Activity-based annotation: The emergence of systems biochemistry

Abstract

Current tools to annotate protein function have failed to keep pace with the speed of DNA sequencing and exponentially growing number of proteins of unknown function (PUFs). A major contributing factor to this mismatch is the historical lack of high throughput methods to experimentally determine biochemical activity. Activity-based methods, such as activity-based metabolite and -protein profiling, are emerging as new approaches for unbiased, global, biochemical annotation of protein function. In this review, we highlight recent experimental, activity-based approaches that offer new opportunities to determine protein function in a biologically agnostic and systems-level manner.

The growing challenge of protein annotation

Nucleic acid sequencing technologies have permeated virtually every field of biological research, producing huge amounts of genomic data every day. The advent of metagenomic technologies has more recently made it possible to access the genomes of unculturable organisms and complex communities and generate hundreds to thousands of different genome sequences from a single analysis. This explosion of genomic data has been accompanied by an expansion of protein sequences from model organisms to ever more diverse members of the tree of life. The Global Ocean Sampling expedition, for example, almost doubled the number of known protein sequences, accessing a deep cache of proteins from aquatic microorganisms that were vastly underrepresented in the collective genome data [1]. This expansion of DNA sequence data to non-model organisms has been accompanied by a growing number of genes that lie beyond the reach of standard similarity-based bioinformatic annotation methods. While indispensable for annotation of newly sequenced genomes, existing bioinformatic methods have failed to annotate as much as 50% of putative protein coding sequences and left many known and essential predicted protein activities unannotated [2–4]. Accordingly, it has been estimated that over 30% of the proteins in well-characterized organisms such as Saccharomyces cerevisiae and humans are uncharacterized enzymes [5]. Plant genomes are even more sparsely annotated [6–8].

The scope of similarity-based annotations (see Glossary) for all but the most conserved proteins is further limited by drift in protein function and has inadvertently given rise to potentially significant rates of misannotation [9]. Indeed, some estimates indicate that nearly 50% of current similarity-based annotations could be inaccurate at the gene level [10], while as much as >80% provide only general class or protein superfamily level annotations [3,11]. Misannotation unfortunately propagates alongside correct annotation and erodes the overall annotation quality. Existing similarity-based annotations have thus proven insufficient to meet the challenge of protein annotation.

Protein function is a product of multiple factors, such as biochemical activity, cellular localization, pathway context, protein-protein interactions, and other binding partners such as ligands and nucleic acids and manifests as isolated molecular activity to higher level, physiologic phenotype. Function can coincide with different domains of the protein as in many signaling proteins with modular architecture or may even be housed in the same domain such as the phosphatase activity of bacterial histidine kinases [12]. Annotation of protein function thus requires knowledge beyond its isolated coding sequence and full annotation may require several independent annotations.

While a number of tools ranging from functional genomics to structure prediction have helped to fill these gaps on a more global and high-throughput scale, biochemical methods have lagged in their ability to achieve a similar systems-level scale. Here, we highlight two notable exceptions that enable high-throughput, systems biochemistry: Activity-based metabolite (ABMP) and -protein profiling (ABPP). We argue that these emerging tools have the potential to accelerate experimental and unbiased protein annotation and, in conjunction with existing bioinformatic tools, close the gap between protein sequence and biochemical function.

Improvements in existing annotation tools

Sequence-based annotation is becoming increasingly sophisticated and, in addition to homology-based methods, now incorporates context features that can increase the accuracy of prediction [13]. One such example is gene proximity, or synteny, either within or across genomes, which can indicate shared function, a concept that has long been known for bacteria but that also holds for eukaryotes [14]. As such, gene locus features are now part of most automatic annotation pipelines. Additional gene context-related features include co-expression and conserved, sequence-driven protein-protein interactions.

A mainstay of high-throughput functional annotation has been genome-scale mutagenesis coupled to whole genome sequencing. By screening a given library of mutants under a given condition, it has become possible to identify groups of genes whose loss of function is associated with a similar fitness cost, indicating potentially shared or related functions [15]. In cases where such gene sets include already annotated functions, this approach has proven especially informative by enabling direct experimental validation of the putative function [16–19]. Interestingly, the predictive specificity of this approach was recently further expanded through the compilation of datasets from different experimental conditions into a single repository. Doing so allowed for the identification of potential functionally related genes not only through the presence of conserved genes across different gene sets but also through the identification of genes with similar or shared patterns of associated gene sets across experiments (i.e., gene set fingerprints) [20]. The advent of temporally and quantitatively tunable clustered regularly interspaced short palindromic repeats (CRISPR) and CRISPR interference (CRISPRi) [21] technologies have further expanded the experimental scope of such approaches, including for genes required for growth and viability [22,23].

Protein structure is significantly more conserved than nucleic acid and amino acid sequence [24], making structure the better predictor of function, especially for divergent proteins. However, the proportion of protein sequences with known structures remains small. For example, structure for only ~17% of amino acid residues in the human proteome has been determined experimentally to date [25,26]. That said, recent advances in protein structure prediction provide a viable way to scale structure prediction for annotation purposes for all organisms. One particularly groundbreaking advance was the very recent introduction of Google’s DeepMind AlphaFold2 algorithm [27], which makes use of neural network-based deep learning methods, and together with a similar approach from academia, RoseTTAfold [28], achieves near experimental accuracy. Already, AlphaFold2, in collaboration with EMBL’s European Bioinformatics Institute, predicted the structures of >150,000 proteins from high-interest organisms such as human, Plasmodium falciparum, and Mycobacterium tuberculosis (https://alphafold.ebi.ac.uk). With the source codes openly available, many new sequences can now be confidently predicted, and 58% of residues in the human proteome have already been predicted by AlphaFold2 with high confidence [29], providing a wealth of new starting points for functional annotation.

The emergence of systems biochemistry

From a biological perspective, biochemistry is the operational effector of cellular and organismal physiology. However, biochemistry’s breadth exceeds the experimental scope of available technologies. This is because the chemical diversity of biochemistry within a given cell or organism vastly exceeds that of nucleic acid-encoded genetics [30]. As such, experimental studies of biochemistry required focusing on specific, individual proteins that often began with an activity of interest and ideally led to the identification and study of an individual protein in isolation, dissociated from its native physiologic context [31,32]. The advent of molecular biology made it possible to identify proteins of interest on the basis of their genetic phenotype and sequence but then quickly returned to the need for biochemistry and its dependence on empirical candidate, trial-and-error-based experiments [33,34].

From a conceptual perspective, biochemistry can be experimentally conceptualized along two axes: one that spans the biological diversity of proteins studied and another spanning the chemical diversity of their substrates and/or enzymatic activities (Fig. 1). Classical biochemistry historically began with a single orphan enzyme activity detected in a complex cellular protein extract (Fig 1A). Only cumbersome enzyme activity-guided fractionation to purity, when successful, enabled identifying unique enzyme-substrate pairs. Success was limited to empirical, trial-and-error based purification approaches. Molecular biology enabled starting immediately with potential enzyme-substrate pairs though also with often insufficient and biased outside information to enable predictably meaningful progress.

Figure: Classical and activity-based biochemical analytical landscape.

Schematics illustrating classical (A) and activity-based (B) biochemical experiments. A. In classical biochemistry, assays were limited to a single substrate and only cumbersome protein fractionation allowed identification of unique enzyme-substrate pairs. B. Recent activity-based biochemical approaches leverage the analytical power of mass spectrometry to cover the full metabolite and protein space, ABPP using activity-based probes and ABMP using purified protein. The axes on the top and left indicate the number of sampled proteins and metabolites that can be assayed by each method, the arrows indicate the experimental trajectory of proteomics and metabolomics from global to targeted assays. The ideal experimental space for high-throughput functional protein annotation is indicated in the bottom right quadrant. This prospective, true systems biochemistry approach will likely comprise elements of both ABPP and ABMP and additional approaches such as structural proteomics that can sample many proteins against many metabolites simultaneously.

A major barrier to achieving biologically unbiased systems level studies of biochemistry, in contrast to genetics, has been the analytical inability to systematically traverse the full chemical range of protein and biochemical activity space. While assays with relaxed substrate specificity allowed assigning class-level activities to enzymes (e.g., phosphatase, dehydrogenase), finding physiological substrates remained challenging [35]. Mass spectrometry (MS) has emerged as an analytical technique with high sensitivity towards broad chemical classes that is unbiased towards biology. Thankfully, recent advances in high resolution MS (coupled to high performance liquid chromatography) now allow the detection of thousands of proteins and metabolites in a single run and have made it possible to perform activity-based biochemical experiments covering the entire proteome and metabolome (Fig 1B). In particular, two recent examples of these technologies have been applied to enable biologically agnostic systems-level studies of physiologic biochemistry and new protein annotation.

Annotation by activity-based metabolite profiling

Genetic ablation of genes of unknown function in combination with metabolite profiling is becoming an effective technique to discover gene function [36]. However, the inference of actual catalytic activity from cellular metabolomes remains greatly hampered by metabolic plasticity, leading to pleiotropic effects on the metabolome of genetic mutants. ABMP has helped to overcome this challenge. ABMP involves incubating a purified, usually recombinant, protein of interest with a complex orthologous cellular metabolite extract that functions as substrate library (Fig. 1B) [37].Over the course of incubation, metabolite levels are monitored in an unbiased fashion to identify protein-dependent substrate consumption and product formation [37,38]. Unlike metabolite profiling of intact cells, ABMP thus provides direct, substrate-level enzyme annotation. Since its first reported use in 2006 [37], ABMP has been instrumental in the specific annotation of enzymes from a wide variety of enzyme classes, such as phosphatases [37,39], aminotransferases [16,40], lyases [41], monooxygenases [42–44], and transketolases [45].

In its most basic and exploratory form, ABMP only involves a purified enzyme and a metabolite extract [39,41,45,46]. For example, Shen et al. demonstrated that CLYBL, a conserved mitochondrial enzyme of unknown function, has citramalyl-CoA lyase activity using this approach [41]. Similarly, de Carvalho et al. used basic ABMP to reannotate Rv1248, a protein from M. tuberculosis that was annotated as thiamine diphosphate-dependent alpha-ketoglutarate decarboxylase based on its sequence, as 2-hydroxy-3-oxoadipate synthase [45]. Metabolite extracts are expected to contain enzyme cofactors and substrates, but their concentrations can be at levels that prevent substrate or product detection. To increase the chance of annotation by ABMP, mixtures containing general cofactors such as NADH, ATP, and SAM have been used [37,47]. Similarly, low substrate concentration in metabolite extracts can be overcome by using metabolite extracts derived from gene knockouts that are expected to accumulate substrates immediately downstream of inactivated genes [48].

In cases where enzymes are annotated at the class-level, metabolite extracts can be altered in a more focused manner. In search for the substrates of UDP-glucosyltransferases from grapes, Bönisch et al. used a metabolite extract that was first treated with hydrolases to convert grape glycoconjugates into aglycone substrates for glucosyltransferases [49]. Using this approach, they were able to identify two putative UDP-glucosyltransferases from grape cv Pinot Noir as UDP-glucose:monoterpenol β-d-glucosyltransferases. Class-level knowledge on enzyme function additionally allows the use of labeled co-substrates. Jansen et al., for example, supplemented ABMP reactions of uncharacterized aminotransferases with ¹⁵N-labeled amino acid substrates which allowed the detection of ¹⁵N-amino acid products that remained undetectable using unlabeled substrates [40]. This approach led to the substrate-level annotation of three aminotransferases, a group of enzymes that are notoriously hard to functionally annotate based on sequence. The most extensive use of labeled co-substrates has been in the characterization of orphan cytochrome P450 enzymes. Although the addition of labeled oxygen is not essential [50], the addition of ¹⁸O₂ in combination with isotope detection software allows specific filtering for oxygenation products [51] and has led to the characterization of several cytochrome P450 families [42–44,50].

Finally, ABMP can be applied to confirm the specificity of functional annotation made by other techniques. Black et al., for example, used ABMP to confirm that a putative D-amino acid transaminase only accepted the expected D-amino acids [16]. Liscombe et al. similarly used a metabolite extract from Madagascar periwinkle to confirm the expected activity an S-adenosyl-L-methionine-dependent N-methyltransferase that catalyzes a nitrogen methylation involved in vindoline biosynthesis [52].

Together, the above examples demonstrate the versatility and power of ABMP in activity-based protein annotation. Importantly, the accessible enzyme chemistry is not limited by chemical probes and requires little to no prior knowledge of substrates. The rate of annotation is, however, severely hampered by the need for protein expression and purification. In a technical tour de force, Sévin et al. overcame this limitation by using an automated approach to express 1,275 orphan E. coli proteins and performing high-throughput ABMP [53]. This high-throughput but technically challenging approach led to the prediction of 241 protein functions, of which 12 were validated. Other limitations relate to the fact that metabolite extracts only represent a single metabolic state of the organism and are often limited to polar metabolites, thus underrepresenting the full biological metabolomic complexity. To overcome this limitation, different extractions can be performed on cells grown under various conditions – preferably linked to expression levels of the gene of interest.

At the same time, highly abundant but poor substrates can outcompete low abundant but preferred substrates [54]. ABMP experiments with three different aminotransferases, for example, all led to dominant formation of ketoglutaramate, the keto acid of glutamine. More controlled single-substrate assays later revealed that glutamine is actually a poor substrate and that the dominance of the observed ketoglutaramate was, in fact, an in vitro artefact [40].

Annotation by activity-based protein profiling

Similar to ABMP, the challenge of unbiased, experimental function finding is elegantly solved by an emerging chemoproteomic approach that directly detects biochemical activity through chemical probes - ABPP [55]. In ABPP, a selective chemical probe identifies a specific catalytic mechanism (Fig. 1B). Such probes require three elements: The reactive moiety that recognizes the catalytic mechanism, a selectivity scaffold, and a tag for detection and/or purification. The probe binds the catalytic site of the target protein in a mechanism-dependent way and thus assigns a biochemical function to target proteins proteome-wide. The labeling of the target protein, which is typically covalent, occurs by approaches such as photocrosslinking or in a way reminiscent of suicide substrates. For example, a well-characterized activity-based probe (ABP) modeled on ATP developed initially for the labeling of kinases (ATP-ABP) was designed by appending a reactive acylphosphate in the γ-phosphate position of ATP [56]. The resulting probe is a close mimic of ATP that binds to most ATP binding sites. Instead of transferring phosphates to a substrate, the probe appends an acylphosphate to one of two nucleophilic arginines that are typically found in kinases’ ATP binding sites to coordinate the β- and γ-phosphates of ATP, resulting in C-O bond cleavage and a stable acetamide-protein adduct. Through this mechanism, the probe identifies ATPases in a mechanism-dependent way with only small modifications to the natural substrate. By changing the adenosine scaffold of the probe, the selectivity can be switched to recognize other nucleotide binding proteins [57].

Importantly, ABPP by itself is not a high-throughput approach. ABPP only unfolds its full annotation potential in combination with MS, which can translate biochemical activity into peptide abundance and identifies the labeled proteins against the background of a complex proteome. The potential of combining ABPP with MS for annotation purposes was realized early [58,59]. For example, in a series of studies, the Cravatt lab identified ~80% of predicted human serine hydrolases using a fluorophosphonate probe [59]. Expanding this approach to microbial pathogens, two studies using the ATP-ABP in combination with quantitative shotgun proteomics identified 72 proteins of unknown function (PUFs) in M. tuberculosis [60] and 37 PUFs in P. falciparum [61] as putative ATPases. The same studies also confirmed the existing ATPase annotation of 240 and 141 proteins, respectively. As ABPP is agnostic of sequence, even the most distant enzyme outliers of a family can be detected. In a striking example of this independence from sequence, a serine hydrolase probe identified an enzyme without a conserved Ser that instead used a Thr-based nucleophile for catalyzing lipid hydrolysis [62], defining a distant hydrolase family. This example illustrates how activity-based biochemistry can unearth the most divergent (and convergent) members of a functional class that are invisible to sequence-based methods.

Detail of mechanistic insight and breadth of proteome coverage are tradeoffs in ABPP. For broad coverage, more probes for major classes of enzymes (i.e., hydrolase, oxidase) would be desirable. Even probes that only identify reactive amino acid residues can be of use and identify candidate enzymes. For example, reactive cysteines are characteristic identifiers for many enzyme classes, and their identification can be a first step towards those enzymes’ identification. The utility of such broad labeling was illustrated in studies that used iodoacetamide to identify and rank the reactivity of cellular cysteines [63] and sulfotetrafluorophenyl to detect reactive lysines [64].

Further, for annotation purposes, ABPP should ideally be broad and specific at the same time. One possible solution to this conundrum can perhaps be found in competitive ABPP. The idea of competitive ABPP is that the readout of binding of a more generic probe to a proteome can be refined by competition with ligands characteristic of smaller subfamilies of enzymes. For example, a widely used and very broadly reactive ABP detects serine hydrolases based on the reactivity of fluorophosphonate with serine nucleophiles. Serine hydrolases comprise many large subfamilies of enzymes such that the serine hydrolase annotation itself is of limited use. However, by competing for probe binding with substrates, for example with peptides, the serine hydrolases with protein or peptide substrates such as aminopeptidases might be identified. Similarly, probes with slightly different binding selectivity can parse activities more finely. This approach has been employed for the characterization of metalloproteases [65] and seems particularly useful for characterizing protease activities in general.

Most biochemical assays are targeted, and biochemical function-finding without any a priori information is often a lost cause. ABPP can provide such a priori information to point towards the relevant biochemistry. Once a general enzyme activity is established, the step to more targeted biochemical assays is usually short. For example, Ortega et al. followed a serine hydrolase screen in M. tuberculosis by a simple assay using a generic protease substrate, which identified several likely proteases among the hydrolases [66]. In this way, especially for initial annotation, breadth of probe binding, even at the expense of detail, can be desirable, and while the general assignment of an enzyme family can still be a long way from a full understanding of the protein’s function, it does inform and focus further studies. Every global ABPP-MS experiment is inherently an exercise in annotation: Any PUFs that are detected by the probe can be tentatively assigned to the protein family defined by the probe, and any annotation that differs from that suggested by the probe may warrant another look into the strength of the conflicting annotation.

Applying structural proteomics to annotation

ABPP is bottlenecked by the availability of suitable probes, and ABMP by the availability of catalytically active recombinant protein. Another central limitation of ABPP and ABMP is that non-catalytic protein families such as structural proteins, transcription factors, or transporters are not tractable by these activity-based approaches. What other approaches might access PUFs more comprehensively? Another set of promising systems biochemistry approaches that probe ligand binding to proteins proteome-wide has recently been developed and could address several limitations of the activity-based approaches. A protein’s function is defined to a large part by its ligands, and for an enzyme, by its substrate(s). Ligand binding often affects the structure and/or the stability of a protein, and such changes have now been measured proteome-wide by structural proteomics approaches such as thermal protein profiling (TPP) [67] and limited proteolysis-MS (LiP-MS) [68,69]. A protein that binds a ligand is more (sometimes less) stable towards heat denaturation, altering the protein’s abundance in solution after heating and precipitation. In the case of LiP-MS, ligand binding translates into altered susceptibility to proteolysis, which produces different types and quantities of peptides in the apo- and ligand-bound forms. Conveniently, proteolytic peptides are also the currency of MS, allowing for seamless analysis. This type of global ligand binding analysis appears particularly useful for ligands that cannot readily be turned into probes such as metal ions. While metal binding, for example, may not offer much in terms of annotation, the binding of ligands such as a metabolite, on the other hand, might give substantial clues to function. In fact, in several cases, ligand identification has served as the first step for the full characterization of a PUF’s function [70–72]. Structural proteomics is already beginning to be used to map the full matrix of cellular protein-metabolite interactions [73] and together with the activity-based approaches ABMP and ABPP adds another potentially powerful new systems biochemistry approach to the annotation toolkit.

Concluding remarks

PUFs make up more than one third of proteins even in model organisms, and many more in non-model organisms. Current annotation, initially, is based on inferring function by detecting sequence similarity to proteins with known functions. This approach, while indispensable, is inherently biased toward the annotation of the evolutionarily most conserved functions. New approaches for experimental and high-throughput annotation are needed to understand the vast diversity of protein function, the biology it enables, and to access its potential benefits. Several recent biochemical approaches are poised to make this step towards systems biochemistry. We anticipate that the impact of these new tools, in particular structural and activity-based proteomics and -metabolomics, will be particularly high for protein annotation. There are currently 200 million proteins in Uniprot, a number that grows by ~30 million each year. The global function-finding capabilities of ABPP and ABMP can explore this vast unknown enzyme space with high throughput and identify biochemical functions of even non-canonical and unusual enzymes. They offer opportunities to escape the limitations of current annotation approaches, to accomplish the technical leap that is needed if we want to not only sequence genes but also understand their products’ functions, and to unlock truly new biochemistry.

Outstanding questions.

• Biochemistry has historically centered on single enzymes. How can biochemistry transform into an experimental systems-level discipline and integrate with other systems-level technologies?

• ABPP surveys proteins, ABMP surveys metabolites. Can we simultaneously survey and link proteins and metabolites on a systems-level scale?

• How can we generate ample probes to cover new and/or unknown enzyme space by ABPP?

• Structural proteomics can globally identify ligand binding. How far can ligand binding be exploited for annotation purposes?

• What unknown biology is hidden in the millions of PUFs, and how can we access it?

Highlights.

• The number of genes encoding proteins of unknown function is growing exponentially

• Protein annotation is a central challenge for the molecular life sciences

• Existing computational tools for protein annotation are invaluable but also inadequate to meet the challenge of protein annotation

• Activity-based biochemical approaches such as activity-based protein profiling and activity-based metabolite profiling offer systems-level experimental methods for determining protein activity

• Further development and application of such systems biochemistry approaches are needed to solve the annotation problem

Glossary

Activity-based protein profiling (ABPP)

a chemical proteomics approach to detect enzyme function through activity-based probes, often combined with mass spectrometry.

Activity-based metabolite profiling (ABMP)

a functional metabolomic approach that measures global changes in the metabolome upon addition of recombinant enzyme or a specific enzyme perturbation.

Activity-based probe

a chemical probe that binds to enzymes that have a specific, shared catalytic mechanism and thus identifies the sub-proteome with that activity.

Functional genomics

collective term for genome-wide or systems approaches to identify gene and protein function. Examples include genetic interaction mapping, protein-protein interaction mapping, but also ABMP and ABPP.

Homology- or similarity-based annotation

the transfer of a known protein function to a gene with similar sequence but unknown function.

Limited proteolysis mass spectrometry (LiP-MS)

a proteomic approach by which the proteolytic susceptibility of a proteome upon a perturbation, often ligand binding, is measured. The perturbation induces structural changes in affected proteins that lead to differences in proteolytic patterns, which are then detected by mass spectrometry.

Metabolomics

the global analysis of metabolites and metabolite changes in a biological system.

Neural network-based deep learning methods

a machine learning approach based on algorithms that mimic information processing by neural networks. Deep learning with these multi-layered artificial neural networks can extract patterns from large datasets and predict outcomes.

Protein of unknown function (PUF)

a protein that is expressed but has no characterized function, sometimes also called hypothetical protein.

Synteny

association of genes in a common location or pattern on a chromosome, which can be indicative of linked function.

Thermal proteome profiling (TPP)

a proteomic approach that measures structural changes in a proteome upon perturbation, often ligand binding, based on the change of a protein’s melting temperature and thus change in solubility. Related to LiP-MS, TPP involves precipitation of proteins after heating to separate protein populations with differential solubility.