The Base Substitution and Frameshift Fidelity ofEscherichia coli DNA Polymerase III Holoenzyme in Vitro *

We have investigated the in vitrofidelity of Escherichia coli DNA polymerase III holoenzyme from a wild-type and a proofreading-impaired mutD5 strain. Exonuclease assays showed the mutD5 holoenzyme to have a 30–50-fold reduced 3′→5′-exonuclease activity. Fidelity was assayed during gap-filling synthesis across the lacI dforward mutational target. The error rate for both enzymes was lowest at low dNTP concentrations (10–50 μM) and highest at high dNTP concentration (1000 μM). The mutD5 proofreading defect increased the error rate by only 3–5-fold. Both enzymes produced a high level of (−1)-frameshift mutations in addition to base substitutions. The base substitutions were mainly C→T, G→T, and G→C, but dNTP pool imbalances suggested that these may reflect misincorporations opposite damaged template bases and that, instead, T→C, G→A, and C→T transitions represent the normal polymerase III-mediated base·base mispairs. The frequent (−1)-frameshift mutations do not result from direct slippage but may be generated via a mechanism involving “misincorporation plus slippage.” Measurements of the fidelity of wild-type and mutD5 holoenzyme during M13 in vivo replication revealed significant differences between the in vivo and in vitro fidelity with regard to both the frequency of frameshift errors and the extent of proofreading.

Understanding how organisms maintain low, acceptable levels of mutations is an important problem in biology. One major determinant of mutation rate is the fidelity of DNA replication, and the study of the precise mechanisms by which replication fidelity is achieved is therefore important. The rates of spontaneous mutation are typically as low as ϳ10 Ϫ9 to 10 Ϫ10 mutations per base pair replicated (1). This high fidelity is generally achieved through at least three mechanisms operating consecutively as follows: base selection (the discrimination between correct and incorrect nucleotides by the DNA polymerase during the nucleotide insertion step); exonucleolytic proofreading (the editing of newly incorporated nucleotides by an associated 3Ј35Ј-exonuclease activity); and post-replicative mismatch repair (recognition and correction of mispairs resulting from DNA replication errors).
In the bacterium Escherichia coli, DNA polymerase III ho-loenzyme (HE) 1 is responsible for replication of the chromosome (2). The HE is a large dimeric complex composed of two DNA polymerase III core assemblies, one for each strand at the replication fork, and several auxiliary subunits (2)(3)(4). The core contains three tightly bound subunits as follows: the polymerase or ␣ subunit (encoded by the dnaE gene); the ⑀ or 3Ј35Јsubunit (encoded by the dnaQ gene); and the subunit (encoded by the holE gene), whose role is currently unclear. The auxiliary subunits are responsible for high processivity and for interaction with other replication proteins to achieve coordinated, simultaneous synthesis of leading and lagging strands (5)(6)(7)(8). The auxiliary subunits can be divided into the following two sub-assemblies: ␤ dimers (sliding clamps) that form a donut-shaped structure encircling the DNA (9) tethering each core enzyme to the template (10,11), and dnaX complex, which is responsible for loading the sliding clamp onto the DNA (12,13). The dnaX complex also connects the two core polymerases through the subunit. We are interested in understanding the mechanisms underlying the fidelity and specificity of the pol III HE, including the possible contributions of the various subunits to fidelity. We have recently reported on our studies on the in vitro fidelity of the purified ␣ subunit (14). We report here on our analysis of the in vitro fidelity of HE. The fidelity of HE has been investigated previously using a X174 reversion assay and has provided information about base substitution fidelity at a small number of template nucleotides (15)(16)(17)(18). More recently, a gel kinetic assay was used to determine the HE misincorporation rates opposite a template G nucleotide and also to provide insights into the possible role of the pol III accessory factors (19). In the present study, we use a previously developed forward mutational assay based on the N-terminal part of the lacI gene as a mutational target (14). This target allows us to assay a broad variety of different DNA replication errors and to measure them at a large number of DNA sequence sites. Application of this system to the isolated ␣ subunit (14) has revealed an unusual specificity not observed for any other enzyme: an excess of (Ϫ1)-frameshift mutations was produced at generally non-reiterated sites, which we proposed to occur through a misincorporation ϩ slippage mechanism. Thus, one of the objectives of the current work was to investigate to what extent HE is capable of avoiding (or correcting) these types of mistakes. One other aspect of our study was the use of both a wild-type and a proofreading-deficient enzyme, purified from a mutD5 strain, to allow estimation of the contribution of proofreading as was previously done in vivo (20). Our results show that HE produces relatively more base substitutions than the ␣ * The costs of publication of this article were defrayed in part by the payment of page charges. This article must therefore be hereby marked "advertisement" in accordance with 18 U.S.C. Section 1734 solely to indicate this fact. subunit but retains the high propensity for frameshift mutagenesis. The results are discussed in light of the possible underlying mechanisms.
Purification of DNA Polymerase III Holoenzyme from Wild-type and mutD5 Strains-Wild-type and MutD5 DNA polymerase III HE were purified from strains NR9474 and NR9669, respectively (see above). An inoculum for a large scale fermentation was prepared by growth overnight in 20 liters of M9 medium containing 0.25 g of tryptophan/liter and 0.1 g of streptomycin/liter at 37°C. Cells were transferred to 100 liters containing 14.5 g/liter yeast extract, 9 g/liter peptone, 11 g/liter K 2 HPO 4 , 1.625 g/liter KH 2 PO 4 , and 12.5 g/liter glucose. The pH was maintained between 7.0 and 7.2. After 60 and 150 min, 6 g of glucose was added per liter of fermentation medium. Cells were grown to an A 600 of 5.0 (mutD5) and 6.2 (wild type), respectively, and harvested within 45 min through a Sharples AS16 continuous flow centrifuge. Cells were chilled to 16°C through a heat exchanger during harvest. After harvest, cells were suspended immediately in an equal weight of 50 mM Tris-HCl, pH 7.5, 10% sucrose and frozen by pouring into liquid N 2 . Cells were stored at Ϫ20°C, and 500 g of each strain were lysed as described (25), except the incubation with lysozyme was shortened to 30 min on ice. The ammonium sulfate fraction (fraction II) was produced as described (26) except the second 0.17 ammonium sulfate backwash was substituted with an ammonium sulfate backwash solution prepared by addition of 0.20 g ammonium sulfate to each milliliter of buffer. All other purification steps were as described (26). The specific activities of the final fraction V were 338,000 units/mg for the wild-type and 570,000 units/mg for the MutD5 HE. Fraction V was subjected to SDS-polyacrylamide gel electrophoresis. From the scan of the SDS gels, the MutD5 and wild-type HE preparations were 63 and 65% pure, respectively. From densitometric scans of the Coomassie-stained gels, the uncorrected staining ratio of ⑀ to ␣ was identical for both preparations (0.27 and 0.28 for the MutD5 and wild-type preparations, respectively).
Preparation of Gapped mRS65 DNA-DNA substrates used in this study for gap-filling reactions represent mRS65 double-stranded circular molecules with a 310-bp gap in the lacI d target (position Ϫ62 to ϩ248 of lacI). These gapped substrates were prepared essentially as described by Kunkel and Soni (27), as follows. Phage mRS65 was plated on minimal medium plates using NR9099 as host. A single plaque was added to 200 ml of LB medium containing 2 ml of an overnight culture of NR9099. The infected cells were grown for 6 -8 h at 37°C with vigorous shaking. Cells were harvested by centrifugation at 5000 ϫ g for 20 min and washed twice with STE buffer (100 mM NaCl, 10 mM Tris-HCl, pH 8.0, 1 mM EDTA). Replicative form (RF) of mRS65 was prepared using a Qiagen plasmid kit (Qiagen, Inc.). mRS65 ssDNA was purified from the supernatant of the same culture by phenol extraction (28). mRS65 RF DNA was digested to completion with restriction endonuclease NsiI, which cuts lacI at position Ϫ62. After extraction twice with Strataclean Resin (Stratagene) to remove NsiI, the linearized DNA was further digested with SphI (cut site at position ϩ250 of lacI) to produce two fragments, 8169 and 310 base pairs in length. The solution was treated twice with Strataclean Resin to remove SphI. Purification of the large fragment away from the small fragment was achieved by adjusting the solution to 5.5% PEG-8000, 0.55 M NaCl, and 0.5 mg/ml DNA concentration and incubation for 2 h at 37°C, under which conditions the large fragment is precipitated selectively. The precipitated fragment was collected by centrifugation, resuspended in TE buffer, precipitated with ethanol, and resuspended in TE. To form the gapped template, the purified 8169-bp fragment was hybridized to single-stranded circular viral (plus) DNA of mRS65 (27).
Gap-filling Reactions-The following components were mixed in a  (29), 10 l of the mixture was loaded onto an 0.8% agarose gel, and electrophoresis was carried out at 90 V overnight. The product of completed gap-filling reaction can be seen as a single band comigrating with an RF II mRS65 marker DNA (14). Transfection, Plating, and Calculation of in Vitro Error Rates-The products of gap-filling reactions were diluted 10-fold, and 1 or 2 l was used to transfect competent cells prepared from strain MC1061 by electroporation using a Bio-Rad Gene Pulser (Bio-Rad) set at 2.01 kV, 400 watts, and 25 microfarads. Immediately following electroporation, 1 ml of SOC medium was added. Plating was performed by adding an aliquot of the transfected cells to 3 ml of melted soft agar (42°C) containing 2.5 g of X-gal and 0.25 ml of a mid-log culture of NR9099. This mixture was poured onto minimal agar plates. The plates were inverted and incubated for 24 h at 37°C. A sufficient amount of transfection mixture was used to obtain 500 -1000 plaques per plate. At this density, blue (mutant) plaques are easily detected against the background of colorless wild-type plaques. The mutant frequency was calculated by dividing the number of blue plaques by the total plaque number. For calculation of error rates (the number of errors of any given kind per base per round of replication), the mutant frequency was divided by 0.55 (the average mutant expression efficiency (14,30)) and by the number of sites at which the mutations of interest are detectable in the lacI d target (220 for the (Ϫ1)-frameshifts and ϳ100 for the base substitutions). The number of sites at which individual base substitutions are detectable was derived from Schaaper (20).
DNA Sequence Analysis-Mutant (blue) plaques were selected randomly for sequence analysis. Plaques were purified by replating on NR9099 to confirm the mutant phenotype. Single-stranded DNA from the mutant phages was purified and used for dideoxy sequencing reactions carried out with Sequenase version 2.0 using the protocol provided by U. S. Biochemical Corp. A 16-mer primer, 5Ј-CACCACCACGCTG-GCA-3Ј, complementary to positions ϩ317 to ϩ301 of lacI was used, and the region between positions Ϫ50 and ϩ250 of lacI was sequenced.
Exonuclease Assays-Four 19-mer oligonucleotides 5Ј-AACGACGT-TTTGCCCGCCN-3Ј with a variable 3Ј-terminal base N (A, T, G, or C) complementary to positions ϩ195 to ϩ213 of lacI (3Ј-base of the oligomer is opposite position ϩ195) were purchased from Research Genetics, Inc. (Huntsville, AL). The oligomers were purified on a 15% (w/v) polyacrylamide, 8 M urea gel. The purified 19-mers were 5Ј-phosphorylated using T4 polynucleotide kinase and [␥-32 P]ATP according to a procedure by Sambrook et al. (28). The labeled oligomers were annealed to mRS65 ssDNA at a 1:1.5 molar ratio (mRS65 DNA in excess). Reaction mixtures (15 l) containing 30 mM Hepes⅐KOH, pH 7.5, 16 mM Tris, pH 7.5, 6 mM dithiothreitol, 10 mM MgCl 2 , 100 g/ml bovine serum albumin, 500 M ATP, and 10 nM labeled DNA were preincubated for at least 2 min at 37°C. Reactions were initiated by adding either 6 units of wild-type HE or 13.5 units of MutD5 HE. At appropriate time points, aliquots (2 l) were removed and quenched in 90% formamide, 20 mM EDTA. Samples were heated to 100°C for 5 min, and oligonucleotides were separated by electrophoresis in 15% polyacrylamide gels containing 8 M urea (28). The gels were dried, and the intensity of each band was measured using a Molecular Dynamics PhosphorImager and Im-ageQuant (version 2.0) software. Exonuclease activity was calculated from the fraction of 19-mer converted to shorter products, normalized for the total intensity present per lane to account for loading differences.
Measurement of in Vivo Mutation Rates-To estimate the fidelity of HE in vivo, phage mRS65 was propagated on wild-type (NR9102) and mutD5 (NR11594) strains as follows. mRS65 was plated on minimal medium plates using NR9102 or NR11594 as host cells. After incubation at 37°C overnight, 150 single plaques for each strain were transferred by sterile Pasteur pipette into 1 ml of 0.9% NaCl. Appropriate dilutions were plated on minimal plates containing 50 g/ml X-gal and strain NR9099 as ␣-complementation host. After incubation at 37°C for 24 h, mutant (blue) plaques were scored, and one blue plaque (if available) was picked randomly from each dilution. ssDNA of the mutant phage was prepared and sequenced as described above. Generally, a mutation in the lacI d region (position Ϫ50 to ϩ250 of lacI) was found in only about one-third of the sequenced mutants. Therefore, lacI d mutant frequencies were obtained by dividing the average blue plaque frequency by 3. Approximate mutation rates () for the lacI d target were calculated by the equation ϭ f/ln(N) (1), in which f is the measured mutation frequency at (phage) population size N (ϭ 5 ϫ 10 9 ).

3Ј35Ј-Exonuclease Activity of Wild-type and MutD5 polIII
Holoenzyme-pol III HE was purified from both a wild-type and the corresponding mutD5 strain. mutD5 strains carry a defect in the dnaQ gene encoding the ⑀-proofreading subunit, resulting in a strong mutator effect (20,(31)(32)(33)(34)(35)(36). The purification properties of the two enzymes proved similar, yielding similar specific polymerase activities in the final fractions, 338 and 570 units/g for wild-type and MutD5 enzyme, respectively (see "Experimental Procedures"). The slightly higher specific activity of the MutD5 HE may result from the reduced exonuclease of this enzyme. The subunit composition for the two enzymes as determined on SDS gels yielded identical results (data not shown). Importantly, the ratio of ␣ to ⑀ subunit was identical in the two preparations. This is consistent with the dominant character of the mutD5 mutation in genetic complementation experiments (31,34). This dominance has been interpreted to indicate that the MutD5 ⑀ subunit, although catalytically deficient, retains normal binding to the other core subunits (31,34). Recent studies using the yeast two-hybrid system have also indicated normal ␣-⑀ binding for the mutD5-derived ⑀ subunit (37).
The two holoenzymes were examined for their 3Ј35Ј-exonuclease activity. The assay was carried out by hybridizing 5Јend-labeled 19-mers to mRS65 ssDNA using four different oligonucleotides with the 3Ј-terminal base as either A, T, C, or G opposite the template T at position 195 of lacI and assaying the rate of disappearance of the terminal nucleotide. The experiments were carried out under conditions of excess of DNA. The reaction rate was proportional to the enzyme concentration and remained constant during the assay time (data not shown). The products of the reaction were analyzed on a 15% denaturing polyacrylamide gel (Fig. 1). The data show that the MutD5 HE has strongly reduced 3Ј35Ј-exonuclease activity compared with the wild-type enzyme, although it has some residual activity. The latter was also inferred from in vivo experiments (36). The data are summarized in Table I. For the four substrates, the exonuclease ratio of wild-type versus mutD5 is between 26-and 47-fold. The assay was also carried out in the presence of dGMP, a known competitive inhibitor of the 3Јexonuclease activity (38). Table I shows that addition of 10 mM dGMP decreased the exonuclease activity of the wild-type enzyme by 7-10-fold and of the MutD5 enzyme by up to 4-fold.
Fidelity of pol III Holoenzyme during in Vitro Gap-filling DNA Synthesis-To measure the fidelity of the HE we used a forward mutational assay using the N-terminal region of the lacI gene as mutational target (lacI d target). This assay was first developed to determine the fidelity of the purified ␣ subunit (14). Gapped molecules were constructed of phage mRS65 (M13lacI ϩ lacZ␣) containing a 310-bp gap in the minus strand (positions Ϫ62 to ϩ248 of lacI). The gap is filled by HE, and completed RF II products are transfected into competent cells. Since phage mRS65 carries the lacZ␣ gene, it is capable of ␣-complementation on a lacZ⌬M15 host such as NR9099 (14); however, since it also carries the lacI gene encoding the lac repressor, ␣-complementation depends on the status of the lacI gene. On plates containing X-gal, mRS65 (which is lacI ϩ ) will produce colorless plaques, whereas lacI mutants generated by DNA synthesis errors will produce blue plaques. Thus, the fraction of blue plaques is a measure of the polymerization accuracy.
The fidelity of wild-type and MutD5 HE was determined at different dNTP concentrations in the reaction (Table II). For both enzymes the mutant frequency increases with increasing dNTP concentration. The dependence of in vitro error rates on dNTP concentration is a well known phenomenon (15, 39 -41). It reflects the enhanced rate of mispair extension at higher dNTP concentrations, which allows favorable competition of the extension reaction (necessary for mutation observance) with the proofreading reaction (39) or polymerase dissociation (40). The wild-type and MutD5 enzyme appear similarly susceptible to the effect of increasing dNTP concentration, the mutant frequencies (after DNA sequencing) increasing 5.5-(wild-type) or 7-fold (MutD5) when dNTPs are increased from 10 to 1000 M (Table II). Interestingly, the difference between the wild-type and the MutD5 enzyme is only moderate, between 2.2-and 4.5-fold, regardless of dNTP concentration.
The reactions described in this paper were performed without the presence of SSB. However, one complete set of gapfilling reactions, including DNA sequencing of the generated mutants (see later), was also performed in the presence of various amounts of SSB including saturating concentrations. No effect of SSB was noted on the extent of DNA synthesis, the mutation frequencies, or the specific types of mutations created by HE (data not shown).
Mutational Specificity of Wild-type and MutD5 pol III Holoenzyme-DNA sequencing of mutant phage (Table III) revealed that HE produces several types of mutations as follows: base substitutions, frameshifts, deletions, and complex mutations, but the large majority are (Ϫ1)-frameshifts and base substitutions. Spectra for the two HEs recorded at 50 M dNTPs are shown in Fig. 2. In Table IV we present the calcu- The substrate DNA is shown at the top. The four 19-mer oligonucleotides each with a different 3Ј-terminal base were 5Ј-labeled and annealed to mRS65 ssDNA. Exonuclease reactions were carried out as described under "Experimental Procedures." At indicated time points, aliquots were removed and subjected to 15% denaturing polyacrylamide gel electrophoresis. The 19-mer bands represent intact oligonucleotide. The 18-mer or smaller bands represent products of the exonuclease reaction. The gel shown is for reactions using the 3Јterminal T⅐G mispair. lated error rates (per base replicated) for base substitutions and (Ϫ1)-frameshifts at each dNTP concentration. Several conclusions can be drawn. Both base substitutions and frameshifts generally increase with increasing dNTP concentration. A (dNTP-independent) plateau value of about 6 ϫ 10 Ϫ6 appears present for the base substitutions, which the wild-type enzyme is able to exceed only at 1000 M. The MutD5 enzyme is above this level at all dNTP concentrations, although only barely so at 10 M dNTPs. This plateau may reflect mutations occurring at damaged DNA bases, a possible consequence of the construction of the gapped molecule (see "Discussion"). Each enzyme appears to produce base substitutions and frameshifts approximately equally efficiently.
There is no difference in base substitution specificity between the two holoenzymes (Table III). Under reaction conditions where base substitutions are significantly elevated above the background (1000 M dNTP for wild-type and 200 or 1000 M for mutD5), both transitions and transversions are produced. The most frequent types of base substitutions are G3 C (G⅐dGMP mispair) and G3 T (G⅐dAMP mispair) transversions followed by C3 T (C⅐dAMP mispair) and T3 C (T⅐dGMP mispair) transitions. Other base substitutions are rare or not observed. Likewise, the frameshift specificity of the two enzymes is similar; there is preferential loss of G, followed by loss of T; loss of templates C and A is relatively rare.
Comparing the wild-type and MutD5 HE, the proofreading a 3Ј-Terminal base opposite template T (see Fig. 1).   Base substitutions  42  11  31  16  17  13  22  16  14  Transitions  A3G  4  0  0  0  0  0  1  1  0  T3C  3  2  2  3  3  2  1  4  2  G3A  1  3  1  1  1  0  0  0  0  C3T  1 7  5  1 6  7  3  4  9  3  3  Transversions  G3T  6  0  5  2  4  2  2  2  4  C3A  1  0  0  0  0  0  3  1  0  G3C  7  0  6  2  6  1  5  5  5  C3G 1 0 defect of mutD5 increases the base substitution rate by about 3-fold, at least at 200 and 1000 M dNTP (Table IV). The effect cannot be estimated at the lower dNTP levels due to the proximity to the background. The effect on (Ϫ1)-frameshifts is likewise limited, 3-6-fold at 50, 200, and 1000 M dNTP, although it appears larger (ϳ14-fold) at the lowest dNTP level (10 M). In general, these results suggest only a limited contribution of proofreading (3-6-fold) to fidelity under most conditions. This effect may be contrasted to the 30 -50-fold reduction in exonuclease activity of MutD5 HE (Table I) or the 40 -200-fold increase in error rates by mutD5 in E. coli in vivo (20). Analysis of (Ϫ1)-Frameshift Mutations-(Ϫ1)-Frameshifts mutations are a frequent type of mutations produced by HE and were a dominant event among mutations induced in vitro by polIII ␣ subunit (14). In contrast, they are much less prevalent among in vivo DNA replication errors (20). We have therefore carried out a more detailed analysis of the frameshifts produced in vitro. Two general models have been advanced in the literature to explain the occurrence of (Ϫ1)frameshift mutations during DNA replication (for review, see Ref. 42). According to the Streisinger model (43,44), frameshifts arise in homopolymeric runs as the result of slippage of primer and template strands relative to one another. The second model, suggested by Kunkel and co-workers (45,46), states that certain (Ϫ1)-frameshifts are initiated by a nucleotide misincorporation. In cases where the misincorporated base is complementary to the next template base, a relocation of this base to the next template position is possible, which will produce a primer terminus with a correctly paired 3Ј terminus and a looped-out template base. Extension of this intermediate will yield a (Ϫ1)-frameshift mutation. Below, we will analyze our frameshift data within the context of these two distinct models.  Table III as "others".  Table II and the  sequencing data of Table III. For comparison, the rates of the uncopied control were 3.4 ϫ 10 Ϫ6 and 0.18 ϫ 10 Ϫ6 for base substitutions and frameshifts, respectively. These values were not subtracted from the frequencies in this table, as they reflect the consequences of gap filling in vivo.

Fidelity of DNA Polymerase III Holoenzyme
Under the "Discussion," we will also consider a third possible model for frameshift mutagenesis based on a recently proposed "dNTP-stabilized DNA misalignment" for polymerase misincorporation errors by HE (19). One feature of the Streisinger slippage model is that the mutation frequency at homopolymeric runs increases with the length of the run (42)(43)(44). We have analyzed this relationship for the present mutations. This analysis (data not shown, but see also Fig. 2) indicates that the frequency of base loss in a run of 2 or 3 bases is about 2-fold higher than at non-run bases, but there is almost no difference between runs of length 2 and 3, and for runs of length 4 and 5 a decreasing frequency is actually observed. These data suggest that the majority of frameshifts are unlikely generated by the direct slippage mechanism.
In contrast, there are several indications that most of the present frameshifts may be generated by the "misincorporation plus slippage" mechanism. First, the frameshifts appear proofread with equal efficiency as the base substitutions (Table IV). Second, there is a positive correlation between the sequences at which base substitutions and (Ϫ1)-frameshifts occur; base substitutions occur most frequently at template Gs followed by Ts and Cs, with errors at As least frequent. The same order is followed by the (Ϫ1)-frameshifts; loss of G is most frequent, followed by loss of T, with loss of C and A being least frequent (Table III). Third, there is a clear 5Ј-neighbor effect for the template sites at which the frameshifts occur; most of them are found at sites which have C or T as 5Ј-template neighbor (see Fig. 2; analysis not shown). This preference is consistent with the misincorporation plus slippage model, as purine⅐purine oppositions are the most susceptible substrates for misalignment mutagenesis (45,46). More direct evidence is presented below.
Effect of dNTP Pool Biases on the Fidelity of pol III Holoenzyme-To obtain more insight into the mechanisms by which HE generates mutations, we conducted a series of experiments involving dNTP pool biases. In these experiments, one of the four deoxynucleoside triphosphates is present in excess (or deficit) over the others. This imbalance affects the competition between correct and incorrect nucleotides and therefore the misinsertion rate. (A second, more indirect effect of dNTP biases would be an effect on the extension from misinserted bases analogous to the "dNTP effect" observed for the data of Tables  II-IV). A series of reactions was performed at 50 M dNTPs with one of the four dNTPs enhanced 20-fold to 1000 M (20ϫ bias). The data (not shown) indicated that, compared with the reaction where all four dNTPs were equal at 50 M, the 20ϫ biases are mutagenic for the MutD5 HE by 2-2.5-fold, whereas the wild-type enzyme remains essentially at background levels. We analyzed the consequences of these pool imbalances in more detail by sequencing mutants from the MutD5 reactions (Table V).
Increasing the concentration of dGTP, dATP, or dTTP increased the frequency of both base substitutions and frameshifts by severalfold (Table V). Increased dCTP had no effect on base substitutions, although it significantly enhanced the frameshifts. The dNTP pool imbalances also produced highly specific changes in the pattern of substitutions, consistent with the general predictions for such experiments as follows: high dATP will enhance misincorporation of dAMP opposite any given template residue, predicting increased substitutions to T from G, C or A; likewise, high dGTP should promote changes to C; high dTTP, changes to A; and high dCTP, changes to G. Indeed, Table V shows that the 20-fold increase in dATP increases C3 T transition by 7-fold; the 20-fold increase in dGTP increases T3 C and G3 C by more than 7-fold; and the 20-fold increase in dTTP enhances G3 A and C3 A by more than 3-fold. Excess of dCTP did not cause specific increases in any type of substitution.
Effect of dNTP Pool Imbalances on (Ϫ1)-Frameshifts-If, as suggested above, many of the (Ϫ1)-frameshift mutations created by HE in vitro are initiated by nucleotide misincorporation, specific predictions can be made for the effect of the dNTP pool imbalances on frameshift mutations, particularly with regard to the identity of the 5Ј-neighbor of the lost template base. For example, high dATP would increase the probability of misincorporation of dAMP, but since A is complementary to only T, we would predict to see increased (Ϫ1)-frameshifts only at template sites that have a 5Ј-neighbor T. Similarly, high dGTP would be predicted to promote loss of bases with a 5Јneighbor C, and so on. The data in Table VI clearly confirm these predictions. The loss of G or C with 5Ј-neighbor T is significantly increased in the reaction with 20-fold increased dATP (underlined in Table VI) but not in the reaction with 20-fold increased dGTP, dTTP, or dCTP. Similarly, a significant increase at sites with 5Ј-neighbor A is seen only in the high dTTP reaction, and increased loss of bases with 5Ј-neighbor G is seen only in reactions with excess dCTP. Thus, these data are in excellent agreement with the misincorporation plus slippage model.
The Fidelity of pol III Holoenzyme in Vivo-To compare more directly our in vitro fidelity results with those occurring in vivo, we propagated phage mRS65, whose lacI gene provides the target for the in vitro studies, in wild-type (NR9102) and mutD5 (NR11594) strains and measured the mutation rate for the same lacI d target. The data in Table VII indicate that there are two significant in vivo/in vitro differences. First, the mutD5 proofreading deficiency reduces the fidelity of the HE in vivo by 44 -100-fold, compared with much smaller values (3-6-fold) in vitro (Table IV). Second, there is a clear difference in specificity; in vivo most mutations are base substitutions, rather than frameshifts. Furthermore, analysis of the few (Ϫ1)-frameshifts observed in vivo showed them to occur in majority in runs of more than 2 bases (data not shown, see also Refs. 47 and 48), probably reflecting the Streisinger mode of frameshift mutagenesis. The observed in vivo/in vitro differences are not due to the operation in vivo of DNA mismatch repair because experiments in mismatch-repair-defective strains yielded similar results (data not shown).

DISCUSSION
The purpose of this work is to understand the mechanisms by which E. coli DNA polymerase III HE achieves its high accuracy of chromosomal replication. This accuracy has been estimated on the basis of genetic experiments in the lacI gene on FЈprolac to be approximately 5 ϫ 10 Ϫ8 per base pair per round of replication (prior to the action of mismatch repair), although the precise accuracy generally depends on the type of error (20). The contribution of proofreading to this fidelity in these studies was estimated to range from 40-to 200-fold (20). Here we have used an in vitro gap-filling assay that was previously used to assay the fidelity of the isolated ␣ subunit (14). To investigate the possible contribution of exonucleolytic proofreading, we purified the HE from both a wild-type and the proofreading-impaired mutD5 mutator strain.
The Exonuclease Activity of the Wild-type and MutD5 Holoenzyme-Exonuclease measurements on matched and mismatched primer termini confirmed the proofreading deficiency associated with the MutD5 HE. The exonuclease activity was reduced between 30-and 50-fold (Table I). These results are very consistent with previous measurements with partially purified polIII* preparations from the two strains (49), although similar experiments with pol III core have suggested a smaller difference (32). Our data also indicate that mutD5 has a small amount of remaining exonuclease activity (3.5% or less). As the proofreading deficiency of mutD5 has been used previously to estimate the contribution of proofreading to in vivo replication fidelity (50 -200-fold) (20), those previous estimates must be considered underestimates (see also Ref. 36).

The Fidelity and Specificity of Wild-type and MutD5
Holoenzyme-In our fidelity assay, HE produced two main types of mutations, base substitutions and (Ϫ1)-frameshifts, which occurred generally at similar rates (see Table IV). When viewed from the perspective of the mutations produced by HE in vivo, deduced from in vivo mutation spectra (see Ref. 20 and Table  VII), this is an intriguing result. In vivo, base pair substitutions strongly outnumber frameshifts. The simplest interpretation of these data, supported by analysis of the respective error rates (see below), is that HE in vitro produces an unexpectedly high frequency of (Ϫ1)-frameshift mutations. In the following, we will discuss various aspects of our data, including the base substitution fidelity and specificity of HE, the mechanisms of frameshift mutagenesis, the contribution of proofreading to fidelity, the comparison of HE to the ␣ subunit (whose fidelity was studied previously in the same system), and the in vitro/in vivo comparison of HE.
The Base Substitution Fidelity and Specificity of pol III HE-The base substitution fidelity of wild-type HE is 6 -7 ϫ 10 Ϫ6 per nucleotide replicated at the series of lowest dNTP concentrations (Table IV). However, the presence of a plateau (at 10, 50, and 200 M dNTPs) suggests that this represents a background level. A background is expected for two reasons. First, the DNA substrate used in the in vitro reactions is the product of multiple rounds of replication by HE in vivo. The accumulated mutations from these rounds present a background against which a single round in vitro (if equally accurate) may not be detectable. Second, the gapped DNA, due to its in vitro construction, is subject to in vitro manipulations that are likely to introduce some low level of mutagenic DNA damage. The latter possibility is also suggested by the pattern of the mutations at the plateau level: a majority of C3 T, G3 T, and G3 C mutations (see Table III). These mutations are actually a constant, low level presence in the base substitution spectrum of  Ͻ2.0 Ͻ5.9 Ͻ5.9 Ͻ4.5 Ͻ5.9 5Ј C-A (21) 4.1 Ͻ5.9 Ͻ5.9 Ͻ4.5 11.7 5Ј G-A (24) Ͻ2.0 Ͻ5.9 11.5 8. virtually all enzymes tested in vitro (30) 2 and are generally consistent with (i) deamination of cytosines (producing C3 T) and (ii) damage at guanine sites (possibly oxidative damage, causing both G3 T and G3 C mutations). Indeed, transfection of the uncopied control DNA displays the same frequency and specificity (Tables II and III). In view of these considerations, only a minimum estimate for the base substitution fidelity of HE at low dNTP can be obtained from our measurements, corresponding to a maximum error rate of 1-3 ϫ 10 Ϫ6 . The actual rate could be significantly less. Note that the estimated in vivo rate is ϳ5 ϫ 10 Ϫ8 (20). Also, the local dNTP concentrations at the E. coli replication fork are unknown, although average cell concentrations range between 10 and 50 M, depending on the nucleotide (51). The MutD5 HE produces base substitutions significantly above the background level at dNTP concentrations of 50 M or more. At 50 M dNTPs, the base substitution rate can be estimated at ϳ2 ϫ 10 Ϫ6 (corrected for the plateau level, see Table IV), compared with the inferred rate of 3.5 ϫ 10 Ϫ6 for a single round by MutD5 HE in vivo (Table VII). However, since the in vitro spectrum of base substitutions at 50 M dNTPs still resembles the background pattern (high C3 T, G3 T, and G3 C, see Table III), caution is required. It is likely that some of the mutagenesis at DNA damage sites is proofreading-dependent. If so, the true in vitro base substitution spectrum of MutD5 HE may still remain hidden. In vivo mutation spectra recorded in mutD5 strains on both the E. coli chromosome (20,48) and single-stranded phage (this study, data not shown) do not reveal the inferred damage-specific mutations (C3 T, G3 T, and G3 C), but instead show a predominance of transitions (both A⅐T3 G⅐C and G⅐C3 A⅐T).
Better insights into the base substitution fidelity of HE are obtained from the pool bias experiments (Table V). When applying a 20-fold excess of either dNTP, no increases were observed for the wild-type enzyme, but significant increases were seen for the MutD5 enzyme. DNA sequencing revealed that the following errors were specifically enhanced (from Table V): G3 A (Ͼ16-fold), T3 C (27-fold), C3 T (8-fold), G3 C (7-fold), and C3 A (4-fold), suggesting that these errors are part of the normal error spectrum of MutD5 HE. One complication of these experiments is that in certain cases the elevated dNTP also represents the next correct nucleotide, creating a possible "next nucleotide" effect (squared response) (15,39,41). Inspection of all the mutable sites in the target (data not shown) indicates that only two cases suffer from this complication. Among the G3 C mutations induced by a 20ϫ dGTP pool bias, 8 of 10 occurred at six 5Ј-CG sequences, whereas only 2 occurred at the other 10 scorable, non-5Ј-CG sequences. Second, among C3 T errors more than half (15/25) occurred at a single 5Ј-TC site (position 90). Eliminating all G3 C mutations and the specific C 3 T mutations at position 90, we suggest that the following mispairing errors are part of the normal MutD5 HE error spectrum: G3 A (4.1 ϫ 10 Ϫ6 ), T3 C (5.2 ϫ 10 Ϫ6 ), C3 T (9.0 ϫ 10 Ϫ6 ), and C3 A (1.3 ϫ 10 Ϫ6 ), the numbers in parentheses indicating the extrapolated rate under equal pool conditions (for example, for G3 A transitions under the 20ϫ dTTP condition (from Table V): 31.5 ϫ 10 Ϫ5 /(20 ϫ 7 ϫ 0.55) ϭ 4.1 ϫ 10 Ϫ6 ). These extrapolated rates are generally consistent with the mutD5 mutation rate in vivo. They are not only in the expected range (1-5 ϫ 10 Ϫ6 ) but also have the expected specificity (20), as they favor the production of transition over transversion mismatches. The recent in vitro study by Bloom et al. (19) also indicated that the G⅐T transition mismatch was produced more readily than the G⅐A or G⅐G transversion mismatch.
Our conclusions on the specificity of base mispairing specificity, as deduced from the pool imbalance experiments, differ from those of Fersht and Knill-Jones (16,17), who argued, based on fidelity studies of in vitro ⌽X174 RF3ssDNA synthesis by HE, that transversions occur at equal frequencies as transitions and that they are mediated primarily by purine⅐purine mispairings. In our view, the observed purine⅐purine mispairings are likely in vitro artifacts due to lesions that are inevitably present in purified DNAs (note that a lesion incidence in the range of one per million may be sufficient to give a measurable response). We also note that Fersht and Knill-Jones (17), when comparing ⌽X174 am16 replication in vivo and in vitro, concluded that the most prominent errors observed in vitro were, in fact, not the prominent errors that they had deduced to occur in vivo. Similarly, sequencing studies on the lacI mutants generated in phage M13 (Table VII) 3 have revealed features that are best explained by assuming that even carefully purified DNAs contain damaged purines (in addition to deaminated cytosines) which give rise to purine⅐purine mispairings in vitro.
The Mechanism of (Ϫ1)-Frameshift Mutations Produced by polIII HE-Among the two models that have been formally proposed to account for polymerase-mediated frameshift errors, Streisinger slippage and "misincorporation followed by slippage," our data are most consistent with the latter model, as already indicated under "Results." The model is delineated in Fig. 3A. In case of the ␣ subunit, which produces a similarly high level of (Ϫ1)-frameshift mutations (14), we have proposed that this enzyme has particular difficulty extending base⅐base mismatches (as suggested by the virtual absence of base substitutions among the mutations produced by this enzyme), and this difficulty has since been confirmed experimentally. 4 Instead of directly extending the mismatch, misalignment of the incorrectly inserted base on the next template base, if complementary, provides a matched primer terminus, which can be more readily extended, yielding the (Ϫ1)-frameshift. We suggest that HE produces frameshift mutations in the same manner. The strongest support for this comes from the dNTP pool bias experiments discussed above (Table VI), which promoted specific patterns of frameshift mutations that could be predicted by the applied pool bias and the nature of the 5Ј-template base. Also, the loss of G, the preferred mutation, in the context 5Ј-TG and 5Ј-CG is generally consistent with the fact that transversion mismatches, particularly purine⅐purine oppositions, are more difficult to extend (52,53), increasing the probability of indirect extension through the misalignment pathway.
An alternative model to explain the frequent production of (Ϫ1)-frameshift mutations is suggested by the recent work of Bloom et al. (19). These authors measured misincorporation by HE opposite G on an oligonucleotide template and noted a correlation between frequent misincorporation and the nature of the next template base: G misincorporation was high when the next nucleotide was C, whereas T misincorporation was high when the next nucleotide was A. It was suggested that these events actually comprise correct incorporations opposite the next template base with the polymerase simply "skipping over" the template G (see Fig. 3B). This mechanism was termed dNTP-stabilized DNA misalignment. Although this study focused on misincorporation and did not measure frameshift mutations, the proposed misaligned intermediate is formally identical to the misaligned intermediate created by misincorporation and slippage as proposed here (compare Figs. 3, A and  B). Experimental distinction between the two models may be difficult since they share a common misaligned intermediate, and the subsequent processing of the intermediate is likely to be identical. In favor of the dNTP-stabilized misalignment model, the apparent K m for misincorporation in the "error prone" context was observed to be in the range more typical of correct incorporations (19). Curiously, the dNTP-stabilized misalignment model applied only to HE reconstituted with the polIII core but not when reconstituted with the purified ␣ subunit. In the present experimental system, both ␣ subunit (14) and HE (this study) display the propensity for (Ϫ1)-frameshift mutations. Furthermore, experiments in which the ␣ subunit was presented with a preformed terminal mismatch in a sequence context where misalignment of the mismatched base on the next template base was possible showed that misalignment extension was greatly favored over direct extension (14). Similar results have now been obtained with HE. 5 These observations suggest that, at least, the minimum requirements for the misincorporation plus slippage mechanism have been met. However, further experiments will be required to settle this interesting question. Possibly, both mechanisms contribute to the high frequency of frameshift mutagenesis in vitro.
The Contribution of Proofreading to Fidelity-Our data indicate an apparently low contribution of exonucleolytic proofreading to fidelity in vitro. For example, the data of Table IV suggest an about 3-fold contribution for the base substitutions (at 200 and 1000 M dNTP), and a 3-to 14-fold contribution for the (Ϫ1)-frameshifts (10 -1000 M dNTP). We have also performed fidelity assays (at 50 M dNTP) in the presence of various concentrations of dGMP (up to 20 mM). Only very limited increases in error rate (less than 2-fold) were observed for the wild-type HE and no increases for the MutD5 enzyme (results not shown). Sequencing of mutants from these reactions showed no obvious change in specificity. These data are generally consistent with a limited effectiveness of proofreading during in vitro DNA synthesis.
The limited extent of proofreading in vitro is intriguing in view of the 40 -200-fold contribution of proofreading for various base substitutions in vivo (20) or the 100-fold contribution for frameshifts (e.g. Table VII) in the same target in vivo. We cannot exclude the possibility that at dNTP concentrations below 10 M, the contribution of proofreading will be larger, possibly approaching in vivo levels. Unfortunately, at lower dNTP levels both the wild-type and the MutD5 HE become too accurate to be measured in vitro. The effective dNTP concentrations at the replication fork are not known, although the average concentrations in the cell have been measured to be in the range of 10 -50 M (51). Another possible explanation relates to our suggestion that the majority of base substitutions (for either enzyme) at unbiased dNTP levels reflect mutations occurring at DNA lesions. Proofreading probably plays a more limited role at DNA lesions than at normal base⅐base mispairs, as can be inferred from several studies (54 -58). (This likely reflects the fact that proofreading at lesions removes the incorporated base, but not the lesion, and that hence multiple attempts can contribute to successful bypass.) In addition, no proofreading is expected in the case of C3 T transitions resulting from cytosine deamination, as the intermediate U⅐A mispair is "correct." It is likely, however, that the (Ϫ1)-frameshifts are mediated by genuine base⅐base mismatching, as they respond in a proportionate manner to pool biases (Table VI).
Our data show that pol III HE exonuclease is highly efficient in removing a terminal mismatch in the direct exonuclease assay (Table I). This indicates that the exonuclease is very active when first binding to a 3Ј terminus, but the observed proofreading efficiency during in vitro synthesis suggests that this activity is much reduced during ongoing DNA synthesis. One possible explanation for this discrepancy may be that during the initial binding of the polymerase to a primer terminus the 3Ј-terminal base is preferentially bound in the exonuclease site, as has been demonstrated to be the case for T7 DNA polymerase (59). This would lead to a high probability of excision. In contrast, during ongoing DNA synthesis the terminal mismatch is located in the polymerase active site, and a kinetic barrier may exist for transfer to the exonuclease active site, limiting the effectiveness of the exonuclease for proofreading. It is likely that such limitation, if indeed occurring, would be negated or circumvented efficiently within the replication complex in vivo (see also below).
Comparison of the Fidelity of pol III HE and ␣ Subunit-In our previous study of the fidelity of the isolated ␣ subunit, the enzyme was found to produce base substitutions and frameshifts (at 1000 M dNTPs) at a rate of 11 ϫ 10 Ϫ6 and 41 ϫ 10 Ϫ6 , respectively (14). These rates can be compared with 35 ϫ 10 Ϫ6 and 32 ϫ 10 Ϫ6 , respectively, for the proofreading-defective MutD5 HE (Table IV). This comparison indicates that the fidelity of HE is not drastically different from that of the ␣ subunit (at least at 1000 M dNTP), implying, interestingly, that the pol III HE auxiliary subunits (␤ subunit and DnaX complex) do not contribute measurably to replication fidelity in vitro.
For base substitutions, it appears that, in fact, HE produces significantly more mutations than the ␣ subunit (35 ϫ 10 Ϫ6 versus 11 ϫ 10 Ϫ6 ). This is most simply explained based on the dissociative properties of the single ␣ subunit. As this enzyme has very limited capacity to extend base⅐base mispairs, the most likely event following a misincorporation will be dissociation from the template, particularly at sequences that do not allow misalignment, preventing mutation fixation. However, the presence of the ␤-clamp in HE will prevent or strongly reduce dissociation, increasing the chance for mutation fixation. That processivity factors can be intrinsically mutagenic has also been pointed out by others (50,60).
The fact that ␣ subunit and HE display the same propensity for frameshift mutagenesis in vitro suggests that this propensity reflects an intrinsic property of the ␣ subunit that is retained in the HE. ␣ subunit has particular difficulty extending base⅐base mismatches compared with most other enzymes (14), but to account for frequent frameshift production it must also have a high propensity for either forward misalignment or extension of misaligned structures. In view of the difficulty of ␣ subunit to extend mispaired termini, the former may be most likely. It is probable that such misalignment would require only limited "breathing" of the primer terminus within the bound polymerase. The tendency to allow breathing is likely an intrinsic property of the polymerase and would therefore be 4 ?. Mo and R. M. Schaaper, manuscript in preparation. largely unaffected by the presence of the auxiliary subunits.
Comparison of the Fidelity of Pol III Holoenzyme in Vitro and in Vivo-When comparing the fidelity of HE in vitro and in vivo, two general differences are noted in our study as follows: an increased production in vitro of a distinct class of (Ϫ1)frameshift mutations and a significantly reduced efficiency in vitro of proofreading. In the simplest case, the two phenomena are related, reflecting a particular property of HE that is expressed differentially in vitro and in vivo. The in vivo replication complex functions in a highly coordinated fashion, performing coordinate synthesis of both leading and lagging strands and involving further coordination with the helicase (DnaB) and primase (DnaG) activities. It is also likely that the production of an error in one strand and the possible resulting delay have repercussions for the replication rate in the other strand and for the rate of unwinding by the helicase. Thus, we propose that under the highly controlled conditions of the in vivo replication fork additional constraints are placed on the behavior of the HE such that the efficiency of error processing is enhanced. This could lead to both reduced production of misalignment errors and increased proofreading. It will be of interest to investigate the precise nature of these additional in vivo mechanisms that have the potential to extend the study of replication fidelity beyond the generally accepted two-step scheme of base selection and proofreading.