Molecular Systems Biology

Metabolic profiles are used to quantify disturbance, classify disease and infer biochemical function, yet these endpoints retain different information. We combine secondary analysis of 386 metabolites in six Arabidopsis thaliana mutation accumulation lines, known-disease classification in 129 public human metabolic-disease profiles, and an analytically solvable kinetic model. A conditional occupancy model separates annotation-bin breadth from hit count and measured bin sizes: the plant low-versus-high fitness comparison changes from a two-sided exact permutation p = 0.10 to p = 0.30 after calibration. Unsigned metabolite profiles achieve mean macro recall of 0.911 and 0.895 in separate EDTA and heparin subsets, compared with 0.624 and 0.382 for a burden baseline. However, controlled loss of 50% of held-out measurements reverses representation ranking in heparin, giving signed profiles a recall advantage of 0.108 with a conditional bootstrap interval of 0.063-0.155. A spectral decomposition identifies the within-pathway contrasts attenuated by smoothing. The kinetic model then shows that lower concentration and faster relaxation can coexist with reduced desired product flux, while a null-space criterion states when observations identify a metabolic task. The results connect assay coverage, diagnostic information and functional identifiability. Human performance is internally validated, and the kinetic predictions remain uncalibrated to patients.

Enzyme kinetic parameters inform metabolic models, yet experimental measurements are sparse. A growing body of work predicts them from protein and substrate features, but software fragmentation hinders adoption, so downstream tools lock into the most accessible method. We present OpenKinetics Predictor (at predictor.openkinetics.org), an open-source platform integrating thirteen methods in isolated environments behind one interface. The platform optionally reports similarity between query proteins and each method's training data to contextualise reliability. A common featurisation-prediction abstraction keeps it extensible, and independent parties, including original authors, contributed many methods. We pair it with a data portal (at data.openkinetics.org) that exposes CatLog, a curated kinetic dataset, with precomputed embeddings, predicted binding sites, and standardised splits. Both offer a web interface and an API, and the GECKO modelling toolbox already calls the predictor API. As a case study, we predict across an E. coli model and find inter-predictor agreement varies with metabolic context and data availability.

The rapid growth of omics studies offers new opportunities to uncover disease mechanisms, yet findings from individual studies or single omics may be influenced by cohort-, platform- and layer-specific variation. Integrating evidence across studies and omics layers can reveal robust biological signals that are reproducible across molecular layers, but differences in assay platforms, sample sizes and feature coverage complicate direct combination of summary statistics. Here we present the Omics Rank-Based Integration Tool (ORBIT), a direction-aware framework for prioritizing concordant signals across omics layers using only within-dataset feature ranks and effect directions. ORBIT tests directional rank consistency across omics layers against a closed-form variance-gamma null distribution while accounting for inter-dataset correlation and incomplete feature overlap. Simulations show that ORBIT maintains well-calibrated false-positive rates under inter-omics correlation, increasing numbers of omics layers and missingness, while detecting concordant signals with increasing power as the number of layers grows. We applied ORBIT to tubulointerstitial transcriptomic data from patients with chronic kidney disease in the C-PROBE cohort and proteomic data from the Kidney Precision Medicine Project, identifying a concordant injury programme marked by inflammatory and fibrotic remodeling together with impaired energy metabolism. We also used ORBIT to integrate ten transcriptomic, four proteomic and one translatomic datasets in dilated cardiomyopathy heart tissue, identifying extracellular matrix remodeling and mitochondrial energy failure as the dominant concordant biological programmes. In both diseases, ORBIT prioritized concordant genes and identified pathways that were not significant in any single omics layer alone. These results show that directional rank integration can be used for integrating heterogeneous omics summary data and prioritizing robust molecular mechanisms.

Objective Single-cell RNA sequencing has transformed our understanding of cellular heterogeneity within osteoarthritic (OA) joint tissues, yet transcript abundance provides only an indirect measure of cellular phenotype and function. Recent advances in mass spectrometry now enable proteins to be measured directly at single-cell resolution. Here, we applied single-cell proteomics (SCP) to primary human OA cartilage and synovium to determine whether cellular heterogeneity within the diseased joint can be resolved directly at the proteome level. Design Cartilage and synovium were obtained from patients undergoing total knee arthroplasty for OA. Low-input proteomic profiling was first performed to establish proteome depth from limited cell numbers and validate biologically relevant clustering. SCP was then applied to individual cells, followed by unsupervised clustering, differential protein abundance analysis and iterative biological annotation. Results Low-input proteomics identified 7,800 proteins from 1,000 cells while retaining tissue- and cell-associated biological differences. Across the SCP dataset, ~2,300 unique proteins were detected, resolving chondrocyte, synovial fibroblast and immune-cell populations that closely paralleled those previously defined by single-cell transcriptomics. Higher-resolution analysis distinguished lining and sublining fibroblasts, discrete immune populations and multiple protein-defined chondrocyte states. Conclusions SCP enables direct characterisation of cellular heterogeneity within the human OA joint at the level of the proteome. By independently recovering established cellular populations while revealing protein-defined chondrocyte states, our findings establish SCP as a powerful complement to single-cell transcriptomics and open a new route to investigating functional cellular states and disease mechanisms that cannot be inferred from RNA abundance alone.

High-content imaging-based screens with small molecule or genetic perturbations rely on rich optical phenotypes, yet how distinct subcellular markers contribute complementary biological information remains unclear. Here, we introduce CellMAPP, a multiplexed organelle imaging platform that uses optical pooled screens with iterative antibody staining to map cellular architecture under perturbation. Using 17 markers across 524 genetic perturbations in MCF-7 cells, we systematically examined how individual, and combinations of, markers capture diverse morphological responses. We find that markers provide complementary views of perturbation effects, with increased multiplexing improving signal and the breadth of hit recovery, although some phenotypes are better recovered from a subset of features rather than a full profile. Because non-molecular image-based profiles can be challenging to link with interpretable biology, we developed CHART (available at https://github.com/Genentech/CHART), an archetype analysis framework that identifies distinct reference morphological states for each marker and uses multimodal large language models (MLLMs) to automatically annotate them, connecting cellular images to recognizable organelle phenotypes. Integrating archetypes across multiple markers revealed modular morphological programs that describe coordinated organelle reorganization in response to perturbation. This framework identified regulators of defined organelle states such as micronuclei, nominated novel gene functions and revealed complex patterns of multi-organelle reposition. Together, optical profiling with MAPP and CHART provides an interpretable, scalable framework for image-based screens that sheds light on rules governing cellular structural reprogramming.

Lung adenocarcinoma (LUAD) develops through a defined morphological sequence from atypical adenomatous hyperplasia (AAH) to invasive adenocarcinoma, but the transcriptional programmes accompanying this transition have mostly been characterised in bulk tissue or in tumour tissue alone. We profiled paired samples spanning the full sequence by single-nucleus RNA sequencing (413,697 nuclei, 23 patients) and spatial transcriptomics (56 Visium sections, 25 patients), asking how the transcriptional state of the tissue changes as invasion proceeds and whether that state can be reversed in silico by known perturbations. The atlas resolves into 39 second-level subtypes and the tissue into seven recurring niche-domain archetypes (K* = 7), six of which are ordinary lung structures and only one of which is restricted to invasive disease. A confound then shaped the remaining analyses: domain identity, marker stability and the apparent stage-dependence of the spatial programmes all scale with sequencing depth, and depth is in these sections a proxy for tissue density rather than a purely technical artefact (rho=-0.89 between a domain's depth and the fraction of its markers surviving depth matching). Several apparently positive results disappeared once depth was matched. The invasive domain's ECM/interstitial identity survives that check: within patients the fibroblast ECM programme is higher in invasive disease in 20 of 22 paired cases (P=0.0033 against random gene sets), and the result is unchanged when the genes added by hand are removed and under every leave-one-out variant. Querying a patient-paired disease axis against the LINCS compound library with three scoring schemes returned 34 compounds in common, dominated by PI3K/mTOR inhibitors, with a split-half reproducibility of rho=0.962 against a permutation null of +/-0.045 and an agreement of rho=0.735 between a single-cell-derived and a spatially derived axis. A positive control, the recovery of known regulators of an official benchmark (TBXT ranked 1st of 8,450), establishes that the pipeline recovers true positives, so the negative results reported here are properties of the data rather than of the method. Those negatives are stated explicitly: single-cell copy-number inference produced no usable malignant/non-malignant separator, the cohort-level spatial copy-number difference did not survive depth matching, a developmental trajectory ran opposite to that of the source study, gene-level prediction returned no signal on the pooled axis and only one perturbation survives compartment-wise gating, and a ligand-level analysis produced no usable pair.

Aging is a network-level phenomenon, with its hallmarks interacting through dense feedback loops across vastly different timescales. Decades of reductionist research have produced thousands of mechanistic models of narrow subsystems, but no straightforward way to integrate them into one comprehensive system. Consequently, whole-cell and multi-hallmark aging models remain rare, manually constructed, and relatively small, despite broad agreement on their importance. The emergence of autonomous research agents offers a way to distribute this modeling effort across humans and artificial intelligence and thereby greatly accelerate it, but only if the underlying substrate allows agents to readily compose and analyze models with built-in validation checks and modeling guidance, without extensive custom code. We present hallsim, a JAX-native compositional simulation framework designed as the mechanistic substrate for agent-orchestrated research. It provides composability, end-to-end differentiability, GPU execution, and automated analysis and model-selection tools that facilitate composite model construction and fine tuning. The framework makes it easier to propose a mechanistic hypothesis, integrate it into an existing composite, reparametrize it against observed data, and batch-test it across different initial conditions. We demonstrate the co-simulation and in silico perturbation of three independently published kinetic models spanning genomic instability, nutrient sensing, and proteostasis. The models are connected by four edges and calibrated against a public dataset. Additionally, we train a Neural ODE surrogate and compose it alongside the mechanistic modules in a hybrid composite, demonstrating that mechanistic and neural models can function as complementary components of the same system.

Large language model (LLM) agents can propose combination therapies and construct supporting mechanistic models far quicker than either can be verified. To address this gap, we built a gated agentic AI quantitative systems pharmacology (Ai QSP) workflow where no hypothesis reaches a report until it clears strict hurdles for precedent, evidence, structural integrity, and release. We applied this pipeline end to end to healthy-ageing interventions. It began with a 34 state model calibrated on clinical trial data for semaglutide and metformin. Locked and evaluated against an independent post hoc DNA methylation trial, the model captured the direction of change across all 9 organ and system clocks, but the magnitude was far too small: it predicted about 0.14 years per year where 1.6 were observed, halting tier promotion. The workflow then generated seven nutritional combination hypotheses (H1H7) with 48 new parameters encoded in SBML. Identifiability analysis showed that 37 of these were completely unconstrained by available data. Trimming the model down to its lead pairing - a food derived bioactive (Agent A) plus a single-strain probiotic (Agent B left six identifiable parameters and locked the uncalibrated interaction term at zero. From this reduced model, we derived an 84 day trial with a 28 day washout, requiring 356 participants for 80% power at an interaction ratio of 0.75. A 9,000 person virtual trial confirmed the design (80.5% power), but revealed an operational trap: an assay floor at 24 mg/kg slashed power to 33% and skewed estimates toward the null, forcing a protocol-level censoring condition before study lock. Testing the pipeline across downstream operational stages and comparative oncology programmes showed that unconstrained generation consistently over-reports novelty, topological entry points can masquerade as biological mechanism, and governance gates remain vulnerable to operator overrides. Ultimately, the workflow delivered no silver bullet: its output is a single testable interaction, an assay hardened trial design, and an audit trail documenting why the other six hypotheses were shelved. Agent identities, doses, and agent specific sources are withheld in this version pending patent filing.

Decades of antibody-based protein profiling have generated valuable cohorts across evolving cytometry, sequencing and spatial imaging platforms. Differences in marker coverage, signal scale and antibody performance hinder joint analysis and reuse of cohorts with long-term clinical follow-up. Widely used RNA-based integration methods prioritize shared embeddings rather than directly comparable protein measurements. To fill in the gap, we developed RAMP (Rank-preserving Alignment of Multi-platform Proteomics), an unsupervised framework that aligns shared proteins without cell-type annotations. It combines sample-specific anchoring and cross-platform cell matching with bounded monotone transformations. The corrected values retain each protein marker's expression order across cells within a sample. Across seven integration tasks, RAMP achieved the strongest overall balance of batch correction and cell-type preservation. RAMP also improved marker-threshold transfer while preventing expression-order reversals observed with some comparison methods. These reversals made cell types with originally high marker expression appear lower-expressing than other populations. Such overcorrection can mislead cell identification and protein comparisons, even when the integrated embedding shows improved dataset mixing. By preserving each marker's expression order within a sample, RAMP protects the relationships needed to interpret normalized protein measurements. To investigate whether better harmonized inputs improve downstream learning, we used linear, nearest-neighbor and small neural-network models as a controlled, small-scale testbed. This approach tests an input quality question relevant to proteomic foundation models without undertaking large-scale pretraining. Under matched supervision, RAMP improved cross-platform annotation across all three model families and improved missing-marker prediction over unaligned inputs. These gains demonstrate the value of harmonized inputs for predictive learning and motivate evaluating RAMP as preprocessing for proteomic foundation models. In bone marrow, we used high-quality single-cell protein measurements to impute the failed CODEX CD19 channel, recovering B-cell spatial distributions. The completed panel improved B-cell separation from natural killer cells in the joint embedding. In melanoma, a proliferating CD8 T-cell state near tumor cells was associated with longer survival. Transfer such a spatial information to CITE-seq connected tumor proliferation and infiltration state to an adhesion and activation program involving PD-1, LFA-1 and CD2. In lung cancer, we used single-cell CyTOF measurements to predict unmeasured HLA-ABC states in vascular cancer-associated fibroblasts profiled by spatial imaging mass cytometry (IMC). Higher predicted scores were associated with lower immune-cell fractions and worse disease-free survival in the IMC cohort. In head and neck cancer, transferring a fibroblast/macrophage spatial niche score to CITE-seq identified a CD39/TIGIT-associated CD8 T-cell phenotype. Protein and RNA profiles supported checkpoint-associated features, and transfer back to CODEX showed correspondence with measured LAG3. RAMP enables existing single-cell and spatial cohorts to support new discoveries by connecting molecular programs, tissue organization and clinical outcomes across platforms.

Proteins are often classified as toxic or non-toxic without measuring the abundance reached, leaving constraints on tolerable protein abundance unresolved. We established a limit-pushing approach in Saccharomyces cerevisiae combining strong inducible expression with gTOW-mediated high-copy selection to counteract copy-number compensation while measuring protein abundance and growth. Nearly all of approximately 80 chromosome I proteins severely inhibited growth or reduced viability at sufficiently high abundance. We established IE50, the expression level associated with a 50% reduction in growth rate, to quantify their widely varying overexpression tolerance. IE50 was positively associated with predicted structural order and cytoplasmic localization propensity and negatively associated with sulphur content. Single-cell imaging linked higher tolerance to proteins remaining cytoplasmic without becoming aggregation-positive and revealed abundance-dependent changes in localization and organelle morphology. At extreme abundance, Fun12, Nup60, and Pex22 generated distinct large-scale intracellular states through specific sequence regions. These findings establish overexpression toxicity as a quantitative property linked to protein characteristics and reveal both constraints on tolerable abundance and sequence-dependent capacities for intracellular organization.

Spatial transcriptomics is increasingly used to infer cellular phenotypes, molecular interactions, and functional tissue organization from RNA, although most of these biological processes are ultimately mediated by proteins. Here, we directly tested the interchangeability of spatial RNA and protein measurements using paired Xenium profiling of 21 matched targets across ~1.8 million cells from primary lung adenocarcinoma and metastatic castration-resistant prostate cancer. After accounting for marker- and dataset-specific protein background, we quantified concordance at multiple levels, including single-cell detection, expression magnitude, cell-type distribution, patient-level abundance, and spatial organization. Agreement at single-cell level was strongly marker-dependent. Canonical lineage markers generally showed concordant RNA and protein patterns, whereas several molecules, including in our study PD-1, PD-L1, and LAG-3, showed little concordance across modalities. Importantly, spatial organization remained substantially more discordant between modalities than overall abundance, indicating that spatial biomarkers are not directly transferable across modalities. Together, our results show that spatial RNA and protein measurements provide largely overlapping information for lineage markers and overall abundance, but can diverse for specific molecules and for higher-order spatial features, where the two layers provide complementary rather than equivalent readouts. These findings provide practical guidance on when spatial transcriptomic readouts can be interpreted as proxies for protein and when orthogonal protein validation is advisable.

Protein stability, the capacity to maintain a folded structure and function, matters as much for transporters and regulators as for enzymes, and heat erodes it. Wheat, rice, maize and soybean supply two-thirds of human caloric intake, and warming reduces yields. Rubisco activase variants selected for thermostability increased photosynthesis, biomass and yield under heat. They came from gene shuffling; no method proposes residues at proteome scale. Genome editing has made such changes routine, yet predictors estimate melting temperature, name no residue, and have not been reported for a plant proteome. Here we present CropStabDB: 1,425,716 proteins from 30 plant species scored on six interpretable components, with an Engineering Atlas of 2,097,576 candidate substitutions in 51,440 proteins, 785,622 at Engineering Confidence Score [≥] 0.70. Every candidate is a residue a more stable orthologue carries at that column. Substitution-type scores match the direction of measured folding free energies in 842 of 926 domains (90.9%). On three proteomes from families absent from training, the model reaches an area under the curve of 0.919. Applied blind to wheat, the score reproduces the measured thermostability ranking of the Rubisco activase isoforms in all three subgenomes. CropStabDB is freely available without registration at https://cropstab.org.

Drosophila melanogaster is a powerful model for studying systemic metabolism. Although metabolism is compartmentalized across tissues, prior work in adult Drosophila has mostly measured it in dissected tissues rather than in situ, limiting the spatial information that can be gained. Here, we combine whole-fly mass-spectrometry imaging with stable-isotope tracing and flux modeling to quantify metabolic activities across the adult fruit fly in a spatially resolved manner. Across 68 tracer-fed Oregon-R female flies and 187,709 pixels, we map 81 metabolites onto eight annotated regions, including muscle, fat body, gut, and nervous system. Glucose labeling and fitted relative fluxes indicate higher glycolytic and tricarboxylic acid cycle turnover in thoracic muscle than in abdominal fat body; proline labels the nervous system and eye; glutamine the abdominal reproductive system and gut. Alanine carries glucose-derived carbon through the hemolymph and transfers nitrogen into glutamate. In flies bearing yorkie-driven gut tumors, alanine contribution to malate rises in gut and abdomen, and reducing gut alanine uptake lowers tumor burden. Together, this work establishes a framework for quantitative spatial metabolic analysis, showing compartmentalized nutrient use across the adult fruit fly.

Metaproteomics analyses require protein sequence databases, yet the microbial proteins present in complex communities are not known in advance. Searching reference catalogs containing hundreds of millions of candidate sequences reduces identification sensitivity, while generating sample-specific databases through metagenomic sequencing adds time and cost. We introduce FastaLake, a computational workflow that uses full-length de novo peptide predictions by AlphaNovo to retrieve complete proteins from broad sequence resources and selects compact databases for conventional searching. In a defined eight-member community, FastaLake approached reference-search peptide yields without knowing its members. Using public sequence catalogs, FastaLake identified about as many peptides as matched metagenomic databases in two stool cohorts and 20-35% more in intestinal fluid at the same nominal 1% peptide false discovery rate. The resulting proteomes reveal personal functional profiles that persist across intestinal regions despite differences in leading species between intestinal fluid and stool. DNA, RNA and protein profiles agreed on broad differences between patients but only weakly on individual functions. All 726 acquisition-specific databases were smaller than the human reference proteome. Compact databases allowed non-specific searches of all 347 CapScan acquisitions, revealing histidine enrichment at peptide C termini in intestinal fluid. FastaLake enables metaproteomics analysis without matched metagenomic sequencing by constructing search databases from sequence resources and de novo peptide sequencing.

Tissue function is shaped by both cellular composition and the spatial organization of cells. Spatial single-cell technologies capture this organization across tissue sections, enabling tissue architecture to be linked to disease phenotypes, prognosis and treatment response. However, many current graph-based predictive models depend on assay-specific molecular features, limiting joint analysis of cohorts profiled with incompatible panels, and either aggregate local predictions uniformly or rely on post hoc importance scores that do not directly decompose the basis of a sample-level prediction. Here we introduce GRASS-MIL, a graph multiple-instance learning framework that learns the relative relevance and class-specific evidence of individual cellular neighborhoods, predicts phenotypes across platforms, and decomposes each sample-level prediction exactly into signed contribution from each local neighborhood. These contributions are terms of the prediction itself rather than a post hoc approximation and sum exactly to the predicted logit. In simulations, GRASS-MIL distinguished phenotypes with identical global cell-type composition but different spatial organization, whereas composition-based classifiers performed no better than chance. A single model trained jointly on CosMx spatial molecular imaging, Xenium and MERFISH datasets, harmonized only through a shared cell-type vocabulary, predicted treatment status in high-grade serous ovarian carcinoma with an area under the receiver operating characteristic curve of 0.74 and associated this prediction with two stromal architectures: an immune-rich niche admixed with malignant cells and a desmoplastic niche characterized by malignant-cell exclusion. Applied to spatial proteomics data from non-small-cell lung cancer, GRASS-MIL identified distinct niches associated with relapse (area under the receiver operating characteristic curve, 0.57), including myeloid-rich niches in relapsing tumors and adaptive immune niches in non-relapsing tumors. In a separate lung adenocarcinoma cohort, a similar lymphoid-myeloid polarity was observed when predicting 3-year mortality (area under the receiver operating characteristic curve, 0.62), with a B-cell and CD4+ T-cell niche associated with favorable outcome and a neutrophil and macrophage niche associated with adverse outcome. GRASS-MIL thus combines cross-platform phenotype prediction with an exact, signed decomposition of each prediction into the cellular neighborhoods on which it depends, providing a direct route from sample-level prediction to testable hypotheses linking tissue architecture with patient phenotypes.

Protein homeostasis, or proteostasis, is essential for cellular health, and its disruption contributes to cancer, aging, and neurodegenerative disease. Proteostasis is maintained largely by two intracellular degradation systems: the ubiquitin proteasome system (UPS) and the autophagy lysosomal pathway (ALP). Here, we combine pulse stable isotope labeling by amino acids in cell culture (pSILAC) with selective inhibition of the UPS or ALP to quantify pathway-dependent protein turnover and identify substrates that are preferentially cleared through each route. Direct comparison of these perturbations is complicated by their markedly different response distributions. To address the mismatch in dynamic range, we developed a normalization strategy based on within-condition ranks. Protein stabilization values are converted to percentile scores before comparison across pathways, placing both perturbations on a common relative response scale. Across two independent pSILAC experiments, 5,518 proteins were quantified in at least one condition, and 3,814 yielded high confidence half-life measurements in control, proteasome-inhibited, and lysosome-inhibited conditions. Rank normalization resolved pathway sensitivities that were obscured by direct comparison of absolute half-life changes. Importantly, the method reports relative pathway preference rather than absolute pathway contribution, exchanging effect size magnitude for comparability across perturbations with unequal response distributions. Integration with a curated proteostasis network annotation revealed coherent pathway preferences, including coordinated stabilization of proteasome core and regulatory subunits following lysosomal inhibition. This finding provides proteome-wide quantitative evidence for proteaphagy that complements previous targeted observations. Together, these results establish rank normalization as a practical strategy for comparing degradation resolved proteome turnover across perturbations with different dynamic ranges.

Scientific discovery seeks mechanisms that explain observations and predict beyond the measurements that produced them. Whereas large language models (LLMs) encode knowledge implicitly, mechanistic world models organise it as a parsimonious set of explicit, modular, reusable mechanisms whose predictions can be scored against data. In biology, where measurements are sparse and noisy, mechanistic world modelling must discover latent states and governing equations jointly, yet the prior knowledge that could constrain this search is largely unstructured. We introduce gemot, an agentic framework for mechanistic world modelling, and evaluate it on 18 published biological problems spanning molecular biology, epidemiology, and immune-cell differentiation, with up to 1,755 training measurements, 65 observables, 170 experimental conditions, and 3 data modalities per problem. gemot is auditable by construction: a semantic layer records each hypothesis, and a Model Context Protocol layer separates hypothesis formulation from numerical evaluation, so hypothesis scoring cannot be fabricated. On every problem, autonomously constructed models match or exceed the reference models in fit and parsimony, and generalise where held-outs exist. We demonstrate that gemot can formulate novel, biologically plausible mechanistic hypotheses. This work will enable decoding of interventional biological data into competing mechanistic hypotheses and design experiments that distinguish between them, and may serve as a blueprint for mechanistic world modelling in other domains.

Drug combination models must represent cellular context in a form that remains informative when the tested cells or compounds differ from those used for training. Transcription factor regulons offer a biologically structured representation, but their contribution can depend on the accompanying features and the intended prediction task. We developed a symmetric three-class predictor of DrugComb ZIP interactions and compared all seven combinations of landmark expression, Hallmark pathway scores and CollecTRI regulon activities. The common dataset contained 302,042 observations across 3,042 drugs and 155 cellular contexts. We evaluated three group-held-out generalization settings: unseen cell lines, unseen drug scaffolds, and simultaneous exclusion of both, with five training seeds per representation and partition. Pathways plus regulons achieved the highest mean Macro-F1 on unseen cell lines (0.4737, compared with 0.4503 for expression alone). Compared with pathways alone, adding regulons increased mean Macro-F1 by 0.0153 on unseen cell lines and 0.0229 on unseen drug scaffolds, with positive differences across all five paired seeds in both settings. In contrast, the same addition reduced the mean under joint context and scaffold exclusion. Representation rankings also depended on the objective: regulons alone achieved the highest synergy average precision on unseen lines, whereas the Macro-F1-leading combination did not maximize precision among the top-ranked candidates. Further compression of pathways into non-negative matrix factorization programs did not exceed the best context-model Macro-F1 in any setting. These results demonstrate that the predictive value of functional context representations depends on both the type of domain shift and the evaluation objective, and that combining additional biological representations does not necessarily improve generalization. They provide a computational basis for studying drug combinations in senescent states, where transfer must subsequently be assessed using state-specific responses and matched controls.

Genetic interactions can reveal gene function and identify disease-relevant synthetic lethals, but resource-intensive screening in human cells constrains experimental scope and limits our understanding of how generalizable sparse hits are. Here, we leverage principles from yeast genetic networks to identify human gene modules involved in receptor tyrosine kinase signaling and DNA damage response that are predicted to be enriched for genetic interactions, then use our enCas12a-based In4mer combinatorial knockout platform to efficiently screen these modules across diverse cancer cell lines. We identify hundreds of unreported synthetic lethals, including a dense network within the protein N-linked glycosylation machinery that shows both evolutionary conservation and network rewiring, and we confirm that interactions discovered in 2D cell culture are maintained in more physiologically relevant models. Critically, a meta-analysis within and between screens highlights sources of variation in published results, reveals the potential for false positives in screens with many essential genes, and demonstrates the need to move beyond DDR for actionable synthetic lethalities.

Cytos Multimodal Pooled Screening (CytosMPS) measures CRISPR guide identity, morphology, native 3' transcriptome, and protein abundance and localization in the same fixed cells. Following IL-1{beta} stimulation, cells with higher nuclear NF-{kappa}B show lower phospho-p38/HSP27 and higher NFKBIA/DUSP1 in the same cell, a measurement neither imaging nor transcriptomics alone provides. TGFBR2 and CTNNB1 perturbations produce distinct multimodal phenotypes. Guide representation and perturbation transcriptional responses agree with an orthogonal single-cell transcriptomic platform.