions to monomer contributions, while gradient-guided screening supports prioritization of permeability-enhancing analogue hypotheses. These results position CycDelta as a transferable and interpretable framework for comparative permeability prediction and cyclic-peptide optimization.

Text-driven medical image segmentation demands unified models based on multi-modal learning, capable of processing diverse imaging modalities while providing spatially reliable prediction estimates for clinical practice. Nevertheless, existing vision-language model solutions suffer from poor domain generalization and exhibit limited capability for uncertainty estimation to model the cumulative unreliability of cross-modal alignment across network layers. In this paper, we propose UniMedSeg, a unified vision-language framework for multi-modal medical image segmentation built upon frozen pre-trained vision-language backbones. Our framework incorporates modality-aware adaptation, cross-layer uncertainty-guided encoding, and a frequency-spatial dual-prompt decoder, wherein the progressive uncertainty propagation mechanism of the encoder suppresses misleading cross-modal features, whereas the decoder disentangles anatomical structure and boundary details in the frequency domain under dual textual prompts. Auxiliary losses further sharpen boundary representations and calibrate pixel-wise uncertainty maps. Extensive experiments demonstrate that UniMedSeg preserves competitive in-domain segmentation performance, improves domain generalization, and yields clinically useful spatial risk estimation.

The pharmacokinetics concentration-time profile encodes ADME dynamics. Its optimization in drug discovery curbs late-stage attrition, yet forecasting human pharmacokinetics from chemical structure remains difficult. Machine learning models ignore continuous ADME dynamics, whereas physiologically based pharmacokinetics (PBPK) models demand unavailable parameters and impose rigid, misspecified compartmental structures. We present PK-MUSE (PharmacoKinetic hybrid Mechanistic Universal System Equation), augmenting a two-compartment model with bounded, state- and time-dependent neural corrections to elimination and intercompartmental distribution. Molecular structure parameterizes compound-specific behavior, while learned corrections refine dynamic equations. Across scaffold and temporal out-of-distribution evaluations, PK-MUSE achieved the lowest mean exposure-endpoint MdFE, reducing MdFE by 7.2%, 45.2%, and 26.2% versus state-of-the-art machine learning (HML- RF), deep learning compartmental (DeepCt), and PBPK (PK-Sim) models. Interrogating the corrections localized misspecification to distribution rather than elimination. By extending mechanistic simulators with data-driven constrained neural dynamics that preserve interpretability and generalize to unseen chemotypes, PK-MUSE advances digital twins for drug discovery.

Recent deep learning (DL) models predict average gene expression levels from DNA sequences with high overall correlation to measured values. We examine these DL models through the lens of disease research. The seemingly high overall performance of DL models is largely due to capturing whether genes are "On" or "Off", and to a lesser extent, disease-relevant expression level changes for "On" genes. Indeed, the more bimodal the gene expression distribution, the better the reported performance. We track the extent of this issue across tissues, molecular systems, and cancers. Compounding the model evaluation issue, we find inconsistencies between the published code and reported performance, highlighting the importance of versioning and publishing performance evaluation code. These findings indicate that the high reported performance of popular DL models falls unexpectedly short in disease applications, and that the problem of personalized genomic prediction remains far from solved in a disease context.