AI-empowered perturbation proteomics for complex biological systems

Summary

The insufficient availability of comprehensive protein-level perturbation data is impeding the widespread adoption of systems biology. In this perspective, we introduce the rationale, essentiality, and practicality of perturbation proteomics. Biological systems are perturbed with diverse biological, chemical, and/or physical factors, followed by proteomic measurements at various levels, including changes in protein expression and turnover, post-translational modifications, protein interactions, transport, and localization, along with phenotypic data. Computational models, employing traditional machine learning or deep learning, identify or predict perturbation responses, mechanisms of action, and protein functions, aiding in therapy selection, compound design, and efficient experiment design. We propose to outline a generic PMMP (perturbation, measurement, modeling to prediction) pipeline and build foundation models or other suitable mathematical models based on large-scale perturbation proteomic data. Finally, we contrast modeling between artificially and naturally perturbed systems and highlight the importance of perturbation proteomics for advancing our understanding and predictive modeling of biological systems.

Systems biology remains under adopted in life sciences, primarily due to scarce perturbation proteomics data. Qian et al. emphasize perturbation proteomics' role in elucidating biological causal relationships and enabling precise intervention predictions. They review cutting-edge high-throughput perturbation proteomics technologies and propose the PMMP (perturbation, measurement, modeling to prediction) framework. The authors advocate for computational models based on large-scale perturbation proteomic datasets, which promise to personalize therapeutic regimens and uncover functions of understudied proteins.

Introduction

Contemporary biology has greatly advanced our understanding of biological processes through a reductionist approach. However, it faces challenges such as the complexity and variability of biological systems, the functional coordination of diverse processes, experimental reproducibility, and gaps in knowledge regarding understudied proteins.^1,2 Recent reports highlight a troubling reproducibility issue in cancer biology research, underscoring the importance of measurement precision, research rigor, and the inherent complexity of biomedicine.^1,3

Most studies have focused on a limited set of widely discussed proteins, such as p53, leaving many human proteins and associated interactions insufficiently investigated. This implies significant gaps in the functional characterization of the human proteome and its ability to react to perturbations.^1,2 This perspective suggests that perturbation proteomics—the systematic and precise measurement of the near-complete proteome across multiple layers in various perturbed states—can now be performed at an unprecedented high throughput, advancing our understanding of biological systems.

Insights from physics

The term "spherical cow" in theoretical physics represents an overly simplified model of complex phenomena. Scientific model development, however, is iterative and incremental, with models evolving to incorporate more detail as we refine our understanding. Perturbations act as catalysts in this process, revealing additional information that simpler models might miss. The 2021 Nobel Prize in Physics recognized contributions to understanding complex physical systems. It involved studying perturbations by human activities in these systems, particularly in the context of the Earth’s climate,⁴ suggesting that sophisticated mathematical modeling, bolstered by comprehensive measurements of the system in different perturbed states, can provide an increasingly complete understanding of these complex systems.

A prime example of this approach is the GraphCast model, a recent innovation in global weather forecasting. By harnessing graph neural networks with 36.7 million parameters, GraphCast can predict a wide array of weather variables for up to 10 days, at a detailed 0.25° geographical resolution, in less than 1 min.⁵ This achievement underscores the power of complex models in predicting the behavior of dynamic systems.

Perturbation theory offers a systematic framework for approximating a complex system. This approach is predicated on the realization that many physical systems are too complicated to model following a reductionist approach. So, it is more promising to start with a basic model and then incrementally add the effects of small “perturbative” changes, yielding a series of increasingly accurate approximations that take into account the underlying mechanics and interactions of the system.⁶ Celestial mechanics, for instance, utilizes perturbation theory to model the gravitational influence on the motion of celestial bodies.⁷ Percolation theory, another physics-derived principle, analyzes the behavior of connected components in a network when some node or edge is added or removed randomly or deliberately.⁸ This theory works well in understanding critical transitions in network connectivity. The application of these physical theories to biological molecular networks, such as protein interaction networks and metabolic pathways, reveals the effects of perturbations on network stability and resilience.^9,10 While physical systems' behavior can often be predicted in an unperturbed state, biological systems remain challenging to understand.

Lessons from systems biology

Can the success of modeling the climate system and celestial mechanics be replicated in biological systems? Systems biology,^11,12 a research field that emerged in the early 2000s, takes a holistic approach to studying complex biological systems as networks of interacting elements. By modeling the consequences of the interactions between individual components, systems biology aims to infer emergent properties that cannot be predicted by studying isolated components. It utilizes experimental and computational techniques, integrating data from genomics, proteomics, and metabolomics with observations of phenotypes to model biological systems at different scales, ranging from molecules to physiological states of tissues, organs, and humans.

However, despite the potential of systems biology, its adoption in the life sciences has been relatively limited. One crucial reason is the lack of comprehensive sets of perturbation experiments and data describing the effects of the perturbations on the states of the system. Systematic perturbation-response data strengthen the ability of researchers to hypothesize causal relationships between variables. By then manipulating variables and observing the resulting changes, researchers can often infer the direction and strength of causal relationships, thus enhancing their understanding of underlying mechanisms and enabling practically useful predictions about system responses to investigational or therapeutic interventions.^13,14,15

Some studies have recognized the importance of controlled perturbation-response studies, producing large datasets like the “Connectivity Map” project^16,17 and the NIH Library of Integrated Network-Based Cellular Signatures (LINCS) initiative.^18,19,20 These have established functional links between diseases, genetic factors, and drug perturbations. However, most of these investigations have been primarily focused on nucleic-acid-based measurements, whereas proteomic measurements, which more directly reflect the functional processes and complexity of biological systems, have been limited in scope.

To construct a cellular-protein-based network, it is necessary to determine both nodes (protein, metabolite, or mRNA expression levels) and edges (e.g., protein-protein/metabolite/mRNA/DNA interactions or correlations).²¹ Existing databases provide some edge information but generally do not capture the dynamic network rewiring resulting from cellular perturbations.^{22,23,24,25,26,27,28} High-throughput quantitative protein measurement in sets of perturbation experiments consisting of protein-protein interactions and the interactions of proteins with nucleic acids, metabolites, or lipids increase the power of computational methods to infer functional interactions.²⁹

Recent advancements in mass spectrometry (MS)-based proteomics technology have substantially improved throughput, depth, and quantitative reproducibility.^30,31,32 Preparation of minute amounts of samples for proteomic analysis has been substantially parallelized and accelerated by emerging techniques such as pressure cycling technology (PCT),³³ filter-aided sample preparation (FASP),³⁴ single-pot solid-phase-enhanced sample preparation (SP3),³⁵ and suspension trapping (S-trap).³⁶ Increasingly, robotic systems are applied to automate the previously labor-intensive sample and error-prone preparation procedures. On the other hand, the throughput of MS analysis has also witnessed unexpected improvements. Stable isotope-empowered data-dependent acquisition (DDA) proteomics allows the simultaneous quantification of up to 18 samples in one batch,³⁷ reaching a proteome depth of over 10,000 proteins in human samples.³⁸ In contrast to DDA-MS, which inherently suffers from stochastic sampling when analyzing complex proteome, data-independent acquisition (DIA)-MS offers a high degree of quantitative reproducibility by rapid acquisition cycles of fragment ion mixtures from multiple precursor ions with a relatively large m/z range.³⁹ While MS offers substantial advantages for systematic profiling and the identification of unexpected proteoforms, affinity-based methods like ELISA, proximity ligation assays, protein microarrays, Olink,⁴⁰ and SOMAscan⁴¹ offer complementary advantages with their high throughput and broad dynamic range. Known protein biomarkers with availability of highly specific antibodies or aptamers can be measured reproducibly in body fluids like plasma specimens. Thus, they currently cannot be used to measure understudied proteoforms and post-translational modifications (PTMs), which are widely present and functional participants in biological processes, unless proteoform-specific affinity reagents are developed and validated.^1,2

In summary, while systems biology offers promising approaches to understanding biological systems, the availability of comprehensive perturbation experiments combined with proteomic measurements would substantially enhance its descriptive and predictive power.

Emergence of perturbation proteomics

In biological systems, perturbations of various types, intensities, and timings impact multiple levels. These effects range from alterations in target proteins and their interactions to changes in protein turnover, localization, and activity. These perturbations initiate signal transduction through PTMs, dynamically rewiring downstream networks and resulting in diverse cellular fates. To investigate these early responses and subsequent changes, a variety of proteomic techniques can be employed to examine alterations in protein expression levels and states of modification. Here, we propose a systems biology pipeline named PMMP, with the four steps perturbations, measurements, modeling, and prediction (Figure 1).

Figure 1.

The PMMP pipeline for perturbation proteomics

This diagram illustrates the PMMP (perturbations, measurements, modeling to prediction) pipeline, which involves perturbing various biological systems—including cells, organelles, tissues, and animal models—through diverse methods such as genetic editing, microbial infections, drug treatments, and physical interventions (e.g., light, temperature, radiation). The dynamic alterations in phenotypic attributes and protein-level characteristics of these systems are then analyzed. Protein-level measurements can encompass a broad spectrum, from direct impacts on target proteins, alterations in protein-protein interactions and protein localization, and changes in post-translational modifications and protein turnover to extensive downstream modifications within the network. Subsequently, machine learning models are constructed based on these assessments, aiming to predict protein functions, modes of action, and system responses to the perturbations. This predictive capability is vital for crafting precise therapeutic interventions and planning synergistic experimental approaches in systems biology research.

Perturbations of biological systems

Perturbations can be categorized into biological, chemical, and physical dimensions and can involve time series, dose series, or combinations of perturbagens. Biological perturbations include pathogen infections and genetic perturbations, such as natural genetic variations from genetic reference panels,^{42,43,44,45,46} RNA interference (RNAi),^47,48 and CRISPR.^{49,50,51,52,53,54,55,56,57} Perturbation of the expression of a particular gene is valuable for inferring gene functions; however, the cost is non-trivial for mammalian cells. Therefore, large-scale CRISPR-based proteomic datasets studying non-essential genes are limited to a few organisms like E. coli⁵⁸ and yeast.^59,60 Recent advancements have successfully integrated unique CRISPR-based perturbations with guide barcodes into single-cell genetic sequencing technologies. This integration has facilitated large-scale screens, including those conducted on human cells.⁵⁴ However, these investigations predominantly employ transcriptomic technology as a readout, which does not account for protein PTMs, protein levels, turnover, and interactions, which are crucial, as they directly mirror biological functions.

Chemical perturbations involve various compounds such as drugs and toxins. Compound perturbation datasets are mostly transcriptomic data or targeted proteomic data. For instance, the Golub group pioneered the first-generation Connectivity Map, which analyzed genome-wide mRNA expression data from 164 unique small-molecule perturbations (including drugs and tool compounds) applied to three cancer cell lines at early time points (6 or 12 h post-perturbation).¹⁶ In their next-generation Connectivity Map, the group aimed to incorporate more chemicals and genetic perturbations across a diverse range of cell types.¹⁷ To achieve this, they developed a high-throughput, cost-effective L1000 platform.¹⁷ This platform profiles a reduced representation of the transcriptome of 978 landmark transcripts and has generated 1,319,138 L1000 profiles from 42,080 perturbagens (19,811 small-molecule compounds, 18,493 short hairpin RNAs, 3,462 cDNAs for gene overexpression, and 314 biologics).¹⁷

The Jaffe team assessed changes in approximately 100 phosphorylated peptides, which serve as early bioactivity indicators for a diverse set of signaling pathways and drug mechanisms, and roughly 60 post-translationally modified peptides from core nucleosomal histone proteins.¹⁸ These assessments were made in response to 90 small molecules spanning various mechanisms of action (MOAs), including focused subsets of kinase inhibitors and epigenetically active compounds, across five cancer cell lines and one neurodevelopmental cell model.¹⁸ Compared to the Connectivity Map based on transcriptomic data, this perturbation-reduced phosphoproteome offers a distinct biological dimension, highlighting intra-class connectivity to a broad array of therapeutic and investigational drugs and revealing cell-line-specific vulnerabilities.¹⁸ The Liang group evaluated 210 clinically relevant proteins using RPPAs (antibody-based reverse-phase protein arrays) for 12,000 perturbed samples in response to approximately 170 preclinical and clinical drug compounds covering diverse MOAs and some of their combinations across 120 cancer cell lines.¹⁹ This study showed that both the baseline and perturbed proteomic data had increased predictive power for drug sensitivity and potentially synergistic drug combinations.¹⁹

Considering the balance between cost and depth, further proteomic datasets with increased depth have been generated.^61,62,63 Recently, the Gygi group profiled the 24-h perturbation proteome of 875 compounds, covering nearly 10,000 proteins, for the deconvolution of MOAs and compound repurposing.⁶¹ The Kuster group reported protein and PTM expression changes in response to drug perturbations in a time- and dose-dependent manner and that these changes measured by MS reflect the MOAs of the drugs.^62,63

Extrinsic physical perturbations, resulting from environmental factors such as light, radiation, temperature, and pressure, can induce substantial proteome changes.⁶⁴ Environmental proteomics studies have investigated their impact on various organisms and ecosystems.^65,66 However, extrinsic perturbations are non-specific perturbations of the entire cell, not targeted to particular genes or gene products. Therefore, the value of extrinsic perturbation experiments for studying cell systems is limited.

Cell lines are commonly used to investigate perturbation effects, although more sophisticated models like organoids, tissue cultures, and animals can provide insights into complex biology. Single-cell studies are advantageous for investigating heterogeneity upon perturbation. Ex vivo tissue perturbations with single-cell RNA sequencing (RNA-seq) can measure relevant molecular-level responses.

Animal models present a more complex system but allow for drug perturbations and genetic alterations. Genetic reference panels for multi- and unicellular organisms provide valuable experimental platforms for studying natural genetic variations and complex disruption scenarios.

When using these models, it is crucial to consider interspecies differences to ensure that the insights are relevant and applicable across species boundaries, enhancing our understanding of biological complexity and aiding in the translation of research findings into potential therapeutic interventions.

The design of perturbation experiments and data analysis are specific to the biological systems of interest. Cell lines are the most commonly accessible models to investigate the effects of perturbations. Other sophisticated experimental models such as organoid, tissue culture, and animals can also be used to understand the complexity of cell, tissue, and organ biology. Even though cell lines cannot capture many functional effects of whole organisms and may have genome modifications, the analysis of omics data from the drug perturbation of cell lines has revealed meaningful correlations and coherence with molecular signatures identified in patient biopsies exposed to the corresponding drug treatments. The cell line data in some cases also correlated well with clinical trial outcomes.¹⁷ This observation underscores the potential utility of perturbed cell lines for understanding and predicting biological systems.

Single-cell studies are advantageous for investigating changes in heterogeneity upon perturbation, particularly in the context of the cell cycle.⁶⁷ Organoids cultured in a three-dimensional (3D) spatial environment are heterogeneous, similar to actual tissues, and if derived from patient tissues, can incorporate additional factors such as different cell types, e.g., tumor cells, various immune cells, and stromal cells. The utilization of human ex vivo tissue perturbations in conjunction with single-cell RNA-seq can measure relevant molecular-level responses that may align with in vivo patient data. This could potentially expedite the process of drug development directly within human tissue, thereby laying the groundwork for gene regulatory network models.⁶⁸ A variety of platforms or strategies have been designed for drug screening on ex vivo tumor slices.^69,70,71 These are intended to capture the complex and multifaceted heterogeneity of a tumor and its surrounding microenvironment while maintaining the ability for high-throughput screening. Further molecular profiling, particularly those conducted at the level of single cells or small numbers of homogeneous, sorted cells based on these ex vivo models with drug perturbations, could provide invaluable insights that could enhance the effectiveness of personalized medicine.

Animal models present a more complex system, with interconnected components such as the immune system and organ crosstalk. While these models readily permit the introduction of drug perturbations, the creation of genetic alterations has traditionally been a laborious and low-throughput endeavor. Recently, adeno-associated virus (AAV)-mediated in vivo single-cell CRISPR screening (AAV-Perturb-seq) has been introduced, facilitating the high-throughput study of genotype-phenotype relationships in intricate tissues.⁷²

Importantly, genetic reference panels for multicellular organisms like the BXD murine lines,^42,43,44 as well as for unicellular organisms such as Saccharomyces cerevisiae or Schizosaccharomyces pombe,^45,46 provide valuable experimental platforms. These systems are distinguished by (1) their reliance on natural genetic variation honed by evolutionary processes, (2) the genetic homogeneity within each strain, which facilitates the introduction of additional perturbations to generate complex disruption scenarios, including genetic and chemical interactions, and (3) the ability to dissect tissue-specific effects of single or combined perturbations, which is particularly pertinent in animal models for probing gene-environment interactions (GxE).

When using these models, it is crucial to consider interspecies differences. This careful approach ensures that the insights gleaned are relevant and applicable across species boundaries, thereby enhancing the accuracy of our understanding of biological complexity and aiding the translation of research findings into potential therapeutic interventions.

Measurements at the protein and phenotype levels

A wide range of proteomic techniques are at our disposal, and their selection hinges on the scientific inquiries at hand. This portfolio of techniques encompasses evaluations of distinct protein targets via perturbation; assessments of protein’s interactions with other proteins, nucleic acids, and other types of biomolecules; PTMs; protein turnover; protein localization; enzyme activities; and exhaustive proteome quantifications prior to and following perturbations (Figure 2). Complementing these examinations, concurrent measurements of related phenotypic data, such as cell viability, form the basis for clinical applications by establishing a quantitative connection between molecular alterations and system-wide phenotypic changes, which can be captured by computational models applied to an optimal choice of drug targets.

Figure 2.

The MS-based technologies for protein-level measurements in the PMMP pipeline

These measurements are commonly conducted post-perturbation within in vitro cell lines, ex vivo tissues, or in vivo models. Additionally, target identification can be carried out on cell lysates post-perturbation.

(A) Direct target identification. For drug incubation scenarios, several methodologies—such as CCCP, ABPP, TPP, and LiP-MS—can be deployed. CCCP and ABPP exploit drug-target competition for target protein identification, whereas TPP and LiP-MS leverage the stability changes of target proteins in response to temperature alterations and non-specific protease cleavage after drug-target binding. The ultimate identification of potential drug-binding targets utilizes differential protein quantification between drug-treated and control samples via mass spectrometry.

(B) Drug-protein interaction analysis. Techniques for studying these interactions include (1) affinity purification of protein complexes via genetic tagging or available antibodies for targeted proteins, (2) proximity labeling and purification of neighboring proteins through genetic tagging with enzymes for targeted proteins, (3) detection of interaction topology via cross-linked peptide quantification using cross-linking agents, and (4) global interactome studies through co-fractionation, leveraging the similar migration profiles of proteins belonging to the same complex.

(C) Downstream alteration studies. These primarily rely on various quantification approaches, including protein abundance quantification, post-translational modification (PTM) quantification following enrichment, investigation of protein synthesis and degradation via pulse-chase experiments with isotopic labeling, protein localization characterization through ultracentrifugal separation of organelles followed by protein quantification, single-cell proteomics research after cell sorting for individual cell isolation, and spatial proteomics research through techniques like laser capture microdissection (LCM) and expansion microscopy for tissue sectioning.

The identification of direct targets or interacting proteins can be achieved through a variety of chemical⁷³ and non-chemical proteomics methods.⁷⁴ For this topic, we refer readers to a recent comprehensive review.³⁰ Briefly, these methods encompass compound-centric chemical proteomics (CCCP),⁷⁵ activity-based protein profiling (ABPP),⁷⁶ thermal proteome profiling (TPP),⁷⁷ limited proteolysis-coupled MS (LiP-MS),⁷⁸ drug affinity responsive target stability (DARTS),⁷⁹ affinity purification,⁸⁰ and photoaffinity labeling.⁸¹ For example, Kinobeads, a CCCP method that uses immobilized kinase inhibitors, captures the kinome (set of protein kinases) and associated proteins targeted by the compounds.⁸² This approach, when coupled with MS-based proteomics, facilitates the subsequent analysis of these protein targets.⁸³ ABPP, on the other hand, employs active-site-directed chemical probes to directly assess the functional state of numerous enzymes in native biological samples, allowing for the characterization of dysregulated enzymatic activities in disease and the discovery of targeted and off-target proteins by drugs.^84,85 Both methods depend on the availability of chemical probes derived from the bioactive compound. In contrast, non-chemical proteomics methods are designed based on the influence of compound binding on protein structure and biophysical properties. For instance, LiP-MS incorporates a double-protease digestion step to discover perturbation-induced protein structural states and drug targets,^78,86,87 while TPP monitors the melting profile of expressed proteins to identify drug targets, protein-small molecule interactions, and PTMs.⁸⁸

In terms of protein interactions, multiple approaches, such as affinity-purification MS,^22,23,24 cross-linker-⁸⁹ and proximity labeling enzyme-based⁹⁰ based reactions to capture neighboring proteins, and protein complex co-fractionation coupled to MS,^91,92 have been utilized to build valuable databases for protein networks in various cell lines. However, only limited data are available on perturbation-induced interaction changes.

After perturbation, the impacted target proteins will potentially cause alterations in downstream protein turnover, localization, and/or enzyme function. Protein turnover is essential for maintaining protein homeostasis.⁹³ High-throughput methods for studying protein turnover mainly combine MS-based proteomics with “pulse-chase” methods, which utilize radioactive, biochemical, or stable-isotope-labeled tracers to measure protein production through tracer incorporation and the replacement of labeled tracers with unlabeled compounds.⁹³ Integrative multi-omics approaches, combining RNA-seq, proteomics, and ribosome footprinting, have also shown promise in studying protein turnover.^94,95 Precise identification of protein subcellular localization is pivotal in perturbation proteomics, given that protein spatial organization aligns with their functional roles. A variety of approaches have been utilized to determine protein localization, including imaging-based methods employing endogenous tagged proteins^96,97 and affinity reagents,^{98,99,100,101,102} as well as MS-based proteomic profiling in conjunction with subcellular compartment enrichment^103,104 and proximity-dependent enzymatic reactions.¹⁰⁵ The use of endogenous tagged proteins typically necessitates gene editing, whereas affinity reagents for subcellular imaging exploit the specific labeling of antibodies conjugated with metal ions or fluorescent oligonucleotides. This approach can surpass the optical diffraction limit of traditional immunohistochemistry (IHC) methods. The recently introduced secondary-label-based, unlimited multiplexed point accumulation in nanoscale topology (PAINT) has achieved speed-optimized imaging at sub-5-nm resolution, effectively achieving single-protein resolution.¹⁰² Enzyme activity assays provide valuable insights into multiple cellular mechanisms. Various methods, including optical, magnetic resonance, MS, and physical sampling approaches, are employed in enzyme activity assays.¹⁰⁶

PTMs play crucial roles in intracellular signal transduction pathways. Large-scale PTM studies have been conducted to investigate the phosphoproteomes of yeast induced by 101 environmental and chemical perturbations,¹⁰⁷ P100 (∼100 phosphorylated peptides) and GCP (∼60 post-translationally modified peptides from histones) profiling of six cell lines induced by 90 drugs,¹⁸ and dose-/time-resolved modulation of phosphorylation, ubiquitination, and acetylation of 13 cell lines induced by 31 drugs in the global proteome scale.⁶²

Last but not least, whole-proteome analysis is currently the most mature technology, as evidenced by perturbed proteomic datasets for various compounds and drugs in different cell lines, including A549,¹⁰⁸ lung cancer cell lines,¹⁰⁹ HCT116,⁶¹ breast cancer cell lines,¹¹⁰ and Jurkat acute T cell leukemia cells.⁶³ The proteome of NCI-60 cell lines has been profiled at the proteome level using both DDA-MS¹¹¹ and DIA-MS¹¹²; however, no perturbations have been performed in these studies. These bottom-up proteomic methods involve the enzymatic digestion of proteins into peptides of smaller sizes before MS analysis. DDA, despite its challenges with data reproducibility and missing values, has seen improvements in throughput due to advancements in stable isotopic labeling technologies like Tandem Mass Tag 16 tag set (TMTpro),¹¹³ which allows for simultaneous analysis of multiple samples. Meanwhile, DIA overcomes the sampling issues of DDA by systematically fragmenting all ionized peptides, providing better reproducibility and accuracy, and enabling high-throughput analysis.^31,114 On the other hand, the emergence of top-down proteomics enables the identification of proteoforms, the various protein products originating from a single gene due to alternative splicing and PTMs.^115,116 This allows for a more profound analysis of the proteome and the discovery of novel proteoforms that could have roles in disease progression and treatment response.

Cell surface activity is crucial in cell behavior and communication. Studying these extracellular factors forms an integral part of proteomic analysis. Secretome analysis emphasizes enrichment or isolation methodologies,^117,118 and subsequent proteomic quantification aligns with the whole-proteome profiling mentioned earlier. Extracellular vesicles (EVs), which can carry specific disease biomarkers, are gaining increasing importance in monitoring cancer progression and therapy. Their isolation is based on ultracentrifugation, ultrafiltration, immunoaffinity capture, precipitation, size-exclusion chromatography, and innovative microfluidics techniques.¹¹⁹

In addition, it is important to acknowledge the complexity introduced by cell heterogeneity, particularly when considering cell cycle effects and induced pluripotent stem cell (iPSC)-derived model systems, like neurons. The heterogeneity inherent in these systems can complicate the interpretation of bulk omics data considerably. In such cases, the adoption of single-cell or small bulk techniques following cell type enrichment can prove beneficial. Recently developed single-cell proteome methods, such as antibody-based CyTOF and CITE-seq (cellular indexing of transcriptomes and epitopes by sequencing),¹²⁰ as well as MS-based methods,^{67,121,122,123,124,125,126} are increasingly being utilized to provide a more comprehensive profile of individual cells. In particular, CITE-seq allows for the simultaneous detection of surface protein expression and transcriptomes in single cells, while MS-based methods provide unbiased coverage of proteome expression within its range of detection.

The selection of omics methods is typically guided by the specific research question at hand and the available technology. While all omics methods are applicable across various biological systems, certain techniques may not be as commonly used in certain studies due to current technological limitations. For instance, despite its potential, phosphoproteomics is less frequently utilized in single-cell studies due to existing technological constraints, a challenge that may be surmounted in the future. For a comprehensive guide to proteomics/omics methods, please refer to specialized reviews.³⁰

Maximizing biological insight often involves the strategic combination of different methods. For instance, coupling CCCP or ABPP with affinity-purification MS can yield valuable data on protein targets and their interactions. Additionally, to monitor downstream changes in protein behavior following perturbation, whole-proteome analysis and PTM studies can be integrated. Lastly, utilizing a multi-omics approach can provide a holistic view of system-wide changes, thus enhancing our understanding of complex biological phenomena.

Phenotypic data can be collected at various levels, tailored to specific research objectives. At the cellular level, extensive datasets have been published, including IC50 values and area under the percentage of viability curves (AUC) following drug treatments,^{20,127,128,129,130} as well as genes essential for cell viability.^49,131,132 These datasets provide valuable insights into drug responses and the molecular underpinnings of cellular functions. At the patient level, researchers collect diverse clinical data points to characterize phenotypes. This includes information on recurrence-free survival and overall survival following therapeutic interventions. These clinical markers help to understand disease progression and evaluate treatment efficacy in patient populations. However, detailed clinical data from drug trials combined with proteomic profiles of bodily fluid or tissue samples are rare.

Prediction

Perturbation proteomics data allow prediction of perturbation response, identifying targets, inferring MOAs, and predicting protein functions. Perturbation response prediction involves forecasting changes in omics signatures and phenotypic changes post perturbation, such as protein expression alterations, IC50 values, and synergy effects of drug combinations. This can aid in selecting optimal clinical therapies. Target identification and MOA inference aim to discern the drug’'s mechanism of action and support drug repurposing. Protein function prediction seeks to predict the roles of understudied proteins based on perturbation proteomics data and functions of well-studied proteins. Computational models can provide insight into the roles of uncharacterized proteins by analyzing proteomic response patterns post gene editing.

Overall, proteome-based prediction of perturbation effects could offer a range of applications, including therapeutic regimen selection, target and MOA identification, and protein function inference, using large-scale perturbation-response proteome and phenotype data.

Computational modeling

Quantifying perturbation-induced changes in protein levels, PTMs, interactions, and localization enables investigation of their correlation with and causation of cellular or organismal phenotypic alterations. Perturbation proteomics data offer an invaluable foundation for the construction of predictive models and aid in the interpretation of predictions.

By selectively perturbing specific variables or features, researchers can pinpoint inputs that exert the most significant influence on the model’s output. This process, in turn, enhances understanding and interpretation of the model’s predictive capacity. To build predictive perturbation response models and interpret the perturbed omics data, various computational methods have been employed, including correlation networks,¹³³ regression models,⁵⁰ fuzzy-logic models,¹³⁴ dynamic Bayesian networks,¹³⁵ causal models,¹³⁶ multiple input-multiple output models,¹³⁷ a belief-propagation-based network inference algorithm,¹³⁸ and ordinary differential equation (ODE) network models,^131,139,140 among others. In addition, ensemble models, by integrating different models, have been reported in some cases to achieve better performance than the best individual one, though careful evaluation of orthogonality is required.¹⁴¹ Some reviews have extensively summarized perturbation modeling using single-cell omics data, which may serve as examples for future perturbation proteome studies.^142,143

The advent of artificial intelligence (AI) technology has brought about significant advancements in the field, with AlphaFold2 being a prominent example.¹⁴⁴ By leveraging existing protein structure data and AI techniques, AlphaFold2 has demonstrated remarkable capabilities in structural biology and beyond.¹⁴⁴ Nevertheless, real-world biological systems encompass thousands of protein types, each with variations in a range of attributes, including copy numbers, subcellular locations, PTMs, and synthesis-degradation homeostasis. The complexity involved in understanding and predicting these aspects of protein function is much higher compared to prediction of the analogs of static crystal protein structures from sequence.¹⁴⁵

To address these challenges, emerging large pretrained models, such as foundation models, have shown promise. Transfer learning, where models learn from a vast amount of data and apply that knowledge to new tasks, has proven successful in the field of omics data analysis.^{146,147,148,149} For example, the Geneformer model, developed by C. Theodoris, P. Ellinor, and their teams, is pretrained on a dataset of 30 million single-cell transcriptomes. This pretrained model effectively predicted potential targets for cardiomyopathy.¹⁴⁶

However, it is crucial to note that while foundational models are powerful, they may not always be the best fit for molecular datasets based on real-world evaluations of their performance in single-cell transcriptomics.¹⁵⁰ There is a need for specialized and context-aware models that are tailored to the complexities of molecular data and the integration of domain-specific knowledge, addressing issues such as data heterogeneity, scale, dimensionality, and the ability to accurately capture causal relationships between molecular entities.

The inference of causal relationships necessitates a substantial number of perturbation experiments, especially when dealing with a large system with numerous variables, as theoretical work has indicated.^151,152,153 The emerging field of causal representation learning^153,154 shows promise in this regard. It may offer a promising approach in predicting unseen proteomic perturbations and their phenotypic impacts, a concept that Pearl describes as the second level of the causality ladder.¹⁵⁵

Despite the presence of a large amount of diverse proteomic data in public databases such as PRIDE¹⁵⁶ and iProx,^157,158 perturbation proteomics datasets sufficient to build a large pretrained model are currently not available. The data requirement varies depending on the model’s complexity and specific tasks. For example, considering the Geneformer model as a reference,¹⁴⁶ we may need more than its scale of 30 million for perturbation proteomics due to the inherent complexity of proteomic data. This includes not only concentration changes of proteins but also alterations in their PTMs, protein-protein interactions, and subcellular localizations under different perturbations. Given the current proteomic data collection throughput, it is unlikely to complete this in a short time frame. However, with the anticipated rapid advancement of multiplexing single-cell proteomics technologies, we are optimistic about the feasibility of acquiring enough data for pretraining models in the near future.

Integration of multimodal data

The success of the PMMP pipeline hinges on the quality and quantity of the input perturbation proteome data. In addition, the inclusion of relevant but orthogonal datasets is also beneficial. These include transcriptomics, metabolomics, chromatin state assessment methods like DNaseI hypersensitivity screening and assay for transposase-accessible chromatin with sequencing (ATAC-seq), Simplified Molecular input Line Entry System (SMILES) of compounds, protein interaction and regulation networks, and imaging data.

Each data type offers a unique perspective on cellular function and response. For example, metabolomics provides a snapshot of the cell’s metabolic landscape, indicative of its functional status and active pathways.^159,160,161 The Zampieri team’s application of untargeted metabolomics in E. coli, coupled with CRISPRi-mediated gene perturbation, facilitated the prediction of compounds' unconventional MOAs.¹⁵⁹ Chromatin accessibility studies, such as DNaseI hypersensitivity screening and ATAC-seq, illuminate gene transcription and cellular responses to perturbations,^115,162 while imaging data, with its visual representation of cellular structures and functions, is useful in evaluating the morphological and phenotypic impacts of perturbations,^163,164,165 as exemplified by the CPJUMP1 dataset, which compiled approximately 3 million images of two cell types (U2OS and A549) under paired chemical and genetic perturbations.¹⁶³

Integrating these diverse data sources provides a net information gain and allows predictive models to better deal with the noise, missing data, and variability inherent in biological systems. However, achieving cost effectiveness in data acquisition methods is crucial for the scalability and feasibility of integrative analyses, particularly in resource-intensive areas like proteomics and imaging.

Integrating these methodologies could foster more robust and transferable predictions across different experimental conditions or datasets. Additionally, the combination of different information layers could empower machine learning algorithms to uncover previously unknown associations and patterns, leading to more causality-oriented answers. This multi-model approach could expedite the discovery of new biomarkers, drug targets, and therapeutic strategies.

Modeling of perturbation experiments versus observational studies

Observational studies involve monitoring a system under natural conditions, without any investigational perturbations,^166,167 posing challenges in discerning causal relationships due to potential unobserved confounding factors. However, techniques like Mendelian randomization and colocalization have made significant progress in inferring causal links with diseases, such as the association between the APOE gene and Alzheimer’s disease.^168,169,170 Despite their limitations, these methodologies represent valuable tools for causal inference in observational settings.

In proteomics, there is an emphasis on systematic proteomic profiling in under-represented populations, particularly those with bottleneck populations or those enriched for consanguinity, such as communities in South Asia¹⁷¹ and East London.¹⁷² These populations provide unique opportunities to study high-impact, rare genetic mutations and their associated proteomic changes, aiding in hypothesis development about causal pathways linking genetic variation to disease phenotypes.

However, concerns about hidden variables affecting the interpretation of observational data persist. Notwithstanding these complexities, observational studies are valuable, offering insights when ethical or practical constraints make perturbations unfeasible.

In conclusion, systematic perturbation experiments provide a much more powerful methodology for inferring causality, while observational studies are essential for examining systems where interventions are either unfeasible or ethically questionable. The integration of both observational and perturbation data presents a potent approach to gain a comprehensive understanding of intricate biological systems, generate causal hypotheses, and derive quantitatively predictive models of system responses to perturbations.

Outlook

Perturbation proteomics promises to reflect drug responses, tailor therapies, elucidate drug MOAs, and uncover functions of understudied proteins.^{19,58,59,60,61,62,63,140} These studies provide crucial insights into disease biology and drug discovery, though rigorous biological and clinical validation is necessary.

The challenge for perturbation proteomics lies in the infinite perturbation space in drug discovery, considering variables like cell type, compound, concentration, treatment duration, and drug combinations. Therefore, ideally, we can train AI models based on existing perturbation proteomics data and obtain the ability to predict perturbation responses. However, few studies have yielded effective models due to data scarcity.^61,62,63

Next-generation proteomics technologies allow the effective generation of perturbation proteomics data. Methods such as isobaric labeling, implemented in the context of bulk cells and single-cell proteomics,^173,174,175 enable us to probe proteomic changes at a cellular granularity never seen before. Advancements in miniaturization and automation within proteomic workflows enhance efficiency and make large-scale, high-resolution proteomic profiling practical. In addition, PTM and proteome interactome data acquisition post-perturbation remain challenging due to the costs and time demands of enrichment before identification.

Data integration is another critical obstacle, requiring alignment of protein expression matrices and addressing variations due to different acquisition methods. Stable isotopic labeling technologies like TMT have enhanced throughput while maintaining depth. However, using ratios alone may result in the loss of inter-protein expression information. To address this, protein abundance is adopted to preserve this information for self-supervised learning in the foundation model. The difference between protein abundance quantified by TMT-labeled proteomics and protein intensity quantified by DIA is several orders of magnitude, which requires careful normalization to preserve the original biological information. Finally, integrating perturbation data for direct targets, protein-protein/RNA/metabolite interactions, downstream data, and multi-omics data is challenging. This requires the development of standardized data integration frameworks, machine learning for pattern recognition, and advanced computational modeling.

The rich data generated from perturbation proteomics, together with cell phenotypes, can be further leveraged for perturbation response prediction, perturbation target identification, and exploration of related molecular mechanisms via computational modeling. On the one hand, unsupervised learning techniques can be utilized to decode the inherent structure and dependencies within data, without the prerequisite of specific labels or annotations. These techniques can be used to pretrain foundational models, which can be further fine-tuned using data with explicit perturbation-response labels. However, these foundational models often pose interpretability challenges due to their complex architectures. On the other hand, relevant models can be established by feeding perturbation proteomics data into mechanistic models grounded in first principles, such as ODEs, or using more data-driven, causal-oriented methods. Such approaches typically provide enhanced interpretability, which is instrumental in facilitating prediction, extrapolation, and understanding of systems biology. Moreover, these computational strategies are essential for designing precise interventions, whether single or combinatorial, aimed at specific outcomes like therapeutic efficacy or synthetic biology applications. The integration of these two modeling paradigms can potentially offer a powerful tool for the study and predictive modeling of complex biological systems.

Furthermore, the development of AlphaFold 3, with its advanced diffusion-based architecture, has expanded our capabilities in modeling the structures of complex biomolecular assemblies. This model’s ability to predict interactions involving proteins, nucleic acids, small molecules, ions, and modified residues is unprecedented. The integration of AlphaFold 3 into the perturbation proteomics toolkit is poised to illuminate the intricate networks of molecular interactions that dictate cellular function and response. This convergence promises to unlock new perspectives in our quest to understand and manipulate biological systems, paving the way for groundbreaking advancements in personalized medicine and bioengineering.

Rapid advancements in proteomic techniques and AI models have opened up opportunities for understanding cellular processes and predicting perturbation responses. To harness these opportunities, collaborations between researchers in proteomics, bioinformatics, mathematical modeling, and AI are needed. There is also a need for standardized proteomic data sharing platforms and industry partnerships to facilitate the translation of scientific discoveries into practical applications.

Acknowledgments

The authors are thankful for the financial support from the following grants: the National Natural Science Foundation of China (Major Research Plan, grant no. 92259201) to T.G.; the China National Postdoctoral Program for Innovative Talents (BX20240330) to L.Q.; the "Pioneer" and "Leading Goose" R&D Programs of Zhejiang (2024SSYS0035) to T.G., the Westlake Education Foundation to T.G., and the National Resource for Network Biology (U.S.-NRNB) to C.S. This perspective was partly inspired by discussions from the π-HuB project.

Author contributions

T.G. and C.S. conceived the idea; L.Q., T.G., and R.S. drafted the article; R.A. and P.B. revised the manuscript; and T.G. and C.S. supervised the project.

Declaration of interests

T.G. is a shareholder of Westlake Omics, Inc. C.S. is on the scientific advisory board of Cytoreason.com.