Massively parallel protein-protein interaction measurement by sequencing (MP3-seq) enables rapid screening of protein heterodimers

Abstract

Protein-protein interactions (PPIs) regulate many cellular processes, and engineered PPIs have cell and gene therapy applications. Here we introduce massively parallel protein-protein interaction measurement by sequencing (MP3-seq), an easy-to-use and highly scalable yeast-two-hybrid approach for measuring PPIs. In MP3-seq, DNA barcodes are associated with specific protein pairs, and barcode enrichment can be read by sequencing to provide a direct measure of interaction strength. We show that MP3-seq is highly quantitative and scales to over 100,000 interactions. We apply MP3-seq to characterize interactions between families of rationally designed heterodimers and to investigate elements conferring specificity to coiled-coil interactions. Finally, we predict coiled heterodimer structures using AlphaFold-Multimer (AF-M) and train linear models on physics simulation energy terms to predict MP3-seq values. We find that AF-M and AF-M complex prediction-based models could be valuable for pre-screening interactions, but that measuring interactions experimentally remains necessary to rank their strengths quantitatively.

Introduction

Synthetic protein binders that can mediate interactions between other proteins or cells have the potential to revolutionize fields from cell therapy (Cho et al., 2018) to synthetic biology (Chen et al., 2020; Chen and Elowitz, 2021; Gao et al., 2018; Groves et al., 2016) and material science (Ben-Sasson et al., 2021; Gonen et al., 2015; Ljubetič et al., 2017b). Early work in synthetic biology often relied on natural interaction domains such as SH3 and PDZ (Kuriyan and Cowburn, 1997). However, such components provide a poor starting point for rationally designing large-scale assemblies due to crosstalk. Consequently, synthetic protein circuits remain much smaller and simpler than biological PPI networks. In order to scale up synthetic protein-based circuits, we need large-scale libraries of modular interaction domains.

Fulfilling this need through rational heterodimer design has yielded promising results, but creating large, totally orthogonal sets of dimers remains challenging. Early rational design work focused on coiled-coil dimers (1×1s): 1 × 1 coil binding is primarily determined by complementary electrostatic and hydrophobic interactions at specific heptad positions. Because the biophysical rules guiding these interactions are relatively well understood and can be captured by predictive models (Fong et al., 2004; Potapov et al., 2015), orthogonal sets of up to six 1 × 1 heterodimers have been generated (Lebar et al., 2020; Thompson et al., 2012). Recently, a set of orthogonal 1×1s (primarily homodimers) was identified in a high-throughput screen (Boldridge et al., 2020). However, the restricted geometry required for 1 × 1 coil interactions limits the number of possible orthogonal interactions. The Baker lab expanded the coiled-coil toolbox by introducing helical bundle heterodimers, with protomers consisting of two helices connected by a hairpin loop, where interactions are determined by designed hydrogen bond networks between the bundles (we will refer to these as 2×2s) (Boyken et al., 2016; Chen et al., 2019). This multi-helix bundle approach increases the alphabet of possible coil interactions and could pave the way to generating larger orthogonal sets. However, reliably minimizing off-target interactions (e.g., due to proteins associating in an unexpected register or orientation) remains challenging as these are not captured in the biophysical design objective to produce 2×2s. A practical alternative is to prepare diverse libraries of de novo 2 × 2 protomers and experimentally measure an all-by-all matrix of interactions. Afterward, an orthogonal set can be extracted (Brodnik et al., 2019).

Many PPI measuring methods have been developed which could be used to run all-by-all PPI screens with varying throughput levels. Mass spectrometry and protein-array based methods can measure PPIs but require laborious protein purification steps (Rao et al., 2014; Smits and Vermeulen, 2016). Yeast- or phage-display methods use next-generation sequencing technologies to increase throughput but are limited to “several-versus-many” screening (Rouet et al., 2018). Alpha-seq, a recent high-throughput method that takes advantage of the yeast mating pathway, overcomes this limitation and allows library-on-library screening (Younger et al., 2017). However, throughput is still limited, and not all proteins correctly fold when displayed on the yeast surface.

Yeast two-hybrid (Y2H) methods are a powerful alternative to surface display for characterizing PPIs (Fields and Song, 1989). In Y2H, one protein of interest is fused to a DNA binding domain (DBD) and the second is fused to a transcriptional activation domain (AD). If the proteins interact, a functional transcription factor is reconstituted and drives the expression of a growth-essential enzyme. Early Y2H approaches tested small numbers of PPIs using plate-based selection, but lab automation and pooling strategies (Luck et al., 2020; Weile et al., 2017) have enabled proteome-scale screens. To further address Y2H scaling issues, high-throughput Y2H(HT-Y2H) and enzyme complementation methods have been developed leveraging next-generation sequencing to read out interaction strength (Diss and Lehner, 2018; Erffelinck et al., 2018; Jin et al., 2007; Rajagopala and Uetz, 2009; Trigg et al., 2017; Weimann et al., 2013; Yachie et al., 2016; F. Yang et al., 2018; J.-S. Yang et al., 2018; Yu et al., 2011). Most of these methods require library construction in E. coli before screening interactions in yeast or rely on yeast mating and thus need separate protein libraries to transform MATα and MATa yeast. High-throughput bacterial two-hybrid assays have been developed, which provide a non-eukaryotic alternative for screening PPIs (Boldridge et al., 2020). Concomitantly with experimental methods, custom workflows for analyzing HT-Y2H data were developed (Banerjee et al., 2021; Velásquez-Zapata et al., 2021).

Recently, deep learning models such as AlphaFold2 (Jumper et al., 2021) and RoseTTAFold (Baek et al., 2021) have been shown to predict protein structures with near experimental accuracy. Structure prediction for proteins with multiple chains is possible, either through the concatenation of target chains with a flexible linker, or with specialized models such as AlphaFold-Multimer (AF-M) (Jumper et al., 2021). AF-M takes in multiple input sequences and has been shown to improve interface predictions over AlphaFold2. Given the speed at which these models operate compared to running Y2H assays, even high throughput ones, determining how these models can be used to pre-screen PPIS or supplement PPI assays is highly attractive. Tools have been developed to run AF-M in an all-by-all manner to assist with PPI pre-screening (Yu et al., 2022), and AF-M error metrics have been demonstrated to be state-of-the-art protein-peptide interaction predictors (Johansson-Åkhe and Wallner, 2022). However, it is unclear how accurately non-structural model derived de novo dimerization can be predicted with AF-M for structurally similar proteins and if AF-M predictions can be used to find orthogonal protein sets.

Here, we introduced MP3-seq, a massively parallel Y2H workflow for measuring PPIs using sequencing. In MP3-seq, the identity of each protein is encoded in a DNA barcode, and the relative barcode-barcode pair abundance before and after a selection experiment serves as a proxy for interaction strength. Plasmids are assembled in yeast through homologous recombination encoding the protein pairs of interest, their associated barcodes, and all other elements required for Y2H experiments. Thus, our experimental workflow bypasses the need for plasmid cloning in E. coli or yeast mating. We first validate MP3-seq using well-characterized coiled-coil heterodimer interactions and synthetic binders for different human BCL2 protein family members. Then, we apply MP3-seq to characterize interactions between rationally designed 2 × 2 and 1 × 2 heterodimers and demonstrate that MP3-seq can measure over 100,000 PPIs in a single experiment. We identify successful designs and use a greedy algorithm to find potentially orthogonal subsets. We delve into the elements expected to confer 2 × 2 interaction specificity by screening variations of a successful 2 × 2 pair with MP3-seq. Finally, we predict complexes for coiled-coil dimers with monomeric AlphaFold2, and versions 1 through 3 of AlphaFold-Multimer. We assess predicted error values (iPTM, pLDDT, pAE) ability to predict coiled-coil interactions and orthogonality, and use said error values and physics-based structure energy terms from Rosetta to train models to predict MP3-seq values and classify interactions.

Results

MP3-seq workflow.

In MP3-seq, all molecular components required for measuring interactions between two specific proteins are encoded on a single plasmid (Figure 1A). A plasmid library is constructed directly through homologous recombination in yeast to measure all possible interactions for a set of proteins. To do so, we transform haploid MATa-type yeast with a mixture of DNA fragments. One of these fragments is a backbone carrying a centromere sequence, the selection marker, a DBD, and an AD. On average, the centromere sequence ensures that only one pair of proteins is expressed in each cell (Scanlon et al., 2009). We use a Cys(2)His(2) zinc-finger domain of the mouse transcription factor Zif268 (Mclsaac et al., 2013) as the DBD and its corresponding promoter to drive the expression of the growth essential enzyme his3. The herpes simplex virus-derived protein domain VP16 is used as the AD. Additional fragments contain one of the proteins of interest and its associated unique barcode separated by a terminator sequence. In the experiments described here, these fragments are ordered as a single oligonucleotide, but for longer proteins a barcode can be added to a protein of interest by PCR. Due to their short length and distinct sequences, we used Tsynth23 and Tsynth27 as terminators in most experiments (Curran et al., 2015).

Figure 1. MP3-seq workflow.

A. Barcoded DNA fragments encoding proteins to be screened, a fragment encoding two terminators, and a backbone fragment bearing a centromere signal, a DBD, and an AD are transformed into yeast cells via electroporation. Homology of fragment 5′ and 3′ ends allows plasmid assembly in yeast. The centromere sequence ensures the expression of one pair of hybrid proteins per cell on average. The cells are split into two populations and one is subjected to growth selection in media lacking histidine. If the hybrid proteins interact, the DBD and AD form a transcription factor that drives HIS3 expression. Therefore, interacting pairs should be enriched in the population post-selection. Cells are lysed, and barcode-containing fragments are amplified and sequenced. B. Enrichment can be calculated for each PPI i using library size normalized read counts with a pseudo count of the minimum detected value per condition. C. After enrichment calculation, each replicate is screened for autoactivators and corrected with Autotune. Replicate pre- and post-selection barcode counts are merged directly with DESeq2 or split into pseudoreplicates and then merged.

After transformation, we select in media without Tryptophan (Trp) to ensure plasmid maintenance. An aliquot of the cells is frozen, while the rest undergo selection in media lacking histidine (His). Plasmid DNA is extracted from pre-and post-His selection cells and the barcode-containing regions are amplified (Figure 1A, right). The barcode-barcode amplicons are sequenced, and their relative enrichment can be calculated from barcode counts to serve as a proxy for interaction strength (Figure 1B). Some proteins’ expression or folding may be impacted by fusion with either the AD or DBD; therefore, we test all proteins fused to both the AD and DBD. Therefore, each PPI screened has two fusion orders (DBD fused to P1, AD fused to P2 or DBD fused to P2, AD fused to P1).

MP3-Seq analysis pipeline.

First, we calculate the enrichment between the pre and post-His selection stages for PPI fusion order. (Figure 1B). Next, we detect autoactivators and replace their post-His selection values with values calculated from the non-autoactivating fusion order (see Autotune in Methods); autoactivation is an error mode in Y2H experiments where high enrichment is observed for a protein for all interaction partners, suggesting non-specific activation of selection marker expression (Shivhare et al., 2021). Typically, this behavior is observed fused to either the AD or the DBD but not in both orientations. Note that if an autoactivator is found its homodimer is not recoverable with Autotune as there is no reverse fusion order His measurement to use for correction.

Following autoactivator removal, enrichment values can be averaged together to obtain interaction strength values. Alternatively, to correct for variation in the read count distribution between experimental replicates resulting from different sequencing depths, selection times, and other experimental factors, we calculate log₂ fold change (LFC) using the DESeq2 package (Love et al., 2014). DESeq2 calculates differential enrichment across multiple replicates and provides Hochberg-adjusted Wald test p-values, identifying PPIs with LFCs significantly different from an LFC of zero.

In some cases, it is desirable to combine MP3-seq measurements for both fusion orderings (e.g., P1-DBD+P2-AD and P2-DBD + P1-AD -> P1-P2). For example, if comparing MP3-seq LFCs to K_ds collected for the pair of interest (P1-P2). The simplest method for merging these measurements is applying an operation to both, such as averaging or taking the max of the values. Here, we use an alternative approach and treat each fusion order as an independent set of measurements, turning them into two pseudoreplicates. These pseudoreplicates are combined with DESeq2 to calculate LFCs (Figure 1B). We refer to these values as pseudoreplicate LFCs (P-LFCs).

MP3-seq benchmarking with orthogonal coiled-coil dimers.

To validate MP3-seq, we screened 144 pairwise interactions between six orthogonal 1×1s in the NICP set (Lebar et al., 2020) (Figure 2A). A pool of 24 DNA fragments where each of the 12 proteins was fused to either the AD or DBD was ordered. Then, all interaction pairs were assembled in a pooled experiment, and interaction strengths were quantified with MP3-seq. We confirmed that selections were occurring as expected by checking pre-His selection barcode counts remained steady over time (Figure S1A), and that applying different levels of 3-Amino-1,2,4 triazole (3-AT), a competitive inhibitor of His3, affected post-His selection counts (Figure S1B–D). LFCs for all orientations were calculated from three experimental replicates, with interactions occurring almost exclusively between intended (on-target) partners (i.e., P1:P2, P3:P4, etc., Figure 2B). Homodimer plasmid assembly is less efficient in MP3-Seq and generally results in lower input barcode counts for these PPIs, but coverage was sufficient for inclusion in the analysis (Figure S1E). This phenomenon is likely due to increased sequence homology, which interferes with DNA fragment order during plasmid assembly. For a more quantitative validation, we correlated MP3-seq LFCs with luciferase expression assay interaction scores in HEK293T cells from (Lebar et al., 2020) and found good agreement (R²=0.83, Figure 2C). Figure 2D represents P-LFC MP3-Seq data for these PPIs as a graph, where each protein is a vertex and edges are significant measured interactions weighted by P-LFC. This graph demonstrates one of the advantages of having adjusted p-values (p_adj) values available to filter data to only higher confidence measurements for understandability.

Figure 2: Validation of MP3-seq with coiled-coil heterodimers.

A. NICP-series 1×1s from (Lebar et al., 2020) with complementary heptads designed to interact shown in like colors. B. MP3-Seq LFC (log₂ fold change) of the NICP-series interactions. All values were calculated from 5 experimental replicates except for those labeled, where labels indicate the number of replicates available. Yellow outlines denote designed on-target interactions. C. Correlation of the on- and off-target NICP-series MP3-seq LFCs with fold activation fluorescence values from (Lebar et al., 2020) The gray bar is the gap separating on- and off-target interactions. Error bars are standard error. D. Filtering NICP-series interactions and E. P, PA, and N series designed coil interactions (Plaper et al., 2021) to include only those with p_adj ≤ 0.05. Line weights correspond to MP3-Seq P-LFCs. F. A designed 1 × 1 and its truncations from (Thomas et al., 2013) G. MP3-Seq enrichment for the A_N and B_N coils and their truncations with three replicates, gray ‘x’ indicates missing interaction. H. Correlation of MP3-Seq P-LFCs with K_d values from (Thomas et al., 2013) for the A_N and B_N truncations, horizontal error bars are standard error, vertical are standard deviation.

In a separate all-by-all experiment with 28 proteins, we screened the NICP 1 × 1 s and two sets of 1 × 1s derived from them (N1, N2, N5-N8, and P5A-P8A) but with increased thermodynamic stability of their on-target interactions (Plaper et al., 2021). In this experiment, we expect related coils to have similar interaction patterns (e.g., P5A can bind to P6A, N6, and P6), which agrees with the interaction graph created from significant MP3-Seq P-LFCs in Fig 2E. The full interaction graph can be found in Fig S2A, where we show minimal off-target cross-talk between NCIP proteins and their variants. We exhibit low correlation with melting temperatures collected for the interactions by (Plaper et al., 2021), but good correlations with split-luciferase and split transcription factor interaction assays for this validation set (Fig S2B–D).

To assess if MP3-seq values provide quantitative information about interaction strength, we measured interactions between a set of varying-length 1 × 1 s designed to span a wide range of K_d values (Thomas et al., 2013) (Figure 2F). We performed three all-by-all replicates; these interactions are not expected to be orthogonal as (Thomas et al., 2013) achieved weaker interactions by truncating the two parent four-heptad binders (A_N⁴, B_N⁴) by a half or full heptad (Figure 2G). The MP3-seq P-LFCs correlated very well with previously measured dissociation constants over approximately three decades (r²=0.94, Figure 2H).

We explored using minimum readcount filtering at different thresholds, and found that filtering brings minimal or no changes to correlation with the (Thomas et al., 2013) K_d values. However, for the NICP series, it does improve correlation (particularly for Spearmans’ rho) (Fig S3A–B). Finally, to examine the performance for weaker interactions, we compared the correlation of enrichment and P-LFC values with the (Thomas et al., 2013) K_d values. We found that enrichments and P-LFC values had high correlations with all data points and similar intermediate correlations when only considering the five weakest interactions (Fig S3C).

MP3-seq benchmarking with BCL2 family binders.

To validate MP3-seq outside of 1 × 1 interactions, we tested a set of proteins previously characterized by biolayer interferometry (Berger et al., 2016) and Alpha-seq (Younger et al., 2017) composed of six homologous proteins from the BCL2 family (Bcl-2, Bcl-xL,BCl-w,Mcl-1,Bfl-1,Bcl-B) and nine de novo designed inhibitors of said homologs. A crystal structure of one of the synthetic binders bound to its target is shown in Figure 3A, while Figure 3B shows the domain organization of the six human proteins. All inhibitors were designed to interact with the BH3 domain binding pocket of the homologs. To better compare with the HT-Y2H method Alpha-seq, a truncated version of Mcl-1 was used.

Figure 3: Validation of MP3-Seq with BCL2 proteins.

A) Colored crystal structure of Bcl-2 and its designed BH3 binding inhibitor from (Berger et al., 2016) (PDB: 5JSN). Binder in blue, Bcl-2 in gray with different domains in the binding region highlighted in color. (B) BH3 binding domain annotations of the six human Bcl homologs measured by (Berger et al., 2016; Younger et al., 2017) with the experimental truncation used by Younger and in this work shown. C) MP3-Seq P-LFCs for inhibitor and Bcl-2 interactions. Yellow boxes highlight intended on-target interactions D) Correlations with K_d measurements from biolayer interferometry. E) MP3-Seq P-LFC (F) and Alpha-Seq distributions for interactions that were undetected by biolayer interferometry due to instrument detection limits (K_d ≥ 25,000 nM). G) Correlation of PUMA peptide and Mcl-1 (t) PPIs with PUMA peptide and full Mcl-1 K_d measurements from (Rogers et al., 2014).

An MP3-seq screen of all BCL2 homologs against all inhibitors is shown in Figure 3C. Version 1 of MP3-seq was used for the first two replicates of this experiment. Unlike version 2, barcodes were inserted into the 3’UTRs of the binders upstream of the terminators and could thus differentially impact mRNA stability and protein levels. However, the inter-replicate correlations suggest consistent interactions between versions (Figure S3A). Only binders beginning with alpha correspond to the final, specific designs of (Berger et al., 2016), while the others are intermediate or failed designs. This can be seen in Figure 3C, with a divide between the largely orthogonal rightmost four columns of the heatmap of MP3-seq P-LFCs and the left, less specific columns.

Our data agree well with dissociation constant measurements obtained from biolayer interferometry by (Berger et al., 2016) (r²=0.61, Figure 3D). We exhibit good agreement with Alpha-Seq percent survival for their low-throughput pairwise assay and high-throughput batched assay on the same interactions (batched r²:0.45, paired r²:0.61, Figure S3B–E). Different measurement conditions can partially explain variation between our results and those published earlier. MP3-seq interactions are measured with proteins expressed in yeast, while biolayer interferometry uses purified proteins, and Alpha-seq displays proteins on the yeast surface.

Some BCL2 inhibitors failed to produce K_d values when measured with biolayer interferometry, likely because the interactions were weak and below detection limits. We examined the undetected (n=11) versus detected interactions (n=43) and found that the mean MP3-seq value of PPIs thatbiolayer interferometry detected was significantly greater than those which were undetected (independent t-test, H₀:μ_detected > μ_undetected) (Figure 3E). Pairwise Alpha-Seq also had the mean of detected interactions significantly greater than undetected ones. However, batched Alpha-Seq, the high throughput version of the assay, did not exhibit this behavior (Figure 3F). Additionally, MP3-Seq correlation improves for batched but not paired Alpha-Seq when including all datapoints (batched r²:0.45, paired r² : 0.69) (Fig S3B–E). Together, these results show that MP3-seq can work with globular proteins in addition to coils and that MP3-seq results agree with those obtained by Alpha-seq and biolayer interferometry. The significant difference between the detected and undetected interactions provides evidence that MP3-seq excels at separating very weak PPIs from others at high throughput.

Finally, we screened a set of PUMA peptides to validate MP3-Seq performance for peptide-protein and disorder-dependent interactions. PUMA is an intrinsically disordered protein that binds to Mcl-1 at the BH3 pocket (Figure 3B), where Mcl-1 binding results in the disordered PUMA BH3 binding motif transitioning to an ordered helix. Peptides of BH3 motif mutants were used to investigate the effects of helicity degree on this IDP-dependent interaction (Rogers et al., 2014). As part of this investigation, stopped-flow fluorescence techniques were used to measure the Kd values of a set of PUMA peptides interacting with the full Mcl-1 protein. We screened a subset of these peptides against our Mcl-1(t) protein and found a good level of correlation (r₂:0.66) (Figure 3G).

Large-scale assay of designed heterodimers.

To explore the feasibility of using MP3-seq for large-scale screening, we prepared a library of designed heterodimers (DHDs) following (Chen et al., 2019). On-target pairs were designed as three- or four-helix bundles with buried hydrogen bond networks (HB-Nets) connecting all helices and then split into a helix-turn-helix hairpin and a single helix (1×2s) or 2×2s. Initially, we chose 100 designs for testing (DHD1, Figure 4A), later adding 52 more designs in a second round of experiments (DHD2, Figure 4A). As controls, we selected nine previously tested 2 × 2s (Chen et al., 2019). We also included 35 pairs of binders derived from these controls by modifying the hairpin loops or truncating the helices (DHDO, Figure 4A). Finally, we added a series derived from a common parent 2 × 2 through truncating heptads (mALb, Figure 4A). For each on-target pair, one protomer is designated “A” and the other “B”. Interactions between and within these groups were tested in MP3-seq experiments of varying sizes. We performed two replicates of an all-by-all screen including DHD0, DHD1, DHD2, and mALb proteins as well as the BCL2 homologs and cognate binders described above, resulting in a matrix of 337 × 337 = 113,569 interactions. We also performed three replicates of a screen using only DHD1 with a subset of DHDO and one with DHDO, DHD2, and mALB. MP3-seq version 1 was used for most experiments in this section. Still, the experiment highlights the large experimental scales achievable with MP3-seq and the inter-replicate correlations suggest that the data is internally consistent (Figure S2 B–C).

Figure 4. High throughput screening of designed heterodimers.

A. Summary of the synthetic binder types. B. Successful designs for the three main sets. Gray bars are possible pairs for the set, and colored bars indicate the number of successes. C. Top 5 successful designs from each set by P-LFC. Yellow boxes show on-target interactions. D. From left to right, pre-DUET positive, p_adj ≤ 0.01 P-LFC PPI network for the DHD1 and DHD0 subset, and the network at DUET iterations 50, 75, and 113. E. Pre-DUET network for the DHD0, DHD2, and mALb sets, and the network at DUET iterations 10, 15, and 37 (final). F. Pre-DUET network for all designs and the network at DUET iterations 50,75, and 114. G. Orthogonality gaps of the DUET final networks without significance filtering. Left and right dashed lines show the MP3-seq orthogonality gaps for BCL-2 and final inhibitors and the NCIP-series, respectively. Yellow squares correspond to half of the starting networks remaining. H. DUET final networks reduced to half their final iteration size.

We applied autotune corrections and symmetrized the read counts to calculate P-LFCs for all possible pairs. On-target interactions, where the A and B protomers were designed to interact, were evaluated for success. In this case, we defined a ‘success’ as any PPI with a p_adj ≤ 0.01 and a P-LFC ≥ 4. The number of successes out of the total possible on-target PPIs in each design set can be seen in Figure 4B. Every design set had an approximately 20% success rate in the largest screen, consistent with past 22% success rates for alpha-helical bundles (Ljubetič et al., 2017a). The top 5 P-LFC on-target interactions for each design category are shown in Figure 4C. The DHD pairs have strong on-target interactions. However dimers are not necessarily limited to the designed on-target pairs, even when examining only the PPIs of the A and B protomers of the top five designs.

Finding orthogonal subsets.

For all four sets of designs (DHD1–3, mALB), we detected a large fraction of strong off-target interactions (e.g., A1 and B2 of DHD1 instead of A1 and B1). For example, in the all-by-all screen with all of DHD1–3 and mALB, only 33 of 914 total PPIs with p_adj ≤ 0.01 and P-LFC ≥ 4 were on-target.

Given our many measured strong interactions, we asked if we could extract potential orthogonal PPI subsets from our data. To this end, we used significant (alpha = 0.01), positive P-LFC values to construct a weighted undirected graph, as in Figure 2D–E. This approach allows us to rephrase the problem as finding a graph of degree one vertices without self-edges which maximizes the sum of the remaining edges. To do this, we developed a simple scoring function that rewards graphs based on existing orthogonal edges or those over a desired orthogonality gap and punishes graphs for non-orthogonal edges. This scoring function was used in a greedy graph reduction method, Deleting Undirected Edges Thoughtfully (DUET), which removes a vertex and its associated edges each iteration until a 1-regular graph remains (see Methods). We applied DUET to our different data sets (the number of degree one vertices and score of total graphs per algorithm step can be seen in Fig S5A–B). For DHD1 and the DHDO subset, we went from 2,001 edges between 202 vertices to 18 DUET pairs (Figure 4D), for the DHD0/DHD2/mALb data from 279 edges between 85 vertices to 11 DUET pairs (Figure 4E), and for all designs (DHD0–2,mALb) from 1,562 edges between 270 vertices to 36 DUET pairs (Figure 4F). Of these, 2, 2, and 4 DUET pairs were on-target for DHD0/DHD1, DHD0/DHD2/mALb, and all designs, respectively.

The DUET final results are only orthogonal if all non-highly significant interactions are considered non-interacting. As this may not be the case, we used all P-LFC > 0 between DUET pair protomers regardless of significance for a more conservative analysis. First, we removed interactions with protomers for which the DUET pair P-LFC was lower than the highest non-DUET pair P-LFC. Then, we reduced the remaining DUET pairs by removing whichever pair had the largest non-DUET P-LFC one by one (Figure 4F). When reduced to half the starting DUET pairs (squares in Figure 4F), the orthogonal sets shown in Figure 4G are left, which all have orthogonality gaps comparable to the NICP set in Figure 2 and the BCL2-inhibitor pairs in Figure 3. An additional undesirable behavior for these potentially orthogonal sets would be if DUET is biased toward selected proteins that have missing interaction data between one another (and, therefore, fewer edges in the graph). To determine if this was the case, we ran permutation tests where we counted the number of missing interactions between the proteins in the initial DUET results and randomly sampled protein sets of the same size. We did not find that DUET results had a significantly higher number of missing interactions (Fig S5C).

Elucidating the rules of specific helix-loop-helix binding.

High-throughput, high-quality data can be used to probe novel design rules to improve protein design. We demonstrate how MP3-Seq can contribute towards a better understanding of binding specificity in designed 2 × 2 pairs by testing variants of different lengths and with a different number of buried hydrogen bond networks. Our target pair, mALb8, is a high-affinity 2 × 2 heterodimer with three HB-Nets between its A and B protomers (Figure 5A). This pair was the basis of the mALb set introduced earlier.

Figure 5: Systematic exploration of length and hydrogen bond network presence on helix-loop-helix binding.

A. The mALb8 helix-loop-helix heterodimer with HB-Nets 1 and 2 shown. B. mALb8 and the two shorter variants (T1; one helical turn shorter, T2; two helical turns shorter). C. Log fold change (LFC) of the truncated mALb8 interactions. Gray X squares indicate insufficient reads. We observed a markedly non-linear response to the length of the binder. T2 variants do not bind. All PPIs had four replicates, except those marked with their replicate number. D. An example of HB-Net removal using the small and large hydrophobic replacement protocols for the first HB-Net E. LFCs of mALb8 with and without removed HB-Nets. S and L designate the replacement protocol used, while R₁, R₂, and R₁₂ denote if the first, second, or both HBNets were replaced. Gray X indicates insufficient reads. The yellow boxed regions show that two HB-Net mismatches are enough to specify orthogonality. Hydrophobic residues alone cannot confer specificity.

First, two mALb protomer truncations were designed, removing one or two turns from the end of each alpha helix (Figure 5B). Truncation 1 (T1) versions are one turn shorter than the original monomers, while Truncation 2 (T2) versions are two full turns shorter (Figure S6A). The original binders and the truncations were screened in an all-by-all screen with MP3-Seq and LFCs were calculated from four replicates. It can be seen from Figure 5B that truncating the binders by two turns (T2) eliminates binding. The relationship between length and binding affinity is highly non-linear, as binding disappears once the minimum length is reached. We can use this data to infer the shortest possible length (3.5 heptades) for 2 × 2 binders containing two buried HB-Nets. This pattern is similar to that seen for the A_N and B_N single helix truncation experiments in Figure 1D–E.

To probe the specificity conferred by the buried HB-Nets in the 2 × 2 binders designed with the approach from (Chen et al., 2019), we have removed either the first, the second, or both buried HB-Nets from the mALb8 dimer. Removal was done by replacing HB-Net amino acids in both A and B protomers with hydrophobic residues that can not form hydrogen bonds. Two Rosetta-design replacement protocols were used: one using smaller (S) and one larger (L) hydrophobic residues for replacement. The networks are shown in Figure 5D, while all replaced designs can be found in Figure S6B. We observe that the designs largely do not homodimerize, as interactions between various A and B protomers are low. These results show that at least two hydrogen bond network mismatches are needed to confer specificity and orthogonality (Figure 5E, boxed areas). For example, A_R1 and B_R2 binding would bury half of the second HB-Net on protomer A and half of the first HB-Net on protomer B, while A_R12 and B_R2 binding buries only the first HB-Net on chain B. The binding data suggest that pairs with one HB-Net mismatch can still exhibit partial binding affinity: for example, B_R12(s) weakly binds to both A_R1 and A_R2. Additionally, hydrophobic residue size alone is not sufficient to confer orthogonality. For example, both B_R1(S) and B_R1(L) bind to both A_R1(L) and A_R1(S), regardless of the size of hydrophobic amino acid residues. The all-by-all screen of original, truncated, and HB-Net removed mALb proteins can be found in Fig S6C, which shows that the first truncation mALb proteins have similar binding patterns as the originals, but that the second continues to have eliminated binding.

Predicting orthogonal coiled-coil complexes and training ridge regressors and classifiers.

To assess AlphaFold’s ability to predict interactions, we predicted complexes for all 61 × 1 NICP pairs displayed in Fig 2B. We used AF-M v1-3 and the monomeric version of AlphaFold2 (with chains connected by a virtual linker (Mirdita et al., 2022)). Five AF2 models were run per dimer, using both orders of feeding in input sequences. We compared error metric (pIDDT, PAE, etc.) correlation with MP3-Seq LFC values and onand off-target classification for the complex predictors. The monomeric AF2 and V1 AF-M performed poorer than later AF-M versions (Fig S7A–B). The interface Predicted TM score (iPTM) score averaged across our five predicted complexes was among the best-performing metrics. The AF-M v3 averaged across all 5 AF models for the NICP interactions can be seen in Fig 6A. The increase in performance for both correlation with LFC values and on/off target classification can be seen in Figure 6B, where v3 improves over V2 in both cases. However, when comparing the performance of the structure predictor error metrics with the 1 × 1 binding predictor iCipa (Boldridge et al., 2020), AF-M still is underperforming (particularly in the classification task). Other 1×1 binding predictor performances can be seen in Fig S7D, which similarly outperform AF-M metrics.

Figure 6: Predicted complex error for coiled dimers.

A. Average predicted interface TM (iPTM) per complex from Alphafold2 multimer v3 (AF-M v3) for the NICP 1×1s. On-target interactions are marked with yellow squares. B. NICP correlation of average iPTM per complex with experimental (Fig 2A) LFC values (top) and area under precision recall curve (AUPRC) of classifying on- and off--target interactions (bottom). The 1×1 binding predictor iCipa (Boldridge et al., 2020) is shown for comparison. C. Average iPTM per complex with AF-M v3 for a subset of the mALb8 interactions. D. Correlation of average iPTM with mALb8 interaction LFCs (Fig 2E). The 1 × 1 predictor iCipa is shown for comparison, and is unable to generalize to 2 × 2 coils unlike AF-M.

We then used the best-performing complex prediction models, AF-M v2 and v3, to predict structures for the truncated, mutated, and original 2 × 2 mALb8 complexes examined in Fig 5. There is no apparent increase in error metric LFC correlation with AF version numbers for this set as in the NICP set (Fig S7F–G). The AF-M v3 averaged across all 5 AF2 models for a subset of mALb8 interactions can be seen in Fig 6C. Heterodimeric interactions have higher confidence than homodimeric complexes, but the effect of HB-Networks on orthogonality is not captured in detail. Although AF-M v3 iPTM captures some binding rules investigated for the mALb8 dimer, the iPTM values it does not have a clear advantage over V2 as in the NICP series. The better generalization ability of structure predictors can be seen in Fig 6D, where both AF-M versions outperform the 1×1 model iCipa when applied to 2 × 2 interactions. Therefore, while AF-M metrics may not outperform specialized binding predictors on the domain they were trained on, they have undoubtable cross-domain generalization abilities.

Encouraged by the demonstrated generalization abilities of AF-M, we set out to determine if the predicted structures can be used to predict LFC values and classify strong interactions. Rosetta was used to collect physics-based metrics (energy of interaction, surface of the interface, etc.) for each simulated dimer complex. Agglomerative hierarchical clustering was used to reduce the number of multicollinear features between the collected energy terms and AF error metrics (Figure S8). Linear least square ridge regression models and logistic ridge regression classifiers were trained with on feature sets of decreasing sizes (see Methods).

We used two train-test approaches to investigate ways these models might be applied. For the first, we ranked all data points by their padj values and used them to partition the dataset into high-quality test set interactions and a mix of high and low-quality training set interactions. This approach was chosen to assess the ability of models using AF-M complex features to fill in missing interactions in an assay. An example test set can be seen in Fig 7A.

Figure 7. Training simple models on predicted dimer features to predict MP3-Seq values.

A. Examples of the two different train-test splitting techniques: held out interactions, where half of the low MP3-seq p_adj LFCs are used for the test set (top) and held out proteins, where all interactions involving selected proteins are part of the test set (bottom). The dashed line shows the performance of the best held-out interaction model with the same number of training points as the held out protein models. Red interactions are train/validation, and non-red are test. Held out interaction (top) and held out protein (bottom) results, with features used for model training split into the number of Rosetta (R) and AF-M metrics (AF) used, for (B) predicting the NICP LFC values, (C) classifying the NICP interactions as on- or off-target, and (D) The mALb8 original, truncated, and HB-Net removed interaction LFCs. The dashed lines indicate the performance of the best held out interaction model with similarly-sized train/test set splits. E. (Top) Performance of predicting the LFC values of mALb interactions using the AF-M models with only AF inputs and comparing them with iPTM. (Bottom) Predicting the NICP LFC values with the AF-M V3 LFC prediction model.

In the other approach, proteins were selected, and all interactions involving those proteins were assigned to the test set. The protein-based split was used to evaluate how AF-M complex-based models could predict interactions involving new proteins instead of filling in missing interactions between known proteins (see Methods for how held-out proteins were selected). Multiple train/test sets were created for this split by holding out different protein sets to get a distribution of test performances. An example test set can be seen in Fig 7A.

The NCIP regression models performed similarly to iCipa for the held out interaction task and had gains over using iPTM alone for prediction (Fig 7B, top). In particular, the V2 and V3 AF-M models performed well, while the monomeric AF and AF-M V1 models were less successful, particularly when smaller training sets were used (Fig S9A). A similar trend was seen for the held out interaction classification models, with the main difference being that the V2 iPTM provided an excellent classifier for this test set, and AF-M V2 performed better than V3 in modeling (Figure 7D, top). The number of Rosetta (R) and AlphaFold (AF) features each model used can be seen in the tables at the bottom of Fig 7B–D, and more features yielded better models (notably when decreasing training set sizes, see Fig S9A–B). All regression models dropped drastically in performance in the held-out protein task. However, this performance decrease can be partially attributed to a smaller training set, where the dashed line shows the performance of the best similar training set size held out interaction model (Figure 7B). An additional thing to note is that in some cases, iPTM outperformed models using only AF error metrics on the test-set due to overfitting on the training set (Fig S10C). The classification task’s performance stayed relatively high between the two tasks even when the training set size was reduced (Fig S10A–B).

In contrast, the more complicated mALb interactions did not show clear trends in model performance with AF-M versions (Fig 7D). It had a much smaller gain in performance in the held-out interaction task between the AF-only models and those with R and AF features. However, interestingly when dropping the training set size, there were gains in model performance, likely due to the held out interaction test set initially consisting of only high LFC interactions (Fig S11). We then used our best AF-feature-only held out interaction NICP models to predict LFC values for the MALb interactions and vice versa (Fig 7E, Fig S9C–F, Fig S11B). We found in both cases that the models performed worse than iPTM when applied to coiled-coil interactions different from those they were trained to model.

Discussion

Here, we introduce MP3-seq, an easy-to-use HT-Y2H method that can measure pairwise PPIs in a single yeast strain. In MP3-seq, the identity of each protein is encoded in a DNA barcode, and the relative barcode-barcode pair abundance before and after a selection experiment serves as a proxy for interaction strength. We also developed a data analysis workflow based on DESeq2 that makes it easy to merge replicates, remove autoactivators, and identify statistically significant interactions. The MP3-seq workflow could be further generalized in future work by adding autoactivator screening or adjusting the selection scheme to eliminate autoactivators (Shivhare et al., 2021). MP3-seq could also be combined with protein stability and expression assays to confirm that proteins tested in interaction screens are folded correctly and that a negative interaction measurement is not due to reduced protein levels. In particular, it is easy to imagine integrating MP3-seq with other growth selection-based workflows such as Stable-seq (Kim et al., 2013) or high-throughput protease assays (Rocklin et al., 2017). We expect stability screening to be more important for human protein than for the synthetic binders tested here, which are generally designed to have very stable folds.

We validated MP3-seq using several sets of proteins for which interactions had been previously characterized. In particular, we used a family of human pro-apoptotic proteins BCL2 and their de novo designed inhibitors, a set of peptides binding to MCL-1, and three different sets of coiled-coil peptides. We found quantitative agreement between our results and those reported previously. We then applied our method to characterize interactions in a pool of de novo 2 × 2 heterodimers that contain buried hydrogen bond networks. We showed that our approach could scale to measuring more than 100,000 interactions in a single experiment. Our computational workflow enabled us to identify potential orthogonal subsets, at various orthogonality gaps, within larger all-by-all screens.

Moreover, by screening interactions between protomers with truncations and modified hydrogen bond networks, we probed design rules for 2 × 2 binders with two buried HB-nets. Firstly, we showed that the minimum length for strong binding is 3.5 heptads and that binding affinity drops sharply for shorter designs. Next, we showed that at least two hydrogen bond network mismatches are needed for orthogonality, thereby setting minimum requirements for future sets of orthogonal 2 × 2 binders with buried HB-nets.

Finally, we assessed the structural prediction models AlphaFold 2 and AlphaFold-Multimer to predict coiled-coil interactions. On the set of 1 × 1 dimers, we found that a sequence-based predictor (iCipa) correlated better with experimental LFC values than any single AF metric. However, AF-M metrics performed better than iCipa for the 2 × 2 dimers. Next, we tested if we could use a combination of AF2 and Rosetta metrics to improve AF2 interaction predictions. We found that regression models could be trained to fill in missing interaction LFCs from an all-by-all screen for the 1 × 1 dimers and, to a lesser extent, the 2 × 2 dimers. However, regression models performed poorly when test sets were selected to mirror the task of predicting interactions of ‘new’ proteins not present in the PPI training sets. In contrast, classification models performed well on filling in test interactions selected from the all by all matrix and classifying PPIs as strong, on-target interactions versus weak. None of our models performed well when predicting PPIs for the other coiled-coil group (predicting 2 × 2 interactions with the 1 × 1 models and 1 × 1 s with the 2 × 2 models). Our results suggest that AF-M-predicted complex energy values may be used with minimal data to differentiate strong from very weak interactions for simple proteins, with the potential to expand to predicting interactions for new proteins unseen in the assay. Therefore, AF-M and AF-M complex prediction-based models could be valuable tools for qualitatively pre-screening interactions, but it remains necessary to measure interactions experimentally to determine their relative quantitative strengths.

We believe that the increased scale and streamlined workflow of MP3-seq will further accelerate the adoption of HT-Y2H methods and facilitate their use in additional applications. For example, high-throughput data collected with MP3-seq could be used to train predictive models of PPIs. Such models could be combined with generative algorithms to iteratively design heterodimers with desirable properties or with neural network interpretation to uncover determinants of binding specificity. Moreover, MP3-seq could be applied to characterize human PPIs or to screen variants’ impact on PPIs at high throughput.

Data and code availability

Barcode count data and scripts necessary to recreate the MP3-seq pipelines, analysis, and models can be found at https://github.com/Seeliglab/MP3-DUET-AUTOTUNE