The new method, called HDOCK-Multimer, or HDM, is described in Nature Computational Science and tackles a problem that has grown increasingly urgent as AlphaFold and its successors have transformed the field. While deep learning systems such as AlphaFold2, AlphaFold-Multimer and AlphaFold3 achieve remarkable accuracy on small and medium-sized complexes, their performance degrades as the number of chains grows. The reasons are partly architectural and partly practical: end-to-end predictors suffer from insufficient conformational diversity when modeling many interacting chains, and they run into hard memory limits on graphics processing units. AlphaFold-Multimer version 2, for example, can handle only around 3,300 residues on a 40-gigabyte NVIDIA A100 GPU, while AlphaFold3 extends its input limit to roughly 5,000 residues on its server and about 5,120 residues for local execution on an 80-gigabyte card. Many biologically important assemblies, including respiratory chain complexes and ATP synthase components, far exceed these thresholds.

Assembly-based methods offer one workaround. These approaches first predict smaller subcomponents with AlphaFold and then stitch them together into the full complex. MoLPC, for instance, uses Monte Carlo tree search to assemble complexes from predicted dimers or trimers, while CombFold applies a hierarchical combinatorial assembly strategy that explores inter-subunit interactions more broadly and yields higher-quality predictions. But both methods inherit a critical weakness: their success depends entirely on the quality of the subcomponents produced by AlphaFold-Multimer. If a predicted subunit or interface is wrong, the entire assembly collapses. Symmetry-aware tools such as EvoDock and SymProFold improve modeling of symmetric assemblies but do not generalize to the heteromeric, asymmetric complexes that dominate cellular biochemistry.

HDOCK-Multimer resolves this tension by fusing two historically separate traditions: classical protein-protein docking and combinatorial assembly. The framework takes subunit sequences as input and proceeds through three stages. First, it generates monomeric and subcomponent structures using AlphaFold2 and AlphaFold-Multimer, defining subcomponents as all possible trimeric combinations of subunits. Trimers were chosen as a deliberate compromise, capturing more cooperative interactions than pairwise combinations while avoiding the prohibitive computational cost and GPU memory overflow that tetrameric predictions would incur. Second, the system selects appropriate modeling strategies, including asymmetric docking, symmetric docking and combinatorial assembly, based on the complex’s homology and pairwise docking results. Third, it constructs the full complex using the selected strategies, evaluates the resulting models with an iterative knowledge-based scoring function called IT-score, and clusters them using a root-mean-square-deviation cut-off of 5 angstroms to eliminate redundancy.

The docking component is where HDM distinguishes itself most sharply. For symmetric assemblies, the method supports cyclic, dihedral, tetrahedral and octahedral symmetries, inferring candidate symmetry types from chain number and pairwise docking results rather than requiring explicit symmetry input. For asymmetric complexes with fewer than five unique chains, it employs a multi-body docking procedure built on the team’s earlier HDOCK software, which uses a fast Fourier transform-based search. Each chain is treated as a node in a graph, pairwise docking interactions become edges, and complex modeling is recast as a minimum spanning tree construction problem that borrows principles from Prim’s and Kruskal’s algorithms. Iterative score and interface screening prune the enormous combinatorial space of possible chain arrangements, which is precisely the bottleneck that has historically limited docking approaches to dimers.

The assembly side is equally sophisticated. HDM extracts representative subunits and pairwise transformations from predicted subcomponents, performs 500 rounds of random assembly sampling using create, append and merge operations, and then refines the resulting population with a genetic algorithm. Crossover operations recombine subunits from parent structures, while mutations rebuild complexes from alternative transformations. A confidence score, weighted by the number of residues involved in each assembly operation, ranks candidate structures, and molecular dynamics relaxation with the ff14SB force field removes steric clashes before final scoring. Ablation experiments on a benchmark of 35 large heteromeric complexes showed that each component contributes: the basic assembly strategy alone already outperforms MoLPC and CombFold, genetic algorithm optimization increases the number of acceptable predictions, and IT-score ranking after relaxation enriches high-quality models among the top-ranked outputs.

The benchmark results are striking. On the first benchmark of 35 heteromeric complexes containing 5 to 20 chains each, HDM achieved top-1 success rates of 83 percent for acceptable accuracy and 66 percent for high-quality predictions, compared with 74 and 54 percent for CombFold, 29 and 11 percent for MoLPC, and 29 and 26 percent for AlphaFold-Multimer v2. Measured by interface quality with the DockQ metric, HDM reached 86 percent top-1 success. In one showcase example, the Polytomella complex I, HDM produced a high-quality prediction with a template modeling score of 0.85, while CombFold and MoLPC managed only 0.64 and 0.32, with CombFold failing primarily because two subunits were misplaced. Notably, AlphaFold-Multimer v2 failed entirely on several targets due to GPU memory limitations, whereas HDM assembled them with varying degrees of accuracy.

On a second benchmark of 172 large complexes with 10 to 30 chains, including 114 homomers and 58 heteromers, HDM achieved top-1 and top-10 success rates of 59 and 66 percent, ahead of CombFold at 41 percent, MoLPC at 30 percent, and AlphaFold3 at 55 and 58 percent. The comparison with AlphaFold3 deserves nuance: the deep learning system produced substantially more high-quality predictions on the complexes it could process, though it could model only 145 of the 172 targets because of input limits, and the benchmark targets predate January 2022 and may overlap with its training data. Where HDM truly shines is scale. For a glycolyl-CoA carboxylase containing 7,062 residues, well beyond the AlphaFold3 server limit, HDM delivered a near-perfect prediction of the resolved regions with a template modeling score of 0.98 while reconstructing the unresolved portion of the alpha subunit with an average predicted confidence of 91.06, recovering 2,749 additional amino acids. For the human GCN5 histone acetyltransferase domain, HDM’s octahedral symmetry docking yielded a high-quality result where AlphaFold3 failed outright.

Tested on seven real targets from the 15th Critical Assessment of Structure Prediction, including two single-chain proteins longer than 3,000 amino acids that were partitioned into domains using disorder predictions, the fully automated pipeline generated acceptable models for six of seven targets, a success rate of 86 percent. CombFold matched that rate, while MoLPC and AlphaFold-Multimer v2 succeeded on only three and one targets respectively. The team also demonstrated that HDM can infer stoichiometry, the number of copies of each subunit, by enumerating possible combinations and using a normalized IT-score, correctly identifying the stoichiometry for 14 of 20 representative complexes, compared with 9 using confidence scores alone. An analysis of interface novelty further showed that prediction accuracy correlates only weakly with similarity to interfaces in AlphaFold’s training data, indicating the method can genuinely assemble complexes it has never seen.

The authors are candid about remaining limitations. Their symmetric docking covers only a subset of symmetries, excluding helical assemblies; model ranking sums pairwise interaction scores rather than capturing higher-order cooperativity; and stoichiometry prediction, while promising, is not absolute and ultimately depends on subcomponent quality. They also note that integrating experimental data such as crosslinking mass spectrometry, which HDM already supports, fluorescence resonance energy transfer and cryo-electron microscopy could further improve accuracy, and that antibody-antigen modeling remains an open question, with the method succeeding on two of four tested targets. Still, with the software freely available on GitHub and full complexes typically built within hours on modest CPU hardware after subcomponent prediction, HDOCK-Multimer represents a pragmatic and powerful hybrid: it keeps the accuracy of deep learning where it works, and deploys decades of docking physics where it does not, opening the door to atomic-level models of the molecular machines that have long eluded both crystallographers and algorithms alike.

Subject of Research: Computational prediction of large protein complex structures by integrating docking and combinatorial assembly

Article Title: HDOCK-Multimer: improved structure prediction of large protein complexes by integrating docking and combinatorial assembly

Article References: Yao, X., Ya, Y., Li, H., & Huang, S.-Y. (2026). HDOCK-Multimer: improved structure prediction of large protein complexes by integrating docking and combinatorial assembly. Nature Computational Science. https://doi.org/10.1038/s43588-026-01062-1

Image Credits: AI Generated

DOI: 10.1038/s43588-026-01062-1

Keywords: protein structure prediction, HDOCK-Multimer, protein docking, AlphaFold, combinatorial assembly, protein complexes, computational biology, deep learning, stoichiometry prediction, structural biology, symmetric assemblies, CASP15

Perturbational similarity maps are among the most ambitious artifacts in modern biology. The idea is elegantly simple: knock out or repress thousands of genes one at a time, record the molecular or cellular consequences of each perturbation, and then compare those consequence profiles to one another. Genes whose disruptions produce similar effects are inferred to act in the same pathways or protein complexes, turning a mountain of raw screening data into a functional network. If such maps were transferable, a perturbation atlas painstakingly assembled in one system, for example the mouse brain, could be used to interpret screens performed in human cell lines, dramatically multiplying the value of expensive experiments and accelerating the translation of animal findings into human biology.

Testing that assumption required a careful matching infrastructure. Wen began by restricting the comparison to genes with strict one-to-one orthology between mouse and human, a conservative choice that retained 71 shared perturbation targets for comparison with the K562 Perturb-seq dataset, 90 for the RPE1 Perturb-seq dataset, 695 for HeLa morphology profiles from the PERISCOPE resource, and between 273 and 311 for nine L1000 cell-line matrices. Even before any statistics were computed, the attrition was striking: the vast majority of perturbations profiled in any single system simply have no directly comparable counterpart in the others, a practical constraint that any cross-system transfer strategy must confront.

The core of the analysis rested on three complementary endpoints, each designed to probe a different notion of transferability. The first was global pair-rank concordance: if two perturbed genes are ranked as similar neighbors within the mouse brain map, do they receive similar ranks in the human datasets? This was quantified with Spearman rank correlation across all pairs of shared targets. The second was source-defined top-k neighbor preservation, a stricter test asking whether a perturbation’s ten nearest neighbors in the source map remain among its top neighbors in the target dataset. The third was within-validation retrieval against CORUM, a curated database of protein complexes, which checks whether external perturbation matrices show internal structure consistent with known molecular assemblies.

The headline numbers were sobering. Under the primary zero-anchored root-mean-square representation of the response features, the Spearman correlation between mouse and human pair ranks was 0.026 in K562 and 0.035 in RPE1, effectively indistinguishable from zero. The two HeLa morphology conditions yielded correlations of 0.0022 and 0.0055, and across the nine L1000 cell lines the median was 0.0055. In other words, knowing which perturbations resemble each other in the mouse whole-brain atlas provided essentially no information about which perturbations resemble each other in any of these human systems. The neighbor-preservation test fared no better: across all thirteen tested contexts, not a single source-defined top-10 preservation result survived statistical correction for multiple comparisons, with the smallest adjusted q-value reaching only 0.240.

One of the study’s most technically important findings concerns how sensitive these conclusions are to the mathematical representation of the data. Perturbation profiles are high-dimensional vectors, and the choice of how to normalize and scale them can reshape the geometry of the entire map. When Wen switched from the primary zero-anchored RMS representation to across-target standardization in the RPE1 dataset, the correlation rose to 0.102, with a 95 percent confidence interval of 0.027 to 0.201, a weak but statistically detectable signal. Omitting the feature RMS scaling step instead produced a descriptive correlation of negative 0.034. The lesson is double-edged: representation choices can manufacture or destroy apparent transferability, which means that any published claim of cross-system concordance must report and justify its preprocessing pipeline in detail.

The study also probed the reliability of the maps themselves, asking how much of the failure to transfer could be blamed on noise within each dataset. Across fifty random cell-level splits of the K562 and RPE1 Perturb-seq data, the median global correlation between independently derived maps was 0.600 and 0.710 respectively, with top-10 neighbor overlap of roughly 0.50. Across fifty mouse-level splits of the K562- and RPE1-matched source profiles, the median global correlations were 0.731 and 0.703. These figures indicate that the individual maps are internally reproducible to a moderate degree, yet they also reveal substantial split-to-split variability. Crucially, Wen notes that a true cross-species noise ceiling, the maximum correlation one could expect even for perfectly transferable biology given measurement noise, remains unestimated, so the near-zero observed correlations cannot yet be cleanly separated from an unfavorable signal-to-noise regime.

The within-validation CORUM analysis added an important nuance. Several of the external human matrices did show evidence of protein-complex structure when analyzed on their own terms, meaning that perturbations targeting members of the same complex tended to cluster together within those datasets. But Wen is careful to point out that this result does not test preservation of the source map’s neighbor structure; a dataset can be internally biologically coherent while remaining globally discordant with the mouse brain atlas. This distinction matters for the field because within-dataset validation is often presented as evidence of quality, when the more demanding question, whether functional structure survives the journey between systems, is a separate and harder test.

The implications reach well beyond cell culture. In vivo Perturb-seq has recently been used to map transcriptional networks in the developing cortex, to dissect astrocyte function, and to link psychiatric disease risk genes to specific molecular mechanisms, and such whole-brain atlases are increasingly proposed as reference resources for interpreting human genetic data. The new results do not invalidate that enterprise, but they do impose a discipline: conclusions drawn from a perturbation map in one species or modality should not be assumed to hold in another without direct empirical verification, and the choice of feature representation must be treated as a first-class scientific decision rather than a technical afterthought.

Wen is equally explicit about the limits of the negative findings. The results are conditional on the specific representations and feature selections tested, they concern non-neural human systems that differ profoundly from brain tissue, and they do not establish that transferability is absent, particularly for neural systems where matched perturbation data do not yet exist at comparable scale. The analysis code, fixed random seeds, processed inputs, and reviewer-requested analyses have been deposited in a versioned Zenodo archive, and the underlying public datasets from the whole-brain atlas, the Replogle Perturb-seq screens, PERISCOPE, and the Connectivity Map remain freely available. That transparency should make it straightforward for other groups to extend the audit, perhaps to neural organoids, assembloid screens, or emerging spatial Perturb-seq platforms, as those resources mature. For now, the study stands as a rigorous caution: in perturbational genomics, the map is not yet the territory, and every boundary between species, cell systems, and assay modalities must be tested rather than assumed.

Subject of Research: Cross-species and cross-platform transferability of perturbational similarity maps in functional genomics

Article Title: Testing the Transferability of Perturbational Similarity Maps Across Species, Cell Systems, and Assay Modalities

Article References: Wen, J. (2026). Testing the Transferability of Perturbational Similarity Maps Across Species, Cell Systems, and Assay Modalities. Neuroinformatics, 24(4), Article 66. https://doi.org/10.1007/s12021-026-09821-1

Image Credits: AI Generated

DOI: 10.1007/s12021-026-09821-1

Keywords: Perturb-seq, Neuroinformatics, perturbation atlas, cross-species comparison, L1000, Connectivity Map, protein complexes, CORUM, reproducibility, representation transfer, CRISPR screens, whole-brain mapping

The core insight behind the approach is disarmingly simple. AlphaFold3 can model the structure of protein complexes, and researchers have increasingly used it to test whether two proteins that have never been shown to interact might nevertheless form a physical pair. The obvious way to apply this to an entire organism is to run the algorithm on every possible pair of proteins one at a time. But the arithmetic quickly becomes brutal. The number of possible pairwise interactions scales quadratically with the size of the proteome, so even a modest bacterium generates tens of thousands of candidate pairs, while the human proteome, with roughly 20,000 proteins, yields around 200 million. Running that many individual AlphaFold jobs is, as the authors of the accompanying commentary note, computationally intractable with realistic resources.

Todor and colleagues sidestepped the problem by refusing to play the game one pair at a time. Instead of enumerating every possible interaction, they grouped proteins into random pools, generally containing between 10 and 25 proteins, and submitted each pool to AlphaFold3’s free online server as a single job. Because AlphaFold3 models all proteins in an input simultaneously, each pooled run produces an all-versus-all comparison within that pool. The pools were designed so that every pair of proteins in the proteome appears together in at least one pool, guaranteeing that the full set of pairwise interactions gets modeled once all the pools have been processed. The logic that makes this work is that true interactions are rare: out of the vast number of possible pairs, only a small fraction are real, so the extra proteins packed into a pool are unlikely to interfere with the evaluation of genuine interactions.

The efficiency gains are striking. Compared with the traditional pairwise approach, pooling reduced the total number of AlphaFold jobs by 100-fold and improved the runtime by twofold. More surprisingly, the pooled approach also improved the overall accuracy of identifying true interactions. For a set of seven proteins, the enumerated pairwise strategy would require 21 separate AlphaFold runs, while the pooling strategy accomplishes the same coverage in just three. In a field where compute budgets routinely dictate the scope of a study, a method that is simultaneously cheaper, faster, and more accurate is a rare and welcome combination.

The organism chosen for the demonstration, Mycoplasma genitalium, is a free-living bacterium with one of the smallest known genomes, containing roughly 475 proteins and therefore 113,050 possible pairwise interactions. That scale is manageable enough to serve as a proof of principle, but it also highlights the limits of the current achievement. The human proteome’s roughly 200 million possible pairs remain far beyond what pooling alone can conquer. Other research groups have attacked the scaling problem from a different direction, prioritizing pairs of proteins that are most likely to interact based on independent evidence. Burke and colleagues, for example, produced AlphaFold2 models of confident protein pairs drawn from publicly available human protein interaction databases such as hu.MAP 2.0 and HuRI, while other teams have applied related strategies to core eukaryotic complexes and to systematic human interactome computations.

Prioritization, however, has an inherent dependency: it requires experimental interaction data to exist for the species being studied, or at least for a close relative. Non-model organisms, which include most of the bacterial and archaeal diversity on the planet, often lack such data entirely. This is precisely where the pooling approach shines, since it needs no prior knowledge to get started. Importantly, the two strategies are not in conflict. When experimental data are available, pooling and prioritization could in principle be combined in future efforts, using evidence-based filters to narrow the candidate space and pooled AlphaFold3 screening to evaluate what remains. For understudied organisms, pooling stands as a valuable new addition to the computational toolkit.

Why pooling improves accuracy remains an open question, and the authors have offered several hypotheses. One appealing idea involves competition among interactions. Because the pooled approach evaluates many potential interactions simultaneously, only the most confident interaction at any given protein interface will be selected in the final model. This competitive filtering could suppress false positives, particularly for promiscuous proteins that have a tendency to produce spurious predictions when considered in isolation. A deeper understanding of the mechanism behind the accuracy boost would be more than an academic curiosity; it would allow researchers to engineer protein pools deliberately, optimizing their composition to maximize predictive performance rather than relying on random assignment.

That same hypothesis, however, exposes a potential weakness. If AlphaFold3 prioritizes some interactions based on spatial constraints, then mutually exclusive interactions, in which two different proteins bind a third protein at the same interface, may cause one partner to be systematically favored over the other. Such exclusivity is common in nature and often underlies molecular functions that switch depending on cellular context. The severity of the problem depends on how many mutually exclusive interactions exist and on the size of the pools. With the current setup of 10 to 25 proteins per pool, the issue is likely to be rare, but if completeness becomes the explicit goal, or if pool sizes are increased to gain further efficiency, it will need to be addressed head-on.

There are also limitations inherited from AlphaFold3 itself. The algorithm depends on multiple sequence alignments to identify co-evolving pairs of amino acids, both within a single protein and between proteins in a complex, and it uses those co-evolutionary signals as distance constraints when building structural models. Proteins without sufficiently deep alignments suffer greatly in accuracy. In Mycoplasma genitalium, around 40 proteins, roughly 8 percent of the proteome, have limited sequence homology and appear to be found only in this species. As a result, nearly 18,000 protein pairs cannot be evaluated by this approach at all. That gap is more than a technical footnote, because species-specific proteins and their interactions may be responsible for the unique biology of the organism. Context-dependent interactions, such as those that form only under particular cellular conditions or in specific cell types, are similarly difficult for the method to discern.

Looking ahead, pooled screening offers an intriguing path toward detecting higher-order interactions, such as trimers and tetramers, which are often more biologically relevant than simple pairwise contacts. The obstacle is probabilistic: purely random pools make it unlikely that three or more proteins that genuinely interact in the cell will land in the same pool. Capturing all pairwise interactions through pooling is likely to be effective, but exhaustively screening all possible three-way combinations, roughly 18 million in Mycoplasma genitalium alone, will remain out of reach for quite a while unless pools are chosen intelligently rather than at random. Even so, the trajectory is clear. Todor and colleagues have brought the field measurably closer to the long-sought goal of complete, all-versus-all computational screening of protein interaction landscapes, for model organisms and neglected ones alike. The efficiency gains and the unexpected accuracy boost suggest that pooled AlphaFold screening will become an indispensable tool in a toolkit that is expanding at a remarkable pace.

Subject of Research: Proteome-wide prediction of protein-protein interactions using pooled AlphaFold3 screening

Article Title: Proteome-wide AlphaFold pool party

Article References: Drew, K. (2026). Proteome-wide AlphaFold pool party. Molecular Systems Biology, 22(4), 477-479. https://doi.org/10.1038/s44320-026-00198-6

Image Credits: AI Generated

DOI: 10.1038/s44320-026-00198-6

Keywords: AlphaFold3, protein-protein interactions, Mycoplasma genitalium, structural biology, computational biology, proteome, deep learning, protein complexes, interactome, multiple sequence alignment, molecular systems biology, Proteome-wide

The study exploited an unusual natural experiment: fourteen HeLa cell line variants, collected from thirteen laboratories around the world, that have drifted genetically over decades of independent culture. These lines display strikingly different behaviors — some grow faster, some are more invasive, and some are far more susceptible to Salmonella infection — despite descending from the same original tumor. Because this genetic drift generated a diverse landscape of genotypes, proteomes, and phenotypes without any deliberate manipulation, the panel serves as a natural perturbation system, ideal for tracing how molecular differences cascade into functional differences. Previous work on these cells had profiled their genomes, transcriptomes, and total proteomes, but never their protein complexes, leaving a critical gap between molecular measurements and cellular behavior.

The team closed that gap using SEC-SWATH-MS, a technique that physically separates proteins by size before quantifying them by mass spectrometry. Because protein complexes elute from the size-exclusion column at positions corresponding to their molecular weight, while monomeric proteins appear at their individual sizes, the method physically distinguishes the assembled and unassembled fractions of the same protein. Applying this approach to the two most behaviorally divergent lines — the invasive, infection-prone HeLa CCL2 and the faster-growing HeLa Kyoto — across 420 mass spectrometry runs, the researchers generated a quantitative matrix covering 7,175 proteins. A computational toolkit called SECAT then translated these co-elution profiles into six distinct metrics per protein: total abundance, assembled abundance, monomer abundance, complex abundance, interactor abundance, and interactor ratio.

The most striking finding concerns a long-standing hypothesis in molecular biology: that protein complex assembly buffers cells against variation in gene copy number and transcription. Earlier studies had inferred this buffering indirectly from bulk correlations between mRNA and protein levels, but never demonstrated it by physically separating the two states of the same molecule. The new data provide exactly that direct evidence. For proteins present in both monomeric and assembled forms — the so-called paired state — the monomeric fraction tracked gene dosage far more tightly than the assembled fraction. Spearman correlations with copy number variation were 0.322 for monomers versus 0.198 for assembled proteins, and with mRNA abundance 0.574 versus 0.484, differences that reached statistical significance. In other words, when copy number or transcription fluctuates, the excess unassembled subunits appear to be selectively degraded by cellular quality-control pathways, while the functional complex-bound pool remains stable.

This buffering is not a passive curiosity but appears central to cellular fitness. When the researchers cross-referenced their protein classifications with CRISPR gene dependency scores from the DepMap project, they found that paired-state proteins — those subject to assembly buffering — are significantly more essential for cell survival than proteins detected only as monomers, with a p-value of 1.35 × 10⁻²⁹ and median dependency probabilities of 0.065 versus 0.032. Complex assembly, the authors conclude, functions as a protective shield insulating core cellular machinery from the noise of genetic drift. This is precisely the kind of mechanistic insight that total protein abundance measurements, however accurate, cannot deliver, because they average together the monomeric and complexed pools into a single number.

To capture additional layers of protein function beyond structure, the team deployed two network-based inference algorithms. VIPER infers transcription factor activity from transcriptomic data by measuring the enrichment of a regulator’s known target genes among differentially expressed transcripts. VESPA applies the same enrichment logic to phosphoproteomic data, estimating kinase and phosphatase activities from the phosphorylation states of their substrates. Applied to the HeLa panel, msVESPA identified 73 of 549 kinases and phosphatases as differentially active between CCL2 and Kyoto cells, while msVIPER flagged 2,009 differentially active regulatory proteins out of 6,156 tested. The researchers then validated these inferred activities against three independent reference datasets: CRISPR dependency scores, subcellular localization annotations from the Human Protein Atlas, and microRNA perturbation responses, confirming that the inferences prioritize essential genes, map to coherent cellular structures, and track phenotypic outcomes.

Bringing all three layers together through network diffusion analysis, the study surfaced two major mechanistic themes distinguishing the cell lines. The first centers on invasion and infection. The actin-nucleating Arp2/3 complex, which drives the formation of invadopodia — protrusive structures that both degrade surrounding tissue and admit bacterial entry — is more abundant in CCL2 cells. Critically, the regulatory proteins WIPF1 and WIPF2, which control Arp2/3 activation through the WASP family, show a decisive stoichiometric shift: WIPF1 predominates in CCL2, promoting invadopodium initiation, while WIPF2 dominates in Kyoto, where it blocks initiation and preserves a stable cortical actin network. This WIPF1/WIPF2–Arp2/3 axis plausibly explains why CCL2 cells are both more invasive and more susceptible to Salmonella, since the same membrane-ruffling machinery serves pathogen and tumor cell alike.

The second theme involves immune adaptation. SECAT revealed that while total proteasome subunit abundance barely differs between the lines, CCL2 cells assemble both constitutive proteasomes and immunoproteasomes — an inflammatory variant that generates different peptide fragments for MHC class I antigen presentation. Kyoto cells, by contrast, rely exclusively on the constitutive form, with higher levels of proteasome assembly chaperones. Such an immunoproteasome switch, invisible to conventional abundance profiling because PSMB8 and PSMB9 showed no significant total-abundance changes, represents a form of immune phenotypic plasticity with obvious implications for how tumor cells evade immune surveillance. Similar abundance-invisible mechanisms appeared throughout the data: the IQGAP2–CDC42 scaffolding complex and COP9 signalosome–DDB1/2 supercomplexes differed in stoichiometry between the lines while neither partner changed detectably in total amount.

The broader significance of the work lies in its reframing of what a protein measurement should mean. A protein, the authors argue, is not a single attribute but a bundle of context-dependent states — its abundance, its assembly, its phosphorylation, and its regulatory activity — each captured by a different molecular network. By reconstructing context-specific protein-protein interaction, kinase-substrate, and gene regulatory networks from the data itself and propagating signals across them, the framework identifies functional drivers rather than abstract variance components, complementing latent-factor approaches such as MOFA. The nominated mechanisms are directly testable: LIMK1 activity can be probed with kinase assays, SCF complex reconfiguration by co-immunoprecipitation, and the WIPF1/WIPF2 ratio by live-cell imaging of invadopodia. The authors acknowledge limitations — the approach still demands substantial sample and instrument time, and reference databases remain biased toward well-studied biology — but they envision the framework as a stepping stone toward artificial intelligence virtual cell models capable of adaptive, mechanism-based predictions of cell state. For now, the message is clear: to understand why two genetically similar cells behave so differently, look past how much protein they make, and ask what that protein is actually doing.

Subject of Research: Multi-omics inference of protein complex assembly and activity states as determinants of cellular phenotypic variability

Article Title: Complex assembly and activity states as multifaceted protein attributes explaining phenotypic variability

Article References: Rosenberger, G., Xue, P., Bludau, I., Martelli, C., Williams, E., Collins, B. C., Califano, A., Liu, Y., & Aebersold, R. (2026). Complex assembly and activity states as multifaceted protein attributes explaining phenotypic variability. Molecular Systems Biology. https://doi.org/10.1038/s44320-026-00228-3

Image Credits: AI Generated

DOI: 10.1038/s44320-026-00228-3

Keywords: proteomics, protein complexes, multi-omics, SEC-SWATH-MS, SECAT, HeLa cells, protein activity, network diffusion, copy number variation, immunoproteasome, invadopodia, systems biology