We first characterized the historical evolution of PPI specificity in the BCL-2 family using ancestral protein reconstruction (Figure 2—figure supplement 2). We inferred the maximum likelihood phylogeny of the family, which recovered the expected sister relationship between the metazoan BCL-2 and MCL-1 classes (Figure 2C, Figure 2—figure supplement 3). We then reconstructed the most recent common ancestor (AncMB1) of the two classes – a gene duplication that occurred before the last common ancestor (LCA) of all animals – and 11 other ancestral proteins that existed along the lineages leading from AncMB1 to human BCL-2 (hsBCL-2) and to human MCL-1 (hsMCL-1) (Supplementary file 1).

We synthesized genes coding for these proteins and experimentally assayed their ability to bind BID and NOXA using a proximity-dependent split RNA polymerase (RNAP) luciferase assay (Figure 2—figure supplement 4; Pu et al., 2017b). AncMB1 bound both BID and NOXA, as did all ancestral proteins in the MCL-1 clade and hsMCL-1 (Figure 2C, Supplementary file 1). Ancestral proteins in the BCL-2 clade that existed before the LCA of deuterostomes also bound both BID and NOXA, whereas BCL-2 ancestors within the deuterostomes bound only BID, just as hsBCL-2 does. This reconstruction of history was robust to uncertainty in the ancestral sequences: experiments on ‘AltAll’ proteins at each ancestral node – which combine all plausible alternative amino acid states (posterior probability > 0.2) in a single ‘worst-case’ alternative reconstruction – also showed that BID specificity arose within the BCL-2 clade (Figure 2—figure supplement 4, Supplementary file 2).

To further test this inferred history, we characterized the coregulator specificity of extant BCL-2 class proteins from taxonomic groups in particularly informative phylogenetic positions. Those from Cnidaria were activated by both BID and NOXA, whereas those from protostomes and invertebrate deuterostomes were BID-specific (Figure 2C, Figure 2—figure supplement 4, Supplementary file 1). These results corroborate the inferences made from ancestral proteins, indicating that BID specificity evolved when the ancestral ability to bind NOXA was lost between AncB2 (in the ancestral eumetazoan) and AncB4 (in the ancestral deuterostome).

To rapidly evolve BCL-2 family proteins to acquire the same PPI specificities that existed during the family’s history, we developed a new PACE system (Esvelt et al., 2011; Figure 3A–B, Figure 3—figure supplement 1). Previous PACE systems have evolved binding to new protein partners using a bacterial 2-hybrid approach (Badran et al., 2016), but evolving PPI specificity requires simultaneous selection for a desired PPI and against an undesired PPI. For this purpose, we used two orthogonal proximity-dependent split RNAPs that recognize different promoters in the same cell and – if reconstituted by a PPI – activate transcription of positive and negative selectable markers. Specifically, the N-terminal fragment of RNAP was fused to the BCL-2 protein of interest and encoded in the phage genome, and two C-terminal RNAP fragments (RNAPc), each fused to a different BCL-2 coregulator, were encoded on host cell plasmids. One RNAPc is fused to the selected-for coregulator and drives expression of an essential viral gene (gIII) when reconstituted by binding to the BCL-2 protein; the other RNAPc, fused to the counter-selected coregulator, drives expression of a dominant-negative version of gIII (Pu et al., 2017a). Phage containing BCL-2 variants that bind the positive selection protein but not the counterselection protein produce infectious phage. After optimizing this system, we used activity-dependent plaque assays and phage growth assays to confirm that it imposes strong selection for the PPI specificity profiles of extant hsBCL-2 and hsMCL1 (Figure 3D).

The simplicity of this platform allowed us to drive extant and reconstructed ancestral proteins to recapitulate or reverse the historical evolution of the BCL-2 family’s PPI specificity in multiple replicates in just days, without severe experimental bottlenecks. Three proteins that bound both BID and NOXA – hsMCL-1, AncM6, and AncB1 – were selected to acquire the derived BCL-2 phenotype, retaining BID binding and losing NOXA binding. Conversely, hsBCL-2, AncB5, and AncB4 were evolved to gain NOXA binding, reverting to the ancestral phenotype (Figure 3C and E, Figure 3—figure supplement 2). For each starting genotype, we performed four replicate experimental evolution trajectories (Supplementary file 3). Each experiment was run for 4 days, corresponding to approximately 100 rounds of viral replication (Esvelt et al., 2011). All trajectories yielded the target PPI specificity, which we confirmed by experimental analysis of randomly isolated phage clones using activity-dependent plaque assays and in vivo and in vitro binding assays (Figure 4A–B, Figure 3—figure supplement 2, Figure 3—figure supplement 3). As in prior PACE experiments, variation in the selected phenotype was observed among individual phage isolates within the final populations (Dickinson et al., 2013), presumably because of large populations, high mutation rates, and/or inadequate time for fixation.

We used deep sequencing to compare the sequence outcomes of evolution across trajectories initiated from the same and different starting points (Figure 4—figure supplement 1). Necessity was almost entirely absent. Across all trajectories, 100 mutant amino acid states at 75 different sites evolved to frequency > 5% in at least one replicate (Figure 4C–D, Figure 4—figure supplement 2, Supplementary file 4). Of these acquired states, 73 appeared in only a single trajectory, and only four arose in more than one replicate from multiple starting points (Figure 4E, Figure 4—figure supplement 3). When selection was imposed for binding to both BID and NOXA, no states were predictably acquired in all trajectories from all starting points. The only mutation universally acquired under any selection regime was a nonsense mutation at codon 271, which was acquired in all trajectories selected for BID specificity, but experimental analysis of this mutation shows that it has no detectable effect on coregulator binding (Figure 4—figure supplement 4).

Both chance and contingency contributed to this pervasive unpredictability. Pairs of trajectories launched from the same starting point differed, on average, at 78% of their acquired states, indicating a strong role for chance. Pairs that were launched from different starting points (but selected for the same PPI specificity) differed at an average of 92% of acquired states, indicating an additional role for contingency.

These starting points are separated by different amounts of evolutionary divergence, so to understand the extent of contingency over the timescale of metazoan evolution, we compared trajectories launched from AncB1 to those launched from hsMCL1 (the two most distant genotypes that were selected for BID specificity). Of 34 states acquired in these experiments, only three occurred in at least one trajectory from both starting points. Of 40 states acquired in trajectories launched from AncB4 and hsBCL-2 (the two most distant proteins that were selected to gain NOXA binding), only one occurred in any trajectories from both starting points. Together, contingency generated across long phylogenetic timescales and chance therefore make sequence evolution in the BCL-2 family almost entirely unpredictable.

These experiments indicate an almost complete lack of necessity in the evolution of PPI specificity in PACE. To gain insight into the extent of necessity in the historical evolution of BCL-2 PPI specificity, we asked whether substitutions that occurred during the phylogenetic interval when NOXA binding was lost (between AncB1 and AncB4) were either repeated or reversed during PACE trajectories to lose or regain NOXA binding from any starting point (Figure 4F, Figure 4—figure supplement 5, Figure 4—figure supplement 6). In PACE experiments to lose NOXA binding from proteins that initially bound both peptides, none of the acquired states recapitulated substitutions from the branch on which NOXA binding was historically lost. In PACE experiments to reacquire NOXA binding from proteins with BCL2-like specificity for BID, only two states reversed historical substitutions that occurred on that branch. Both of these reacquisitions occurred in PACE trajectories launched from AncB4, the immediate daughter node of this branch, suggesting that in other proteins, contingency accumulated over phylogenetic time restricted their accessibility. Furthermore, both of these states were acquired in only a subset of trajectories from AncB4, indicating a role for chance even from this starting point. Some substitutions that occurred during other historical intervals were recapitulated or reversed during PACE trajectories, indicating that these states are compatible with BCL-2 family protein functions, but these substitutions could not have contributed to historical changes in PPI specificity, which remained unchanged on these branches. Our experiments therefore suggest strong effects of chance and contingency in the historical evolution of BCL-2’s derived PPI specificity.

We next sought to directly quantify the relative effects of chance and historically generated contingency on sequence outcomes in our experiments. We analyzed the genetic variance – defined as the probability that two variable sites, chosen at random, are different in state – within and between trajectories from the same and different starting genotypes. To estimate the effects of chance, we compared the genetic variance between replicates initiated from the same starting genotype (V_g) to the within-replicate genetic variance (V_r). We found that V_g was on average 30% greater than V_r, indicating that chance causes evolution to produce divergent genetic outcomes between independent lineages even with strong selection for a change in function (Figure 5A). We quantified contingency by comparing the pooled genetic variance among replicates from different starting genotypes (V_t) to that among replicates from the same starting genotype (V_g). Contingency’s effect was even larger than that of chance, increasing V_t by an average of 80% across all pairs of starting points compared to V_g when selecting for a new function. Together, chance and contingency had a multiplicative effect, increasing the genetic variance among trajectories from different starting genotypes (V_t) by an average of 2.4-fold compared to the genetic variance within trajectories (V_r). The effects of chance and contingency were not significantly different between PACE experiments in which protein interactions were gained and those in which they were lost (Figure 5—figure supplement 1).

The preceding analyses do not account for phylogenetic structure or the extent of divergence between starting points. We therefore assessed how chance and contingency changed with phylogenetic distance using linear regression (Figure 5B, Figure 5—figure supplement 1). We found that the effect of contingency on genetic variance increased significantly with phylogenetic divergence among starting points. The effect of chance did not increase with divergence, but the combined effect of contingency and chance increased even more rapidly than contingency alone because the total impact on genetic variance of these two factors is multiplicative by definition.

We next compared the impact of contingency to that of chance as phylogenetic divergence increases. On the timescale of metazoan evolution, contingency’s effect (an increase in genetic variance by about 100%) was three times greater than that of chance when evolution was launched from extant starting points whose LCA was AncMB1, near the base of Metazoa (Figure 5B). The combined effect of chance and contingency on this timescale was a 3.2-fold increase in variance among single trajectories launched from these starting points. Even across the shortest phylogenetic intervals we studied, contingency’s effect was larger than that of chance, although to a smaller extent. Taken together, these data indicate that contingency, magnified by chance, steadily increases the unpredictability of evolutionary outcomes as protein sequences diverge across history.

Contingency is expected to arise in our experiments if historical substitutions (which separate ancestral starting points) interact epistatically with mutations that occur during PACE, causing the mutations that can confer selected PPI specificities to differ among starting points. To experimentally test this hypothesis and characterize underlying epistatic interactions, we first identified sets of candidate causal mutations that arose repeatedly during PACE replicates from each starting genotype. We then verified their causal effect on specificity by introducing only these mutations into the protein that served as the starting point for the PACE experiment in which they were observed and measuring their effects on BID and NOXA binding. We found that all sets were sufficient to confer the selected-for specificity in their ‘native’ background (Figure 6A, B).

We then introduced these mutations into the other starting proteins that had been subject to the same selection regime and performed the same assay (Figure 6A,B). Eleven of 12 such swaps failed to confer the PPI specificity on other proteins that they conferred in their native backgrounds. These swaps compromised binding of BID, failed to confer the selected-for gain or loss of NOXA binding, or both. The only case in which the mutations that conferred the target phenotype during directed evolution had the same effect in another background was the swap into AncB4 of mutations that evolved in AncB5 – the most similar genotypes of all pairs of starting points in the analysis. Contingency therefore arose because historical substitutions that occurred during the intervals between ancestral proteins made specificity-changing mutations either deleterious or functionally inconsequential when introduced into genetic backgrounds that existed before or after those in which the mutations occurred.

To characterize the timing and effect of these epistatic substitutions during historical evolution, we mapped the observed incompatibilities onto the phylogeny (Figure 6C). We inferred that restrictive substitutions evolved on a branch if mutations that arose during directed evolution of an ancestral protein compromised coregulator binding when swapped into descendants of that branch. Conversely, we inferred that permissive substitutions evolved on a branch if mutations that arose during directed evolution compromised coregulator binding when swapped into more ancient ancestral proteins.

We found that both permissive and restrictive epistatic substitutions occurred on almost every branch of the phylogeny and affected both BID and NOXA binding. The only exception was the branch from AncB4 to AncB5, on which only restrictive substitutions affecting NOXA binding occurred. This is the branch immediately after NOXA function changed during history; it is also the shortest of all branches examined and the one with the smallest effect of contingency on genetic variance (Figure 5B). Even across this branch, however, the PACE mutations that restore the ancestral PPI specificity in AncB4 can no longer do so in AncB5. These results indicate that the paths through sequence space leading to historical PPI specificities changed repeatedly during the BCL-2 family’s history, even during intervals when the proteins’ PPI binding profiles did not evolve.

For chance to strongly influence the outcomes of adaptive evolution, multiple paths to a selected phenotype must be accessible with similar probabilities of being taken. This situation could arise if several different mutations (or sets of mutations) can confer a new function or if mutations that have no effect on function accompany function-changing mutations by chance. To distinguish between these possibilities, we measured the functional effect of different sets of mutations that arose in replicates when hsMCL-1 was evolved to lose NOXA binding (Figure 7A, Figure 7—figure supplement 1). One mutation (v189G) was found at high frequency in all four replicates, but it was always accompanied by other mutations, which varied among trajectories. We found that v189G was a major contributor to the loss of NOXA binding, but it had this effect only in the presence of the other mutations, which did not decrease NOXA binding on their own. Mutation v189G therefore required permissive mutations to occur during directed evolution, and there were multiple sets of mutations with the potential to exert that effect; precisely which permissive mutations occurred in any replicate was a matter of chance. All permissive mutations were located near the NOXA binding cleft, suggesting a common mechanistic basis (Figure 7B).

Other starting genotypes showed a similar pattern of multiple sets of mutations capable of conferring the selected function (Figure 7—figure supplement 2). In addition, when mapped onto the protein structure, all sites that were mutated in more than one replicate either directly contacted the bound peptide or were on secondary structural elements that did so (Figure 7C–D), suggesting a limited number of structural mechanisms by which PPIs can be altered. Taken together, these results indicate that chance arose because from each starting genotype, there were multiple mutational paths to the selected specificity; partial determinism arose because the number of accessible routes was limited by the structure-function relationships required for peptide binding in this family of proteins.

We next analyzed the genetic basis for the limited degree of determinism that we observed in our experiments. Specifically, we sought to distinguish whether, from a given BID-specific starting point, only a few genotypes can confer NOXA binding while retaining BID binding or, alternatively, whether there are many such genotypes, but under strong selection a few are favored over others.

We performed PACE experiments in which we selected hsBCL-2 to retain its BID binding, without selection for or against NOXA binding; we then screened for variants that fortuitously gained NOXA binding using an activity-dependent plaque assay (Figure 8A–B). All four replicate populations produced clones that neutrally gained NOXA binding at a frequency of ~0.1% to 1% – lower than when NOXA binding was directly selected for but five orders of magnitude higher than when NOXA binding was selected against (Figure 8A, Figure 8—figure supplement 1). From each replicate, we then sequenced three NOXA-binding clones and found that all but one of them contained mutation r165L (Figure 8B), which also occurred at high frequency when the same protein was selected to gain NOXA binding (Figure 8—figure supplement 2). We introduced r165L into hsBCL-2 and found that it conferred significant NOXA binding with little effect on BID binding (Figure 8—figure supplement 3). Several other mutations appeared repeatedly in clones that fortuitously acquired NOXA binding, and these mutations were also acquired under selection for NOXA binding (Figure 8B, Figure 8—figure supplement 2). A similar pattern of common mutations was observed in AncB4 and AncB5 clones that fortuitously or selectively evolved NOXA binding (Figure 8—figure supplement 4). These observations indicate that the partial determinism we observed arises because from these starting points only a few mutations have the potential to confer NOXA binding while retaining BID binding.

Although we found that chance and contingency strongly influenced sequence outcomes in our experiments, all trajectories acquired the historically relevant PPI specificities that were selected for, indicating strong necessity at the level of protein function. This was true whether evolution began from more ‘promiscuous’ starting points that bound both BID and NOXA or from more specific proteins that bound only BID.

To further probe the evolutionary accessibility of new functions, we used PACE to select for a PPI specificity that never arose during historical evolution – binding of NOXA but not BID. We found that trajectories launched from hsMCL-1 (which binds both coregulators) readily evolved the selected phenotype, but two PACE-evolved variants of hsBCL-2, which had acquired the same PPI profile as hsMCL-1, went extinct under the same selection conditions (Figure 9, Figure 9—figure supplement 1). The inability of the derived hsBCL-2 genotypes to acquire NOXA specificity was not attributable to a general lack of functional evolvability by these proteins because they successfully evolved in a separate PACE experiment to lose their NOXA binding but retain BID binding (Figure 9—figure supplement 2). These results establish that contingency can influence the accessibility of new functions and that the sequence by which a specific functional phenotype is encoded can play important roles in subsequent phenotypic evolution.

Amino acid sequences of the human BCL-2, BCLW, BCL-xL, MCL-1, NRH, BFL1, BAK, BAX, and BOK paralogs were used as starting points for identifying BCL-2 family members in other species. For each paralog, tblastn and protein BLAST on NCBI BLAST were used to identify orthologous sequences between January and March of 2018 (Altschul et al., 1997). Sequences for each paralog were aligned using MAFFT (G-INS-I) with the –allowshift option and –unalignlevel set at 0.1. For each paralog, phylogenetic structure was determined using fasttree 2.1.11 within Geneious 10.1.3. Missing clades based on known species relationships were then identified, and specific tblastn searches were used within Afrotheria (taxid:311790), Marsupials (taxid:9263), Monotremes (taxid:9255), Squamata (taxid:8509), Archosauria (taxid:8492), Testudinata (taxid:8459), Amphibia (taxid:8292), Chondrichthyes (taxid:7777), Actinopterygii (taxid:7898), Dipnomorpha (taxid:7878), Actinistia (taxid:118072), Agnatha (taxid:1476529), Cephalochordata (taxid:7735), and Tunicata (taxid:7712) as needed. Additional sequences were added by downloading genome and transcriptome data for tuatara (Miller et al., 2012), sharks and rays (Wyffels et al., 2014), gar (Zerbino et al., 2018), ray-finned fish (Hughes et al., 2018), lamprey (Smith et al., 2018), hagfish (Takechi et al., 2011), Ciona savignyi (Zerbino et al., 2018), tunicates (Delsuc et al., 2018), echinoderms (Reich et al., 2015), porifera (Riesgo et al., 2014), and ctenophores (Moroz et al., 2014). In each case, local BLAST databases were created in Geneious and searched using tblastn. Finally, we used BCL-2DB to add missing groups as needed (Rech de Laval et al., 2014).

After collection of sequences, each paralog was realigned using MAFFT (G-INS-I) with the –allowshift option and –unalignlevel set at 0.1. Based on known species relationships, lineage-specific insertions were removed and gaps manually edited. Only a single sequence was kept among pairs of sequences differing by a single amino acid and sequences with more than 25% of missing sites were removed. For difficult to align sequences, sequences were modeled on the structures of human BCL-2 family members using SWISS-Model to identify likely locations of gaps (Waterhouse et al., 2018). Finally, paralogs were profile aligned to each other, and paralog-specific insertions were identified.

In total, 151 amino acid sites from 745 taxa were used to infer the phylogenetic relationships among BCL-2 family paralogs. PROT Test 3.4.2 was used to identify the best-fit model among JTT, LG, and WAG, with combinations of observed amino acid frequencies (+F), gamma distributed rate categories (+G), and an invariant category (+I) (Abascal et al., 2005). From this, JTT + G + F had the highest likelihood and lowest Aikake Information Criterion score. RAXML-ng 0.6.0 was then used to identify the maximum likelihood tree using JTT+G12+F0 (12 gamma rate categories with maximum likelihood estimated amino acid frequencies) (Kozlov and Stamatakis, 2019). Finally, we enforced monophyly within each paralog for the following groups: lobe-finned fish (n = 9), ray-finned fish (n = 9), jawless fish (n = 5), cartilaginous fish (n = 8), tunicates (n = 4), branchiostoma (n = 4), chordates (n = 5), ambulacraria (n = 5, hemichordata +echinodermata), deuterostomia (n = 5), protostomia (n = 5), cnidaria (n = 5), and porifera (n = 4) (values in parenthesis are number of identified paralogs in each group) and used RAXML-ng with JTT + G12 + F0 to identify the best tree given these constraints (Supplementary Data Phylogenetic.Data.zip).

Overall, we recovered three clades: a pro-apoptotic clade; a clade containing the BCL-2, BCLW, and BCLX vertebrate paralogs and BCL non-vertebrate sequences; and a clade containing the MCL-1, BFL1, and NRH vertebrate paralogs and MCL non-vertebrate sequences. We used the pro-apoptotic clade as the outgroup to the two anti-apoptotic clades. Within the BCL-2 clade, the majority of vertebrates contained all three copies. However, the exact relationship among the paralogs was unclear; only two copies were identified within jawless fish and their phylogenetic placement had weak support. Non-vertebrate clades tended to have good support and only a single copy. However, support for these groups following established species relationships was often limited. The MCL-1 clade contained the fastest evolving paralogs of the BCL-2 family. As with the BCL-2-like clade, only two copies were found within the jawless fish and the exact sister relationships among paralogs was unclear. Non-vertebrates contained only a single copy, but as with the BCL-2-like clade, support for relationships following established species relationships was often weak.

The BCL-2-like and MCL-1-like paralogs formed a clade with the BHP1 and BHP2 sequences from porifera. The sister relationships among these four clades were unresolved. In addition, we recovered a sister relationship between the BAK and BAX paralogs. While both paralogs contained copies from porifera, these clades evolved quickly and had relatively low support, and they may be artifactual. We identified only a single clade of ctenophores. Finally, the placement of BOK was unresolved; BOK may be sister to the BAK/BAX clade or an outgroup to all clades and the most ancient copy of the BCL-2 family.

Posterior probabilities of each amino acid at each site were inferred using Lazarus (Finnigan et al., 2012) to run codeml within PAML. We used the same model and alignment as used to infer the phylogeny. We used the branch lengths and topology of the constrained maximum likelihood phylogeny found by raxml-ng.

We first reconstructed the LCAs of all BCL-2 and MCL-1 like sequences, AncMB1-M, using the maximum likelihood state for each alignable site. We then reconstructed a series of ancestors from AncMB1 to modern human MCL-1. These included AncM1 (LCA of MCL-1-related sequences), AncM2 (LCA of MCL-1- related deuterostomes and protostomes), AncM3 (LCA of MCL-1-related deuterostomes), AncM4 (LCA of MCL-1-related urochordates and chordates), AncM5 (LCA of MCL-1, BFL1, and NRH like copies in vertebrates), AncM6 (LCA of MCL-1 and BFL1 like copies), AncMCL-1 (LCA of MCL-1 like copies), AncMCL-1-G (LCA of MCL-1 like Gnathostomes), AncMCL-1-O (LCA of MCL-1 like Osteichthyes), and AncMCL-1-T (LCA of MCL-1 like Tetrapods), AncMCL-1-A (LCA of MCL-1 like Amniotes), and AncMCL-1-M (LCA of MCL-1 like Mammals). In each case, the sequence of each ancestor used the maximum likelihood state at each site, with gaps inserted based on parsimony. We used the modern sequences of human MCL-1 to fill in portions of the sequence that showed poor alignment and could not be reconstructed, including both the N and C terms, as well as the loop between the first and second alpha helices. Average posterior probabilities for ancestors in the MCL-1 clade ranged from 0.73 (AncM6) to 0.98 (AncMCL-1-M) with an average of 0.83 (sd 0.08) (Supplementary file 2).

For the BCL-2 like clade, we also reconstructed AncMB1, this time using human BCL-2 sequence to fill in the N and C terms and the loop between the first and second alpha helices (AncMB1-B). We then reconstructed sequences from AncMB1 to modern human BCL-2. These included AncB1 (LCA of BCL-2-related sequences), AncB2 (LCA of BCL-2-related Bilaterian and Cnidaria), AncB3 (LCA of BCL-2-related deuterostomes and protostomes), AncB4 (LCA of BCL-2 deuterostomes), AncB5 (LCA of BCL-2, BCLW, and BCLX like copies in vertebrates), AncBCL-2 (LCA of BCL-2 like copies), AncBCL-2-G (LCA of BCL-2 like gnathostomes), AncBCL-2-O (LCA of BCL-2 like osteichthyes), and AncBCL-2-T (LCA of BCL-2 like tetrapods), using human BCL-2 sequences for the N and C terms and the loop between the first and second alpha helices. Average posterior probabilities for ancestors in the BCL-2 clade ranged from 0.87 (AncB1) to 0.95 (AncBCL-2-T) with an average of 0.9 (sd 0.04).

To determine the robustness of our conclusions on the phenotype of ancestral sequences, we synthesized and cloned alternative reconstructions for key ancestors. In each case, sequences contained the most likely alternative state with posterior probability > 0.2 for all such sites where such a state existed. Alternative reconstructions contained an average of 24 alternative states and represent a conservative test of function (min: 4, max: 44, Supplementary file 2). In our luciferase assay, all but two alternative reconstructions retained similar BID and NOXA binding as the maximum likelihood ancestral sequences. The first alternative reconstruction that differed from the maximum likelihood reconstruction was AltAncB3, which bound both BID and NOXA, while the ML for AncB3 bound BID, but NOXA only weakly. As a result, the exact branch upon which NOXA binding was lost historically is not resolved by this data.

The second alternative reconstruction that differed from the ML reconstruction was AltAncMB1-B, which had weaker NOXA binding than the ML reconstruction. To further test the robustness of AncMB1-B to alternative reconstructions, we synthesized and tested additional reconstructions that included only alternative amino acids with posterior probabilities greater than 0.4 (n = 3), 0.35 (n = 7), 0.3 (n = 13), and 0.25 (n = 18), and compared these to AncMB1-B and the 0.2 AltAncMB1-B (n = 21) (values in parentheses are number of states that differ from the ML state). We found that the 0.4, 0.35, and 0.3 alternative reconstructions bound both BID and NOXA, while the 0.25 and 0.2 alternative reconstructions had diminished NOXA binding.

Finally, we synthesized and tested modern sequences from key groups to determine the robustness of our inference on the timing of NOXA binding loss. These included BCL-2-related sequences from groups that diverged prior to the predicted loss of NOXA binding (Trichoplax adhaerens and Hydra magnapapillata), sequences from groups that diverged around the time of predicted NOXA binding loss (Octopus bimaculoides and Stegodyphus mimosarum), or sequences from groups predicted to have diverged after NOXA binding lost (Saccoglossus kowalevskii and Branchiostoma belcheri). In each case, we used human BCL-2 sequence to replace extant N and C terms and the loop between the first and second alpha helices. The T. adhaerens and B. belcheri sequences were non-functional in our luciferase assays, binding neither BID nor NOXA. However, recent work has comprehensively characterized binding in BCL-2 family members within T. adhaerens, finding that the BCL copy can bind both BID and NOXA as predicted (Popgeorgiev et al., 2020). H. magnapapillata bound both BID and NOXA in our assay and the remaining sequences bound only BID, suggesting a loss of NOXA binding prior to the divergence of protostomes and deuterostomes in the BCL-2 related clade, consistent with the conclusion drawn using reconstructed proteins.

E. coli 10-beta cells were used for cloning and were cultured in 2xYT media. E. coli BL21 (BE3) cells were used for protein expression and were cultured in Luria-Bertain (LB) broth. E. coli S1030 cells cultured in LB broth were used for activity-dependent plaque assays, phage growth assays, and luciferase assays. S1030 cells cultured in Davis Rich media were used for PACE experiments (Carlson et al., 2014). E. coli 1059 cells were used for cloning phage and assessing phage titers and were cultured in 2xYT media.

Plasmids were constructed by using Q5 DNA Polymerase (NEB) to amplify fragments that were then ligated via Gibson Assembly. Primers were obtained from IDT, and all plasmids were sequenced at the University of Chicago Comprehensive Cancer Center DNA Sequencing and Genotyping Facility. Vectors and gene sequences used in this study are listed in Supplementary file 5, with links to fully annotated vector maps on Benchling. Key vectors are deposited at Addgene, and all vectors are available upon request. The following working concentrations of antibiotics were used: 50 µg/mL carbenicillin, 50 µg/mL spectinomycin, 40 µg/mL kanamycin, and 33 µg/mL chloramphenicol. Protein structures and alignments were generated using the program PyMOL (Schrödinger, 2018).

Cloned expression vectors contained the following: (1) a previously evolved, isopropyl β-D-1-thiogalactopyranoside (IPTG)-inducible N-terminal half of T7 RNAP (Zinkus-Boltz et al., 2019) fused to a BCL-2 family protein; (2) the C-terminal half of T7 RNAP fused to a peptide from a BH3-only protein; and (3) T7 promoter-driven luciferase reporter. Chemically competent S1030 E. coli cells (Carlson et al., 2014) were prepared by culturing to an OD₆₀₀ of 0.3, washing twice with a calcium chloride/HEPES solution (60 mM CaCl₂, 10 mM HEPES pH 7.0, 15% glycerol), and then resuspending in the same solution. Vectors were transformed into chemically competent S1030 cells via heat shock at 42°C for 45 s, followed by 1 hr recovery in 3× volume of 2xYT media, and then plated on agar with the appropriate antibiotics (carbenicillin, spectinomycin, and chloramphenicol) to incubate overnight at 37°C. Individual colonies (three to four biological replicates per condition) were picked and cultured in 1 mL of LB media containing the appropriate antibiotics overnight at 37°C in a shaker. The next morning, 50 µL of each culture was diluted into 450 µL of fresh LB media containing the appropriate antibiotics, as well as 1 µM of IPTG. The cells were incubated in a shaker at 37°C, and OD₆₀₀ and luminescence measurements were recorded between 2.5 and 4.5 hr after the start of the incubation. Measurements were taken on a Synergy Neo2 Microplate Reader (BioTek) by transferring 150 µL of the daytime cultures into Corning black, clear-bottom 96-well plates. Data were analyzed in Microsoft Excel and plotted in GraphPad Prism, as previously reported (Pu et al., 2017a).

Protein expression hsBCL-2, hsMCL-1, and evolved variants were constructed as N-terminal 6xHis-GST tagged proteins. The recombinant proteins were expressed in BL21 E. coli (NEB) and purified following standard Ni-NTA resin purification protocols (ThermoFisher Scientific) (Zhou et al., 2019). Briefly, BL21 E. coli containing an N-terminal 6xHis-GST tagged BCL-2 family protein were cultured in 5 mL LB with carbenicillin overnight. The following day, the culture was added to 0.5 L of LB with carbenicillin, incubated at 37°C until it reached an OD₆₀₀ of 0.6, induced with IPTG (final concentration: 200 µM), and cultured overnight at 16°C. The cell pellet was harvested by centrifugation followed by resuspension in 30 mL of lysis buffer (50 mM Tris 1 M NaCl, 20% glycerol, 10 mM TCEP, pH 7.5) supplemented by protease inhibitors (200 nM Aprotinin, 10 µM Bestatin, 20 µM E-64, 100 µM Leupeptin, 1 mM AEBSF, 20 µM Pepstatin A). Cells were lysed via sonication and were then centrifuged at 12,000 g for 40 min at 4°C. Solubilized proteins, located in the supernatant, were incubated with His60 Ni Superflow Resin (Takara) for 1 hr at 4°C, and the protein was eluted using a gradient of imidazole in lysis buffer (50–250 mM). Fractions with the protein, as determined by SDS-PAGE, were concentrated in Ulta-50 Centrifugal Filter Units (Amicon, EMD Millipore). Proteins were purified via a desalting column with storage buffer (50 mM Tris–HCl [pH 7.5], 300 mM NaCl, 10% glycerol, 1 mM DTT) and further concentrated. The concentration of the purified BCL-2 family proteins was determined by BCA assay (ThermoFisher Scientific), and they were flash-frozen in liquid nitrogen and stored at −80°C.

Fluorescent polarization (FP) was used to measure the affinity of BCL-2 family proteins with peptide fragments of the BH3-only proteins in accordance with previously described methods (Zhang et al., 2002). hsBCL-2, hsMCL-1, and evolved variants were purified as described above. The fluorescent NOXA and BID peptides (95+% purity) were synthesized by GenScript and were N-terminally labeled with 5-FAM-Ahx and C-terminally modified by amidation. These peptides were dissolved and stored in DMSO. Corning black, clear-bottom 384-well plates were used to measure FP, and three replicates were prepared for each data point. Each well contained the following 100 µL reaction: 20 nM BH3-only protein, 0.05 nM to 3 µM of BCL-2 family protein (1/3 serial dilutions), 20 mM Tris (pH 7.5), 100 mM NaCl, 1 mM EDTA, and 0.05% pluronic F-68. FP values (in milli-polarization units; mFP) of each sample were read by a Synergy Neo2 Microplate Reader (BioTek) with the FP 108 filter (485/530) at room temperature 5–15 min after mixing all the components. Data were analyzed in GraphPad Prism 8, using the following customized fitting equation, to calculate K_d (Zhou et al., 2019):y=B+C(D+Kd+x-D+Kd+x2-4Dx)where y is normalized measured FP, x is the concentration of BCL-2 protein, D is the concentration of the BH3-only protein, B and C are parameters related to the FP value of free and bound BH3-only protein, and K_d is the dissociation constant.

PACE was used to evolve hsBCL-2, hsMCL-1, and ancestral proteins in accord with previously reported technical methods (Carlson et al., 2014; Esvelt et al., 2011; Pu et al., 2019; Pu et al., 2017b) using a new vector system. Briefly, combinations of accessory plasmids and the MP6 mutagenesis plasmid (Badran and Liu, 2015) were transformed into S1030 E. coli., plated on agar containing the appropriate antibiotics (carbenicillin, kanamycin, and chloramphenicol) and 10 mM glucose, and incubated overnight at 37°C. Colonies were grown overnight in 5 mL of LB containing the appropriate antibiotics and 20 mM glucose. Davis Rich media was prepared in 5–10 L carboys and autoclaved, and the PACE flasks and corresponding pump tubing were autoclaved as well. The following day, PACE was set up in a 37°C environmental chamber (Forma 3960 environmental chamber, ThermoFisher Scientific). For each replicate, an overnight culture was added to ~150 mL of Davis Rich carboy media in chemostats and grown for 2–3 hr until reaching an OD₆₀₀ of approximately 0.6. Lagoons containing 20 µL of phage from saturated phage stocks (10⁸–10⁹ phage) were then connected to the chemostat. Magnetic stir bars were used to agitate chemostats and lagoons. The chemostat cultures were flowed into the lagoons at a rate of approximately 20 mL/h. Waste output flow rates were adjusted to maintain a constant volume of 20 mL in the lagoons, 150 mL in the chemostat, and an OD₆₀₀ close to 0.6 in the chemostat. A 10% w/v arabinose solution was pumped into the lagoons at a rate of 1 mL/h. If the experiment included a mixing step (two separate chemostats flowed together into one lagoon for a mixed selection pressure), a chemostat was prepared the next day (as described above) and connected to the lagoons. During this step, lagoon volumes were increased to 40 mL, and the arabinose inflow rate was increased to 2 mL/h. After disconnecting the first chemostat the next day, the lagoon volumes and arabinose inflow were both lowered to 20 mL and 1 mL/h, respectively. During the experiment, samples were collected from the lagoons every 24 hr and centrifuged at 13,000 rpm for 3 min to collect the phage-containing supernatant, as well as the cell pellet for DNA extraction. PACE experiments are listed in Supplementary file 3. A single replicate of AncB5 was removed because of contamination. No statistical method was used to determine the number of replicates as only four independent replicate experiments could be performed simultaneously.

During PACE, the media volume of each lagoon turned over once per hour for 4 days, or ~100 times. For a phage population to survive this amount of dilution, a similar number of generations must have occurred between the starting phage and the phage in the lagoon at the end of the experiment (Esvelt et al., 2011). This is expected to be a conservative estimate; as a more fit phage rises in frequency in the population, it will undergo a greater number of generations than less-fit phage in the population. The mutagenesis plasmid MP6 induces a mutation rate of approximately 6 × 10⁻⁶ per bp per generation. The BCL2 family proteins used in the PACE experiments were ~230 amino acids long, indicating that a mutation occurred on average every ~250 phage replications. Phage population sizes ranged from 10⁵ per mL to 10¹⁰ per mL over the course of a PACE experiment, indicating a rate of 400–40,000,000 new mutations every generation. Conservative estimates thus suggest that a during each individual replicate, phage populations sampled at least 40,000 mutations, and upwards of 4 × 10⁹ mutations. While not all mutations were equally likely each generation because MP6 enriches for transitions (i.e. G→A, A→G, C→T, and T→C), the high number of mutations sampled suggests that the vast majority of possible single point mutations (approximately 230*3*4 = 2760 potential mutations) were sampled over the course of each experiment, with higher population sizes generating all potential single point mutations each generation.

Plaque assays were performed on 1059 E. coli cells (Carlson et al., 2014; Hubbard et al., 2015), which supply gene III (gIII) to phage in an activity-independent manner, to measure phage titers. Additionally, activity-dependent plaque assays were done on S1030 E. coli containing the desired accessory plasmids to determine the number of phage encoding a BCL-2 family protein with a given peptide-binding profile. All cells were grown to an OD₆₀₀ of approximately 0.6 during the day. Four serial dilutions were done in Eppendorf tubes by serially pipetting 1 μL of phage into 50 µL of cells to yield the following dilutions: 1/50, 1/2500, 1/125,000, and 1/6,250,000. 650 µL of top agar (0.7% agar with LB media) was added to each tube, which was then immediately spread onto a quad plate containing bottom agar (1.5% agar with LB media). Plates were incubated overnight at 37°C. Plaques were counted the following day, and plaque forming units (PFU) per mL was calculated using the following equation:PFU=1000*A*504-B

Where A is the number of plaques in a given quadrant, and B is the quadrant number where the phage were counted, in which one is the least dilute quadrant and four is the most dilute quadrant.

Phage growth assays were performed by adding the following to a culture tube and shaking at 37°C for 6 hr: 1 mL of LB with the appropriate antibiotics (carbenicillin and kanamycin), 10 µL of saturated S1030 E. coli containing the accessory plasmids of interest, and ~1000 phage. Phage were then isolated by centrifugation at 13,000 rpm for 3 min, and PFU was determined by plaque assays using 1059 E. coli and the plaque assay protocol described above.

PACE samples were collected from each lagoon every 24 hr. The lagoon samples were centrifuged at 13,000 rpm for 3 min on a bench top centrifuge to separate supernatant and cell pellet. The phage-containing supernatants were stored at 4°C prior to the creation of sequencing libraries. To prepare Illumina sequencing libraries, each phage sample was cultured overnight with 1059 E. coli cells, followed by phage DNA purification (Qiagen plasmid purification reagent buffer), P1 (catalog number 19051), P2 (catalog number 19052), N3 (catalog number 19064), PE (catalog number 19065), and spin column for DNA (EconoSpin, catalog number 1920–250). The resulting DNA concentration was ~50 ng/µL. Freshly generated DNA samples were then used as template for PCR amplification. For each library sample, we amplified three overlapping fragments of the BCL-2 family protein, which are 218–241 bp in length (Figure 4—figure supplement 1). Each primer also included 6–9 ‘N’s to introduce length variation (Supplementary file 4). In total, 12 PCRs were used for each library. Phusion DNA polymerases and buffers (ThermoFisher Scientific, catalog number F518L) were used in the first PCR round to amplify all three fragments for all library sequencing. The 25 µL reaction contained: 0.5 µL of 50 mM MgCl₂, 0.75 µL of 10 mM dNTP, 0.75 µL Phusion DNA polymerase, 20 ng library DNA, and 0.5 µL of 10 µM primer (each). The PCR were run on a C1000 Touch Thermal Cycler (Bio-Rad), with the following parameters: 98°C for 1 min, followed by 16 cycles of 98°C for 12 s, 58°C for 15 s and 72°C for 45 s, and finally 72°C for 5 min. PCR were purified using the ZYMO DNA clean and concentrator kit (catalog number D4013) and 96 well filter plate (EconoSpin, catalog number 2020–001). The DNA products were dissolved in 30 µL ddH₂O. All 12 reactions for each library were combined, and 1 µL was used as the template for a second PCR round. PCR components and thermocycler parameters were the same as above, except that the annealing temperature was 56°C, and only 15 rounds of amplification were conducted. The primer and sample combinations are listed in Supplementary file 4. PCRs were then purified following the same procedure as previous step. Equal volumes of all 72 library samples were combined and concentration was measured using a Qubit 4 Fluorometer. The total DNA sample was 2.68 ng/µL (equivalent to 10 nM, according to the average length of PCR fragments). DNA samples were diluted to 4 nM from step 4 following the Illumina MiSeq System Denature and Dilute Libraries Guide and then diluted to 12 pM for high-throughput sequencing. The final sample contained 100 µL of 20 pM PhiX spike-in plus 500 µL of the 12 pM library sample. Sequencing was performed on the Illumina MiSeq System using MiSeq Reagent Kit v3 (600-cycle) with paired-end reads according to the manufacturer’s instructions.

Illumina sequencing yielded 22 million reads, 13 million of which could be matched to a specific sample (Supplementary file 4). One replicate for AncB5 was found to be contaminated and removed from further analysis. To process the remaining data, we first used Trim Galore with default settings to trim reads based on quality (https://www.bioinformatics.babraham.ac.uk/projects/trim_galore/). Then, we used BBMerge, a script in BBTools (https://jgi.doe.gov/data-and-tools/bbtools/), to merge paired-end reads. Next, we used Clumpify to remove repeated barcode sequences. We then used Seal to identify and bin reads by sample and fragment. Finally, we used BBDuk to remove any primer or adapter sequence present. Scripts and reference sequences are available on Github (Thornton, 2021).

Reads were binned by experiment and then aligned to the appropriate WT sequence using Geneious (low sensitivity, five iterations, gaps allowed). Sequences were then processed in R to remove sequences containing ‘N’s or that were not full length. Insertions found in less than 1% of the population and sites that extended outside of the coding region were removed from all sequences. Remaining gaps were standardized among replicates and within an experiment. Finally, allele frequencies were calculated for each site and amino acid, as well as remaining insertions and deletions.