Multivariate discrimination for phenylketonuria (PKU) and non-PKU hyperphenylalaninemia after analysis of newborns' dried blood-spot specimens for six amino acids by ion-exchange chromatography

¹ Structural Pathology Laboratory and
² Laboratory of Newborn Screening and Genetic Services, Wadsworth Center, New York State Department of Health, P.O. Box 509, Albany, NY 12201-0509.
Departments of
³ Biometry and Statistics and
⁴ Biomedical Sciences, School of Public Health, University at Albany, One University Place, Rensselaer, NY 12144-3456.

Abstract

Ion-exchange HPLC was developed for testing dried blood-spot specimens from newborns. The method is suitable for quantitative confirmatory testing of abnormal specimens detected in the New York State Newborn Screening Program. Positive specimens were initially identified among all New York State newborns with semiquantitative bacterial inhibition assays (BIA) for aminoacidopathies, including phenylketonuria (PKU) and non-PKU hyperphenylalaninemia (HP), maple syrup urine disease, and homocystinuria. A selection of 1346 specimens from routine BIA screening, including 131 newborns with PKU or persistent HP, were tested by HPLC. Of 179 BIA results that were falsely positive, 98 (55%) were also falsely positive by HPLC in which the Phe/Tyr ratio was the discriminator and the threshold was set to attain 100% sensitivity. Investigation of three multivariate discriminatory methods revealed that linear discriminant analysis excluded all but 35 (20%) of the BIA false-positives.

The New York State Newborn Screening Program (NYSNSP)¹ screens ~280 000 births annually, using dried blood specimens and bacterial inhibition assays (BIAs) to test for phenylketonuria (PKU) and non-PKU hyperphenylalaninemia (HP), maple syrup urine disease (MSUD), and homocystinuria. The program tries to simultaneously maximize sensitivity and specificity by setting low thresholds to avoid false-negatives (to date, none have been reported to NYSNSP) and performing multiple retests of positive specimens. Nevertheless, false-positives (FP_Phe) occur when the original dried blood-spot specimen gives positive results in three BIAs but the subsequent recall specimen is BIA-normal or the Specialty Center diagnosis is normal (unaffected). Potential causes for FP_Phe are immaturity of the Phe hydroxylase system or a regulatory defect affecting the system, defects of biopterin metabolism, prematurity, liver damage, and parenteral amino acid nutrition (1)(2)(3). Poor accuracy and precision of the testing method, interfering substances in blood, and specimen quality can also increase the FP_Phe rate (4)(5)(6). Investigation of the utility of performing HPLC analysis after a positive BIA result is motivated by the NYSNSP's wish to identify all infants with permanent hyperphenylalaninemia while avoiding the monetary and psychological costs of unnecessarily requesting recall specimens from pediatricians and making Specialty Center referrals. We report here a retrospective study of the HPLC technique, comparing results for infants with false-positive BIA test results (FP_Phe) with those for confirmed PKU and HP newborns.

Screening newborns for PKU with dried blood-spot specimens was initiated by NYSNSP in 1965, which used the Guthrie BIA (7) to measure increased concentrations of Phe. A variety of methods have subsequently been used to measure Phe in dried blood-spot specimens, including fluorometric (8)(9), enzymatic (10), HPLC (11)(12)(13)(14), and tandem mass spectrometry (4)(15). Measurement of Tyr is also desirable, given the observation of Qu et al. (14) that the Phe/Tyr ratio, computed as the ratio of the individually measured concentrations, is useful for discrimination by HPLC. Qu et al. examined 210 newborns, including 9 PKU patients, and a series of 50 plasma specimens from PKU patients; the Phe/Tyr ratios were always >2. Chace et al. (4) obtained similar results with tandem mass spectrometry in a set of 274 subjects that included 8 PKU patients for whom the minimal Phe/Tyr ratio was 2.6. These two studies showed 100% sensitivity and specificity for PKU when the Phe/Tyr ratio threshold was set to 2.

The HPLC method is a modification of the method of Shapira and colleagues, ion-exchange chromatography with postcolumn ninhydrin detection of amino acids (14). The method separates 6 amino acids, Val, Met, Ile, Leu, Tyr, and Phe in 35 min. HPLC offers the advantages of automation, high accuracy, and multiple determinations for hyperphenylalaninemias, MSUD, and homocystinuria in a single run. However, although HPLC yields precise and accurate concentrations for 6 amino acids, it gives no indication of how these measurements should be used to achieve either maximal classification accuracy or minimum numbers of false-positives. Here, we report the utility of previously proposed and new scaler discriminators in an extended database of 131 cases of PKU and HP and 1215 control patients. The performance of 18 newly proposed multivariate discriminators—simultaneously incorporating age, birth weight, and the six HPLC amino acid concentrations—is also presented, and the best of these multivariate discriminators is shown to be superior to any of the scaler discriminators.

The new multivariate discriminators were constructed by three techniques designed to find the combinations of the 8 variables that best separate the populations. Two classical multivariate discriminatory techniques, linear discriminant analysis (LDA) and quadratic discriminant analysis (QDA) (16), were compared with a newer technique, flexible discriminant analysis (FDA) (17). LDA estimates the fewest parameters and is appropriate when all of the precision matrices in the 8 variables are identical. QDA is appropriate when, as in the current data, this condition does not hold but instead requires estimating a large set of parameters. FDA, in requiring more parameters than LDA but fewer than QDA, has the potential to balance the number of estimated parameters against achievable performance. We undertook investigation of all three methods because it was not evident a priori which would be most successful in the discrimination after linear combinations capturing the information contained in this complicated 8-variable data set were introduced as new simplifying variables.

Materials and Methods

equipment and reagents
The 3 x 150 mm high-efficiency cation-exchange column, PCX3100 postcolumn derivatization instrument, lithium citrate pH 2.8 and 7.5 eluents and pH 2.2 diluent, regenerant, and ninhydrin were from Pickering Labs. The Model 717 autosampler, Model 486 detector, and Millennium and Model 845 computer programs were from Waters. The GP40 gradient pump was from Dionex.

Liquid standard and physiological amino acid calibrators were from Sigma Chemical Co. The filter paper used for dried blood specimens was S&S 903 (Schleicher & Schuell). Quality-control blood spots from the 1995 Newborn Screening Quality Assurance Program were obtained from the Centers for Disease Control and Prevention (CDC). In-house calibrators and an in-house control were made from Red Cross packed cells by determining the endogenous concentrations of amino acids, supplementing the cells with the appropriate amino acid concentrations, adjusting the hematocrit to 55%, and spotting 50-µL aliquots onto filter paper. The specimens were then dried overnight and stored desiccated at -14 °C. The volume of blood per 3.2-mm-(1/8-in.-)diameter disk was 3 µL, determined with [I]thyroxine as previously described (18). All dried blood-spot concentrations are given in units of whole blood.

specimen storage and chromatography
The 646 specimens from newborns with PKU or HP, the specimens determined to be false-positives, and those with no BIA growth because of antibiotic inhibition were stored at -14 °C for 13 days to 2.5 years (median 0.36 years) before analysis. The 700 remaining specimens were stored at 4 °C for 10 days to 6 months and thereafter at -14 °C for up to 1.1 year (the median total storage time for these specimens was 28 days). Because others (19)(20)(21) have reported that amino acid concentrations in dried blood-spot specimens change over long periods, we investigated the stability of the six assayed amino acids in our specimens. Decay coefficients were estimated from 43 patients' blood spots analyzed at two times: 0.1–2.4 years (median 0.4) after collection and 1.4–4.0 years (median 2.6) after collection.

All dried blood spots were analyzed according to the following protocol. Two ~3.2-mm-(1/8-in.-)diameter disks were punched from dried blood specimens previously tested by NYSNSP. The contents of the disks were eluted in 100 µL of 1 mmol/L HCl by sonication at room temperature for 30 min in a 1.5-mL microcentrifuge tube. Proteins were precipitated by addition of 100 µL of 60 g/L sulfosalicylic acid solution and kept at 4 °C for 15 min. After the addition of 100 µL of lithium diluent, pH 2.2, the samples were centrifuged at 8000g for 10 min. The supernatants were transferred to assay vials for HPLC analysis.

The column was equilibrated with buffer A (Li pH 2.8):buffer B (Li pH 7.5), 80:20 by vol, at 39 °C at a flow rate of 0.3 mL/min for the buffers and 0.2 mL/min for the ninhydrin. After injection of the sample (100 µL), chromatography proceeded with a linear gradient to 42:58 (by vol) buffers A:B at 17 min. After a 4.5-min wash with regenerant, the column was reequilibrated with 80:20 A:B, as above. Total assay time was 35 min. Absorbance at 570 nm was monitored for elution of amino acids.

chromatography performance
Calibration curves for the liquid amino acid calibrators were linear from 1 to 50 µmol/L, equivalent to blood spot concentrations of 25–1250 µmol/L. For Phe, the calibration parameters were: slope = 61.6 (SE = 0.133); intercept = -1.09 (SE = 3.10); r = 0.9999; and root mean square error (RMSE) = 8.02. For Leu, slope = 63.12 (SE = 0.134); intercept = -0.67 (SE = 3.13); r = 0.9999; and RMSE = 8.09. For Met, slope = 62.15 (SE = 0.116); intercept = 4.18 (SE = 2.71); r = 0.9999; and RMSE = 7.01.

The recoveries of added amino acids from in-house dried blood-spot calibrators ranged from 80% to 103% for Phe in the range 121-1210 µmol/L; from 87% to 100% for Leu (152–1524 µmol/L); and from 80% to 91% for Met (67–804 µmol/L). The between-assay imprecision (CV) of the in-house control for Phe, Leu, and Met at their cutoff values of 180, 305, and 101 µmol/L, respectively, was 6.5%, 7.4%, and 7.6% for 26 replicate samples obtained over a period of 4 months. The within-assay imprecision was <4.2%. The detection limit, measured as the concentration at a signal 3 SD times the baseline noise, was 3.6 µmol/L for Phe, 2.3 µmol/L for Leu, and 2.0 µmol/L for Met. The physiological concentration of the amino acids we obtained in healthy newborns is 27 times the detection limit for Phe, 68 times for Leu, and 15 times for Met.

The accuracy of the method was determined with the blood spots supplied by the CDC Newborn Screening Quality Assurance Program. The HPLC amino acid concentrations were compared with the CDC values. For Phe at CDC-measured concentrations of 66.6–424 µmol/L (n = 11), the regression parameters were: slope = 1.0 (SE = 0.022); intercept = -3.75 (SE = 6.10); r = 0.998; RMSE = 0.287. For Leu over the range 73.7–429 µmol/L (n = 9), they were: slope = 0.95 (SE = 0.02); intercept = 7.30 (SE = 9.25); r = 0.9883; RMSE = 0.30. For Met over the range 73.7–429 µmol/L (n = 15), they were: slope = 0.83 (SE = 0.32); intercept = -8.8 (SE = 7.52); r = 0.9799; RMSE = 0.21. Although the CDC dried blood-spot specimens are spotted from lysed blood, which is equivalent to ~6% less volume per 3.2-mm (1/8-in.) disk than specimens spotted from whole blood (18), we did not correct the above parameters for this difference.

None of the amino acids found in the physiological amino acid mixture interfered with the six amino acids measured. Antibiotics, however, have been reported to interfere with amino acid analyses (5)(22)(23). In our system, some newborn specimens that exhibited antibiotic inhibition of growth in the BIA had a low, broad peak eluting after Phe, resulting in minor positive biases in calculation of Phe concentrations. Autoclaving the specimen removed the antibiotic peak but diminished amino acid recovery by as much as 50%. To ascertain whether HPLC discriminant analysis could avoid BIA recalls, we included 77 unautoclaved specimens of this type in our study.

Example chromatographic profiles from the liquid amino acid calibrators, a specimen from an unaffected newborn, and one from a PKU-affected newborn are shown in Fig. 1 .

View larger version (43K): [in this window] [in a new window]   Figure 1. HPLC chromatograms of (top) liquid amino acid calibrators (250 pmol/L each), (middle) extract from dried blood-spot specimens from an unaffected newborn, and (bottom) sample from an infant with PKU (Phe concentration 339 µmol/L).

population studied
The population used to determine optimal discriminant functions was composed of five groups of infants admitted to the study, all of ages <50 days. All groups were classified on the basis of their BIA screening result. The first group, 700 newborns with "normal" BIA test results (CTRL), were selected from routine specimens, but ''difficult'' babies with marginally increased concentrations of Met, Leu, or Phe; low birth weight; and specimens taken <24 h after birth were preferentially included.

The second and third groups were 131 cases (60 PKU and 71 HP, respectively) with hyperphenylalaninemia identified from routine newborn screening utilizing BIA with a Phe threshold of 180 µmol/L (3 mg/dL) and diagnosed at approved NYS Specialty Centers. These 131 newborns represent all such infants born from 1993 through 1996 except for 15 infants for whom there was insufficient specimen for HPLC analysis. Specialty Center confirmation was attained with follow-up plasma amino acid analysis. Hyperphenylalaninemia is diagnosed in patients whose plasma Phe concentration is 180 µmol/L. These patients are differentiated into HP, for Phe concentrations of 180 to 600–720 µmol/L, and PKU, for Phe concentrations 600–720 µmol/L. Classical PKU newborns have Phe concentrations 1200 µmol/L.

The fourth group was 148 false-positive hyperphenylalaninemias (FP_Phe) included from 183 (81%) FP_Phe babies born in 1995 and 1996 for whom sufficient sample was available to reanalyze and who were found to be negative by either BIA analysis of recall blood spots or, for 10 patients, by a Specialty Center diagnosis. These were augmented with 31 similar patients from 1994 for a total of 179.

The final group (FP_Other) was composed of 232 of 289 (80%) babies born in 1995 and 1996 with sufficient sample quantity to reanalyze, who had BIA increases in Met and Leu, either alone or in combination with each other and (or) Phe. All specimens in this group were found to be negative by BIA after analysis of recall blood spots. These were augmented with 27 similar specimens collected in 1994 and with the 77 antibiotic-inhibited newborns for a total of 336. BIA results are considered positive and patients are recalled for homocystinuria evaluation if Met concentrations are 101 µmol/L (1.5 mg/dL) and for MSUD evaluation if Leu concentrations are 305 µmol/L (4.0 mg/dL).

statistical methods
The objective of the analysis was to estimate discriminating functions that successfully separate PKU, HP, FP_Phe, FP_Other, and CTRLs by finding linear combinations of the following variables: HPLC concentrations of six amino acids (µmol/L); birth weight (g); age (days). Although many discrimination methods have been developed, no theory indicates which method is best. In addition, no theory indicates how many of the above variables might be useful in the discrimination. Finally, some methods, such as QDA, because of the large number of estimated parameters used, are apt to "overfit"; that is, discrimination is excellent in the data used to estimate parameters (training data) but poor when the method is applied to other data sets (test data). For these reasons, we undertook a cross-validation study comparing the methods and an ordered selection of discriminating variables. Even though each method partitions the data into five groups (CTRL, FP_Other, FP_Phe, HP, and PKU), performance measures were based on two numbers: the true positives (HP, PKU) and the true negatives (CTRL, FP_Other, FP_Phe). In particular, the method most successfully separating HP cases from FP_Phe cases was sought. Additionally, however, the proposed protocol allowed an HPLC discrimination analysis only for samples that are BIA-positive. False-positive HPLC CTRLs were therefore reclassified as true negatives to account for the fact that BIA-negative cases do not receive HPLC analysis and therefore will have no reason to be considered anything other than negative.

The cross-validation was initiated by drawing a random sample of one-half of the 1346 newborns as a training sample and estimating parameters for each selection of variables within each method. The diagnostic sensitivity, specificity, positive predictive value, negative predictive value, and diagnostic accuracy attained with the estimates were then computed for both the training sample and the remainder of the data, the test sample. The yearly prevalences required for some of these measures were taken from NYSNSP (x 10¹ ): CTRL = 99 908; PKU = 4; HP = 6; FP_Phe = 31; FP_Other = 51. Bootstrap confidence intervals (24) for each measure were obtained by repeatedly sampling the data with replacement 300 times and using the 7th and 292nd ordered values as the estimated 95% confidence interval.

In the cross-validation, we used the discrimination methods to try to find the best way to combine the patients' variables into just two linear combinations. The values of these combinations could then be computed for each newborn and a discriminant plot created by placing the first combination's values on the abscissa and the second combination's values on the ordinate. The combinations chosen were selected to minimize the amount of misclassification by maximizing the separation of the five populations on the discrimination plot. Because it is desirable to eliminate patients' variables that do not aid the discrimination, within each method we investigated six models by initially including the best scaler discriminators—Phe, Tyr, Leu, and Phe/Sum (where Sum is the sum of the non-Phe amino acids)—and then sequentially adding Ile, Met, birth weight, age, and Val. The order of entry was determined by the amount each variable contributed to the discrimination. The data were log-normally (gaussianly) distributed, so all analyses were run on the log scale. Untransforming the linear combinations therefore led to combinations of variables that are products of ratios. The methods were also permitted to find the best three linear combinations of patients' variables to ascertain whether adding a third dimension to the discrimination plot would significantly reduce the number of FP_Phe.

Results

specimen stability
The stability of the six assayed amino acids was assessed for patients' dried blood-spot specimens examined over various storage times. Decay coefficients were estimated by robust regression (25) of log concentrations on length of storage and converted to percent yearly decay. Yearly decay rates for the amino acids were 1–6%, except for Met, for which the rate was 13%. The Phe/Tyr ratio decreased an estimated 1.4% per year.

patients' characteristics
The newborns' HPLC amino acid concentrations followed a log-normal distribution. The geometric mean Phe concentration in a sample of 493 unaffected newborns selected from the 700 CTRLs on the basis of normal birth weight (>2500 g) and age (1–5 days) was 93 µmol/L (1.5 mg/dL) with a range (±2 SD) of 57–151 µmol/L. These results for blood-spot samples are similar to the range of 38–137 µmol/L reported by Shapira et al. for amino acid concentrations in normal infants' plasma (26). Ranges for Met of 16–55 µmol/L and Ile of 27–84 µmol/L are also close to the plasma concentrations. Ranges for Val (108–305 µmol/L), Leu (84–259 µmol/L), and Tyr (72–254 µmol/L) have higher upper bounds than Shapira reports, but this is in part due to our use of log-normal distributions. Ranges for the ratios were Phe/Tyr 0.41–1.14, Phe/Leu 0.46–0.86, and Phe/Sum 0.12–0.23.

Scaler HPLC discriminators involving Phe or one of the three ratios were evaluated for their ability to eliminate the recalls associated with BIA FP_Phe results. Scatterplots show the performance of the various candidate measures when HPLC thresholds were established at either the lowest value among the 60 PKU newborns or the lowest value among the 71 HP newborns, thereby minimizing false-negatives (Fig. 2 ). If the PKU threshold of 1.93 were used for the Phe/Tyr discriminator, then 5 of the 179 (2.8%) FP_Phe samples would be recalled. If the HP threshold of 0.91 were used, 98 (55%) would be recalled. Similarly, using the discriminators Phe, Phe/Sum, or Phe/Leu with the HP thresholds yields respective recalls of 54%, 53%, and 37%.

View larger version (33K): [in this window] [in a new window]   Figure 2. False-positive errors of five discriminating functions when using PKU and HP HPLC thresholds. Performance of four scaler discriminators and the recommended multivariate LDA discriminator compared for correctly classifying 179 BIA FPPhe newborns with respect to two thresholds. In each case the upper threshold, PKU (first cutoff), is set at the minimum value observed among the PKU cases and the lower threshold, HP (second cutoff), is set at the minimum among the HP cases. Each panel also shows the number (and %) of HPLC discrimination false-positives under the two cutoffs. (To convert Phe cutoffs shown in mg/dL to µmol/L, multiply by 60.5.) Note that the LDA discriminator minimizes errors at both thresholds.

Even though the CTRLs were enriched with newborns having marginally increased amino acids, only 1 CTRL exceeded the PKU threshold for Phe/Tyr, and no CTRLs exceeded any of the other scaler discriminator thresholds. In additional, 21 (3%) of 700 CTRLs exceeded the HPLC Phe HP threshold, 68 (10%) exceeded the Phe/Tyr HP threshold, 64 (9%) exceeded the Phe/Sum HP threshold, and 68 (10%) exceeded the Phe/Leu HP threshold.

By simplifying the data through using only two linear combinations of the eight available variables, all of the multivariate procedures improved on the scaler discriminators but the discrimination regions produced were different (Fig. 3 ). The QDA method gave curved discrimination functions but, after the number of false-negatives were minimized, its regions did not closely coincide with the five groups. Use of FDA markedly improved separation between affected, CTRL, and FP_Other but, unfortunately, not for the FP_Phe. LDA therefore appears to produce the best discrimination. Cross-validation assessment of performance measures showed that LDA (Fig. 4 ) was best for every model (other methods not shown) and that all variables except Val contributed to increased performance. The final LDA linear combinations are given in the Appendix.

View larger version (51K): [in this window] [in a new window]   Figure 3. Two-dimensional scatterplots based on concentrations of six amino acids, birth weight, and age: results of applying three different discrimination techniques—LDA (A), QDA (B), and (C) FDA—to separate the five study populations. Each plot displays coded symbols for PKU (P), HP (H), FPPhe (+), FPOther (), and CTRL () that locate each newborn's results with respect to linear combinations that form each axis.

View larger version (21K): [in this window] [in a new window]   Figure 4. Five measures of performance obtained by cross-validation analysis of LDA. Cross-validation was conducted by randomly splitting the data into two equal portions, using the first portion for training (upper line in each pair of curves) and the remainder for testing (lower line in each pair) each discrimination model. The vertical lines give 95% confidence intervals (upward-pointing for the training data and downward-pointing for the testing data) obtained by repeating the cross-validation 300 times for each model; the remaining lines connect the mean values. Six models are examined. The abscissa value of 1 denotes including Phe/Sum, Phe, Tyr, Leu. The subsequent models cumulatively add: (2) Ile; (3) Met; (4) birth weight; (5) age; and (6) Val. The sensitivity and negative predictive values are similar, as are the remaining 3 plots, owing to the extremely high sensitivity (>98%). Plots for QDA and FDA (not shown) revealed FDA to be superior to QDA and inferior to LDA for every model.

For comparison with the scaler discriminators, the distances from each PKU, HP, and FP_Phe result to the HP/FP_Phe discrimination boundary have been plotted in Fig. 2 . When the HPLC threshold was established at the lowest value among the 71 HP newborns, the LDA procedure excluded all but 35 (20%) of the 179 FP_Phe.

None of the 700 CTRLs exceeded the LDA PKU threshold. However, five HPLC LDA false-positives (<1%) exceeded the HP threshold. Only 2 of the 77 specimens that exhibited antibiotic inhibition of growth in the BIA exceeded the LDA HP threshold; the remainder grouped with the CTRLs.

In addition to testing the above models, we also examined three-dimensional multivariate discriminators. The three-dimensional LDA model excluding Val marginally improves discriminatory performance by reducing the number of FP_Phe classified to CTRL. However, as noted in Statistical methods, discriminatory performance is judged on the basis of correct true-negative classification. Inasmuch as improved classification among true-negative categories does not enhance performance under this criterion, we consider the two-dimensional LDA model adequate.

Discussion

This study shows that multivariate LDA with combinations of Phe/Sum, Phe, Tyr, Leu, Ile, Met, birth weight, and age is superior to any of the previous and even newly proposed scaler (Fig. 2 ) or multivariate (Figs. 3 and 4 ) discriminators because it leads to the greatest reductions in FP_Phe. LDA achieves this success by using two separate combinations of variables (Fig. 3A ) detailed in the Appendix. Additional combinations do not further decrease FP_Phe. The most important LDA combination has Phe/Sum as its major contributor; in the second combination, Phe/Sum, Leu, and a modified Phe/Tyr ratio are the major contributors.

The Phe/Tyr ratio rather than the Phe concentration has been suggested to increase the specificity for screening for PKU. Qu et al. (14) suggested that because Tyr is the product of deficient Phe hydroxylase enzyme, the Phe/Tyr ratio should be noticeably increased (>2.0) even for borderline PKU patients. Chace et al. (6)(27) showed that using the ratio of amino acids normalizes variations from specimen quality and handling. This results in a lower CV for the ratio than for the absolute concentrations, thereby increasing diagnostic accuracy.

Our results are similar to those previously reported (4)(14). We observed 60 newborns with PKU and, of these, 59 had Phe/Tyr ratios >2 (the ratio for the remaining specimen was 1.93). For HP, however, only 56% (40 of 71) had Phe/Tyr ratios >2.0; a Phe/Tyr ratio as low as 0.91 was required to achieve 100% sensitivity. Additionally, 5 of the 179 FP_Phe specimens exceeded the suggested 2.0 threshold and 55% (98 of 179) were above the 0.91 threshold. The best univariate discriminator, the Phe/Leu ratio, achieved an HPLC false-positive rate of 37%, as shown in Fig. 2 . However, many FP_Phe specimens remained clustered near the associated 1.1 threshold—similar to the pattern observed for the other scaler discriminators but in marked contrast to the separation achieved by LDA. In fact, the lower LDA cutoff cleanly divides the FP_Phe into two populations and correctly reclassifies all but 35 as true-negatives. Investigation of the remaining 35 FP_Phe specimens revealed that 47% of these (compared with 14% of the correctly identified specimens) were from low-birth-weight infants. The remaining FP_Phe specimens might thus represent a subpopulation not identified by the current set of variables.

The success of the LDA method can be attributed to its combination of the best separating abilities of the scaler discriminators. The first linear combination (abscissa, Fig. 3A ), which accounted for 70% of the variability in the original 8 variables, has Phe/Sum as the major contributor. A rationale for introducing a ratio has already been noted, and a similar argument justifies the incorporation of Phe/Sum. Many of the FP_Phe samples have high total non-Phe amino acid concentrations. Of the 179 FP_Phe newborns, 35 (20%) had total non-Phe amino acids concentrations >2 SD of the mean for CTRL newborns, and 9 (5%) differed by >3 SD. Apparently the Phe/Sum ratio is superior to the Phe/Tyr ratio in aiding classification of these newborns as negative. The second linear combination (ordinate, Fig. 3A ), which accounted for 29% of the variability, included a product of Phe/Sum with an adjusted Phe/Tyr ratio (antilog scale) and Leu. The observation that the two linear combinations account for 99% of the variation observed in the 8 original variables suggests that only an insignificant improvement could be gained by adding another linear combination in a third dimension of Fig. 3 . Furthermore, the large contributions of each linear combination used show that they represent a minimal simplification of the data.

That Phe and Tyr were among the major discriminatory contributors is reassuring and would be expected on the basis of the known metabolic pathway. However, the cross-validation analysis revealed that incorporating other amino acids (Leu, Ile, and Met) also led to further discriminatory improvement. Some of these variables possibly reflect exogenous influences, such as hematocrit or heavy application of blood. Also, newborns receiving hyperalimentation often have high concentrations of these amino acids. Birth weight and age also made small contributions (Appendix), which coincides with the observation that 36 of the FP_Phe (20%) had birth weights <2500 g.

Although the focus of the current study was on reducing FP_Phe, the LDA method was also remarkably successful in segregating FP_Other results (Fig. 3 ). This may be useful for identifying MSUD and homocystinuria false-positives if sufficient cases can be collected to establish sensitive and specific boundaries for distinguishing patients from FP_Other subjects.

The effect of storage on amino acid concentrations was examined because some of the specimens of affected infants had been stored at -14 °C for as long as 2.5 years before analysis. The specimen stability studies showed, as expected (6), that Met had the maximum percent yearly decay. The 1–6% yearly decay in concentrations of other amino acids in patients' specimens generally agreed with previous reports (19)(20)(21). Because these decays were smaller than the between-assay CVs, we expect that variations in length of storage did not significantly alter our conclusions about the inefficiency of the scaler discriminators such as Phe/Tyr.

We have not addressed the merit of screening for PKU vs HP, even though the recommended discrimination procedure separates these patients. Adult individuals on a normal diet with serum Phe concentrations of 150–720 µmol/L have been classified as non-PKU mild hyperphenylalaninemia (28). Weglage et al. (29), studying a sample of 28 such young adults, reported that IQ, motor skills, and school/career performance were similar to those of healthy controls and therefore concluded that the affected adults do not require dietary treatment if serum concentrations of Phe are <600 µmol/L. On the other hand, others (30), including the Committee on Genetics of the American Academy of Pediatrics (31), report that infants with Phe >360–400 µmol/L should be considered for a Phe-restricted diet. We found one HP infant with an initial blood spot Phe concentration of 370 µmol/L and a Phe/Tyr ratio of 1.2, who had a follow-up serum Phe concentration of 442 µmol/L at age 1 month, 545 µmol/L at 4 months, and 653 µmol/L at 6 months; the infant was thereafter put on a Phe-restricted diet. Another infant with a blood spot Phe concentration of 211 µmol/L, a Phe/Tyr ratio of 1.3, and a follow-up serum Phe concentration of 411 µmol/L at age 1 month had a sibling with a follow-up Phe concentration of 296 µmol/L at age 2 months; the siblings were lost to follow-up at later ages. The delayed increase of Phe in HP infants has been previously reported (32). Identification of HP newborns permits follow-up testing to identify those with delayed increases of Phe, who may require dietary intervention. Genotyping may also assist in identifying at-risk newborns, given that the genotype has been shown to predict clinical phenotype (33). Monitoring cases with Phe/Tyr <2.0 therefore may be essential for achieving 100% diagnostic sensitivity.

The effects of maternal "non-PKU mild hyperphenylalaninemia" are currently being reexamined for possible fetal damage. Because of a considerable feto-maternal Phe gradient, fetal concentrations of Phe are 35% higher than maternal concentrations (34). Preliminary results of the International Maternal Phenylketonuria Collaborative Study indicate that maternal Phe concentrations as low as 360–400 µmol/L may be required to prevent both decreased head circumference at birth and decreased IQ (28)(35)(36). Identification of HP females through newborn screening may therefore also be important for tracking and possibly utilizing dietary restrictions during pregnancy.

In conclusion, HPLC ion-exchange chromatography offers accuracy, precision, and automation of screening for aminoacidopathies. HPLC is suitable for quantitative confirmatory testing of abnormal specimens detected by newborn screening programs and, combined with LDA discrimination, can reduce the recall rate of false-positives from BIA screening by 80%.

Note added in proof: During the first half of 1997, 76 additional newborns' specimens with above-normal BIA Phe values were analyzed by HPLC/LDA. This population's data set was not utilized in constructing the LDA procedure and can thus be used to assess the validity of the method for use with future data. BIA and LDA results agreed on the classification of 18 affected newborns in this group. Of the 58 unaffected, LDA classified 44 as unaffected and thus would exclude all but 24% of the FP_Phe. This is in close agreement with the cross-validation results reported in the present paper (20%).

Acknowledgments

We thank Chester Koblantz of the Newborn Screening and Genetic Services Laboratory for excellent technical support. We are grateful to the following for generously providing patient information: Peggy O'Connor and Alfred Slonim, of North Shore University Hospital, and Cheryl Clow and Marilyn Cowger, of Albany Medical College. We also thank Barbara W. Adam of the CDC for providing quality-control blood spots, and Tony Le, Pickering Laboratory, for helpful suggestions. The Computational Molecular Biology and Statistics Core Facility of the Wadsworth Laboratories provided computational support. This project was partially supported under a cooperative agreement from the CDC through the American Association of Schools of Public Health.

Footnotes

² Address correspondence to this author at: Rm. D224, Wadsworth Center. Fax 518-474-2769, e-mail Andrew.Reilly{at}Wadsworth.Org.

¹ Nonstandard abbreviations: NYSNSP, New York State Newborn Screening Program; BIA, bacterial inhibition assay; PKU, phenylketonuria; HP, non-PKU hyperphenylalaninemia; MSUD, maple syrup urine disease; FP_Phe, BIA false-positive result for hyperphenylalaninemias; LDA, linear discriminant analysis; QDA, quadratic discriminant analysis; FDA, flexible discriminant analysis; RMSE, root mean square error; CTRL, samples giving BIA results within the reference interval for unaffected subjects; FP_Other, BIA false-positive for Met and Leu, possibly in combination with Phe.

References

1. Walker V, Clayton BE, Ersser RS, Francis DEM, Lilly P, Senkins JWT, et al. Hyperphenylalaninaemia of various types among three-quarters of a million neonates tested in a screening programme. Arch Dis Child 1981;56:759-764. [Abstract/Free Full Text]
2. Zschocke J, Carson DJ. Transient neonatal hyperphenylalaninaemia is not related to mutations at the phenylalanine hydroxylase gene. J Inher Metab Dis 1995;18:369-370. [Web of Science][Medline] [Order article via Infotrieve]
3. Scriver CR, Kaufman S, Eisensmith RC, Woo SLC. The hyperphenylalaninemias. Scriver CR Beaudet AL Sly WS Valle D eds. 7th ed. The metabolic and molecular bases of inherited disease 1995;Vol. 1:1015-1075 McGraw-Hill New York. .
4. Chace DH, Millington DS, Terada N, Kahler SG, Roe CR, Lindsay F, Hofman LF. Rapid diagnosis of phenylketonuria by quantitative analysis for phenylalanine and tyrosine in neonatal blood spots by tandem mass spectrometry. Clin Chem 1993;39:66-71. [Abstract]
5. Ulrey JR, Nahata MC. Drug interferences with amino acid analysis. J Pharm Tech 1990;6:64-73.
6. Chace DH, Hillman SL, Millington DS, Kahler SG, Adam BW, Levy HL. Rapid diagnosis of homocystinuria and other hypermethioninemias from newborns' blood spots by tandem mass spectrometry. Clin Chem 1996;42:349-355. [Abstract/Free Full Text]
7. Guthrie R, Susi A. A simple phenylalanine method for detecting phenylketonuria in large populations of newborn infants. Pediatrics 1963;32:338-343. [Abstract/Free Full Text]
8. Hoffman GL, Laessig RH, Hassemer DJ, Makowski ER. Dual-channel, continuous-flow system for determination of phenylalanine and galactose: application to newborn screening. Clin Chem 1984;30:287-290. [Abstract]
9. Gerasimova NS, Steklova IV, Tuuminen T. Fluorometric method for phenylalanine microplate assay adapted for phenylketonuria screening. Clin Chem 1989;35:2112-2115. [Abstract/Free Full Text]
10. Elvers LH, Diependaal GAM, Blonk HJ, Loeber JG. Phenylketonuria screening using the Quantase phenylalanine kit in combination with a microfilter system and the dye tartrazine. Screening 1995;3:209-223.
11. Roesel AR, Blankenship PR, Hommes F. HPLC assay of phenylalanine and tyrosine in blood spots on filter paper. Clin Chim Acta 1986;156:91-96. [Web of Science][Medline] [Order article via Infotrieve]
12. Rudy JL, Rutledge JC, Lewis SL. Phenylalanine and tyrosine in serum and eluates from dried blood spots as determined by reversed-phase liquid chromatography. Clin Chem 1987;33:1152-1154. [Abstract/Free Full Text]
13. Vollmer DW, Jinks DC, Guthrie R. Isocratic reverse-phase liquid chromatography assay for amino acid metabolic disorders using eluates of dried blood spots. Anal Biochem 1990;189:115-121. [Web of Science][Medline] [Order article via Infotrieve]
14. Qu Y, Miller JB, Slocum RH, Shapira E. Rapid automated quantitation of isoleucine, leucine, tyrosine, and phenylalanine from dried blood filter paper specimens. Clin Chim Acta 1991;203:191-198. [Web of Science][Medline] [Order article via Infotrieve]
15. Rashed MS, Ozand PT, Bucknall MP, Little D. Diagnosis of inborn errors of metabolism from blood spots by acylcarnitines and amino acids profiling using automated electrospray tandem mass spectrometry. Pediatr Res 1995;38:324-331. [Web of Science][Medline] [Order article via Infotrieve]
16. Gnanadesikan R, Blashfield RK, Brieman L, Dunn OJ, Friedman JH, Fu K-S, et al. Discriminant analysis and clustering. Stat Sci 1989;4:34-69.
17. Hastie T, Tibshirani R, Buja A. Flexible discriminant analysis by optimal scoring. J Am Stat Assoc 1994;89:1255-1270. [Web of Science]
18. Hannon WH, Aziz KJ, Collier FC, Fisher DA. Blood collection on filter paper for neonatal screening programs, 2nd ed 1992:1-23 National Committee for Clinical Laboratory Standards Villanova, PA. .
19. Levy HL, Simmons JR, MacCready RA. Stability of amino acids and galactose in the newborn screening filter paper specimen. J Pediatr 1985;107:757-760. [Web of Science][Medline] [Order article via Infotrieve]
20. Lundsjö A, Hagelberg S, Palmér K, Linblad BS. Amino acid profiles by HPLC after filter paper sampling: `appropriate technology' for monitoring of nutritional status. Clin Chim Acta 1990;191:201-210. [Web of Science][Medline] [Order article via Infotrieve]
21. Madira WM, Xavier F, Stern J, Wilcox A, Barron JL. Determination and assessment of the stability of phenylalanine and tyrosine in blood spots by HPLC [Letter]. Clin Chem 1992;38:2162-2163. [Web of Science][Medline] [Order article via Infotrieve]
22. Mabry CH, Reid MC, Kuhn RJ. A source of error in phenylketonuria screening. Am J Clin Pathol 1988;90:279-283. [Web of Science][Medline] [Order article via Infotrieve]
23. Hubbard R, Slocum R. Drug interference in amino acid analysis: a graphic presentation of pediatric antibiotics [Abstract]. Clin Chem 1994;40:1093.
24. DiCiccio TJ, Romano JP. A review of bootstrap confidence intervals. J R Stat Soc Ser B 1988;50:338-354.
25. Paulson AS, Delehanty TA. Sensitivity analysis in experimental design. Hiener KW Sacher RS Wilkinson JW eds. Computer science and statistics: proceedings of the 14th symposium on the interface 1982:52-57 Springer-Verlag New York. .
26. Shapira E, Blitzer MG, Miller JB, Africk DK. Biochemical genetics 1989:89-100 Oxford University Press A laboratory manual. New York. .
27. Chace DH, Millington DS, Hillman SL. Further progress in the automation of tandem mass spectrometry used to analyze amino acids and acylcarnitines from neonatal blood spots. Glass M eds. Proceedings of 10th neonatal screening symposium 1994:24-27 Washington State Public Labs. Seattle, WA. .
28. Levy HL, Waisbren SE, Lobbregt D, Allred E, Leviton A, Koch R, et al. Maternal non-phenylketonuric mild hyperphenylalaninemia. Eur J Pediatr 1996;155(Suppl 1):S20-S25.
29. Weglage J, Ullrich K, Pietsch M, Fünders B, Zass R, Koch HG. Untreated non-phenylketonuric hyperphenylalaninemia: intellectual and neurological outcome. Eur J Pediatr 1996;155(Suppl 1):S26-S28.
30. Costello PM, Beasley MG, Tillotson SL, Smith I. Intelligence in mild atypical phenylketonuria. Eur J Pediatr 1994;153:260-263. [Web of Science][Medline] [Order article via Infotrieve]
31. . Committee on Genetics, American Academy of Pediatrics. Newborn screening fact sheets. Pediatrics 1996;98:473-501. [Abstract/Free Full Text]
32. Berlin CM, Levy HL, Hanley WB. Delayed increase in blood phenylalanine concentration in phenylketonuric children initially classified as mild hyperphenylalaninemia. Screening 1995;4:35-39.
33. Okano Y, Eisensmith RC, Güttler F, Lichter-Konecki U, Konecki D, Trefz FK, et al. Molecular basis of phenotypic heterogeneity in phenylketonuria. N Engl J Med 1991;324:1232-1238. [Abstract]
34. Schoonheyt WE, Clarke JTR, Hanley WB, Johnson JM, Lehotay DC. Feto-maternal plasma phenylalanine concentration gradient from 19 weeks gestation to term. Clin Chim Acta 1994;225:165-169. [Web of Science][Medline] [Order article via Infotrieve]
35. Koch R, Levy H, Hanley W, Matalon R, Rouse B, Trefz F, de la Cruz F. Outcome implications of the International Maternal Phenylketonuria Collaborative Study (MPKUCS): 1994. Eur J Pediatr 1996;155(Suppl 1):S162-S164.
36. Koch R, Friedman EG, Azen C. Reproductive outcome of pregnant women with hyperphenylalaninemia. Metabol Currents 1996;8:1–4. (Ross Products Div., Abbott Labs., 625 Cleveland Ave., Columbus, OH 43215)..