Polyketides comprise a large and diverse group of natural products, many of which possess important biological and medicinal properties (1), yet polyketide biosynthesis proceeds by simple, repetitive condensations of acetate or propionate monomers in a manner that closely parallels fatty acid synthesis (2). Structural complexity is introduced by variation in the stereochemistry and the degree of reduction after each condensation as well as by downstream enzymes that catalyze cyclizations, oxidations, alkylations, glycosylations, and other transformations. Although these compounds are an attractive target for drug discovery (1), the complexity of many interesting polyketides impedes the preparation of analogs. Genetic methods for manipulating polyketide synthases (PKSs) show considerable promise for the engineered biosynthesis of novel polyketide molecules (3) but are currently limited in the range of compounds that may be accessed. The challenges involved in total synthesis of macrolides (4) make this approach impractical for the prepara

J. R. Jacobsen, Department of Chemical Engineering, Stanford University. Stanford, CA 94305-6025, USA C. R. Hutchinson, School of Pharmacy and Department of Bacteriology, University of Wisconsin, Madison, WI 53706, USA

D. E. Cane, Department of Chemistry, Box H, Brown University, Providence, RI 02912, USA.

C. Khosia, Departments of Chemical Engineering, Chemistry, and Biochemistry, Stanford University, Stanford, CA 94305-5025, USA.

*To whom correspondence should be addressed.

tion of derivatives. Synthetic modification of macrolides has led to the preparation of interesting compounds (5-7), but this method can also be extremely labor intensive and is limited in the range of transformations that can be selectively performed on these complex natural products. We report the development of a generally applicable, fermentation-based strategy in which chemically synthesized, cell-permeable, non-natural precursors are transformed into molecules resembling natural products by genetically engineered PKSs.

Deoxyerythronolide B synthase (DEBS) produces 6-deoxyerythronolide B (6-dEB) (1 in Fig. 1), the parent macrolactone of the broad-spectrum antibiotic erythromycin. DEBS consists of three large polypeptides (each >300 kD), each containing -10 distinct active sites. A one-to-one correspondence between active sites and chemical steps has been proposed (8, 9), leading to a model for the synthesis of 6-dEB in which each elongation step is handied by a separate enzyine "module" [see figure 1 of (10)]. The modular nature of DEBS and related PKSs (11) suggests potential strategies for genetic manipulation to generate novel natural products. Indeed, the feasibility of generating new polyketides has been demonstrated through the use of module deletion (12), loss-of-function mutagenesis within reductive domains (9, 13, 14), replacement of acyltransferase domains in order to alter starter or extender unit specificity (15), and gain-of-function

REPORTS

mutagenesis to introduce novel catalytic activities within modules (16). Importantly, many experiments show that downstream enzymes can process non-natural intermediates.

Biochemical analysis has also revealed that DEBS has considerable tolerance toward non-natural substrates. For example, primer units such as acetyl and butyryl coenzyme A (COA) (17), or N-acetylcysteamine (NAC) thioesters of their corresponding diketides (18), can be incorporated in vitro into the corresponding analogs of 6-dEB. However, in the course of these studies, it became clear that, even in the absence of externally added propionyl primers, a potential non-natural substrate must compete with propionate primers derived in situ by means of enzyme-catalyzed decarboxylation of methylmalonyl extender units (19). This competition puts severe limits on the priming of DEBS with unnatural thioesters because, for a poorly incorporated substrate, 6-dEB would be expected to be the dominant product.

We focused on incorporating substrates in vivo as cell-permeable NAC thioesters. Although exogenously fed NAC thioesters of advanced intermediates incorporate into several natural products derived from modular PKSs, the degree of specific incorporation was low (<3%) in all cases, presumably because of competing synthesis from metabolically derived intermediates (20-23). Mutational biosynthesis (24) offers the advantage of eliminating such competition. For example, a randomly generated mutant strain of the avermectin producer, in which biosynthesis of branched primer units is blocked, has been used to generate avermectin derivatives of commercial utility (21, 25). However, the unpredictability of random mutagenesis, coupled with the observation that incorporation efficiencies of natural and nonnatural substrates by such a mutant are low (26), precludes the general applicability of this strategy. In contrast, the specific introduction of null mutations, facilitated by the modular nature of DEBS, provides a general method for construction of useful blocked mutants. For example, inactivation of the ketosynthase KSI would be expected to abolish nominal biosynthesis, but polyketide production might still occur if an appropriate diketide (such as 2 in Fig. 1) was supplied as an NAC thioester. This has been demonstrated in the case of an engineered bimodular derivative of DEBS (27). To evaluate the utility of such a mutational strategy for practical precursor-directed biosynthesis of novel, structurally complex molecules, we introduced the same KS1 null mutation in the context of the fuil DEBS system.

(GIMMS) AVHRR NDVI data set at a similar spatial resolution, but at 15-day intervals, for the period January 1982 until the end of December 1990. The Pathfinder data were calibrated to correct for post-launch degradation from estimates of the relative annual degradation rates (in %) of the two channels: 3.6 and 4.3 (NOAA7), 5.9 and 3.5 (NOAA-9) and 1.2 and 2.0 (NOAA-11)". NOAA-9 data were used for inter-satellite normalization. The GIMMS data were independently calibrated"; they are considered here to illustrate how a different calibrating scheme affects derived trends in the AVHRR data.

For each equal-area pixel and at either 10- or 15-day intervals, depending on which of the two satellite data sets was used, the maximum of NDVI with minimal atmospheric effects was retained". The NDVI data from high northern latitudes (>40° N) did not show anomalies related to the El Chichon volcanic eruption during the mid-1982 to 1983 time period. These effects in the low latitude data were not corrected for in either of the two satellite data

sets.

The calibrated Pathfinder NDVI data still showed residual nonvegetation-related variations". We revised them by adjusting the NDVI for a hyper-arid portion of the Sahara desert

letters to nature

(1.42 × 10 km2) which has been found to be invariant as viewed by all three satellites. An alternate correction scheme based on desert pixels from 10° N to 50° N yielded nearly identical results. Importantly, when this desert correction was applied to the NDVI anomaly time series of desert pixels from five-degree latitude bands between 10°N and 50° N, the residuals resembled noise.

Time series of spatially averaged monthly NDVI, evaluated as the mcan of three 10-day maximum value NDVI composites, comprising a 10-year record, are plotted, first directly for reference (Fig. 1a), and then as anomalies to display interannual variability (Fig. 1b). Averaged for regions north of 45° N (uppermost curve in Fig. 1b) the NDVI anomaly shows evidence of increasing amplitude, summer values being low early in the record, high near the end. The NDVI anomaly in the tropics shows a large increase starting from November 1988, which also coincided with the change in satellites from NOAA-9 to NOAA-11. A somewhat smaller increase is seen during the switch from NOAA-7 to NOAA-9 in January 1985, although this increase began in the last months of the NOAA7 record (and the anomaly north of 45° N actually shows a decrease). This raises a question regarding anomalous variations in NDVI from sensor changes. Although efforts have been made to establish

Figure 1 Time variations in normalized difference vegetation index (NDVI) sumpared with changes in a petude of the seasonal cycie u atmcsoneric CO2 for the period from July 1981 to the end of June 1991. The data in a and b have been smoothed by a 3-month running mean. Zonal total NDVI and its anomaly were calculated from pixels having a 10-year monthly average NDVI greater than 0.1 and within 3 of the monthly average. The first condition guaranteed that bare or sparsely vegetated poets were not included in spatial averages, while the second condition removed most of the influence of snow and bed scan-lines. a, Monthly average NDM for selected latitudinal bands and the whole globe. The zonal total NDVI was normalized by the total vegetated land area in the month of August to obtain a zonal average that exhibited seasonality". b, Monthly total anomalies of the above, expressed as departures from the 10-year record averages of monthly NOVI, summed over each latitudinal band" for each month. The vertical scale of the global plot is twice that of individual latitudinal bands, c, Seasonal amplitude of NOVI averaged over selected latitudinal bands. The amplitude, defined as the July

and August average, is a good approximation because, at the northern latitudes Snow, the winarume NOM value is close to zero. Sestial avo-aging was fur fury and August data combined over pixels with 10-year averages of NDVI greater than 0.1, in order to exclude bare areas, such as the great deserts of Asia. Results from both the Pathfinder (left ordinate) and GIMMS (right ordinate) NDVI data sets are shown together with the corresponding rates of increase. The higher rates of increase inferred from GIMMS data may be due to the lack of desert correction for the version of GIMMS data used in this analysis. d. Seasonal amplitude of atmospheric CO2 relative to a base-period of 1961-67 as registered at Point Barrow, Alaska (71°N, 157°W) Linear trend estimates of the increase in seasonal amplitudes of NDVI and CO2 are statistically significant (10% level) for all latitudinal bands shown. However, the limitations of regression analysis on short samples, that is, the determination of trend in the presence of low-frequency variations, must be noted.

Seasonal NDVI amplitude (OIMMS data)

Figure 2 Interannual changes in seasonality of NDVI and in surface temperature, averaged north of 45° N, for selected pairs of years from 1982 to the end of 1900. a, NOVI averaged from 10-day maxima in NDVI for 1962-3, 1985-6, 1987-8 and 198990. Data from 1984 were not included because the number of years was odd. Spatial averaging was similar to that described in Fig. 1a legend. Changes in the timing of the active growing season over this 9-year record period were estimated from differences between the first and last bi-yearly average profiles at six threshold values of NDVI (from 0.1 to 0.35 in 0.05 increments). These values occurred during intervals of about 60 days each in spring and autumn when, respectively, NDVI was increasing and decreasing, at almost a constant rate. The six estimates each of the timing of rise and fall of NDVI may actually be correlated because of low-frequency variations (for example, soil moisture and/or equatorial sea surface temperature oscillations), and therefore, the standard errors given in the text must be interpreted in light of this limitation. We also inferred similar changes in the active growing season duration from an alternate poxel-by-posel and year-by-year analysis". b. Changes in the annual cycle of near-surface ai temperature from 1982 to 1990. Daily thermometer observations of aximum and minimum temperature were averaged in order to approximate daily mean temperatures and interpolated on a 1 x 1 degree grid. The daily data were further averaged over three separate approximately 10-day periods per month to obtain 36 observations per year. These 10-day average temperatures were then linearly regressed on the year (from 1982 to the end of 1990) to obtain the slopes shown tere

proper inter-sensor calibration linkages, some residual effects cannot be ruled out, especially between NOAA-9 and NOAA-11". This situation, for example, confounds proper interpretation of the tropical NDVI anomaly time series. For instance, intense sea surface temperature (SST) oscillatory events in the tropical Pacific and Atlantic oceans from 1982 to early 1989 have been linked to decreased vegetation growth in large regions of the semi-arid tropics". The increase in tropical and global NDVI anomaly starting from late 1988 also coincided with an unprecedented decline in atmospheric CO, anomaly, from a peak value in late 1988 to a minimum in late 1993. Nevertheless, these interpretations, as they involve the NDVI data, are limited by possible sensor change effects. Changes in the amplitude of the seasonal cycle of NDVI at northerly latitudes greater than 35°N are plotted in Fig. 1c, as characterized by changes in the July and August average NDVI. This broad measure of the seasonal maximum approximates the seasonal amplitude because winter-time NDVI at these northern latitudes is close to zero (compare Fig. 1a). The seasonal amplitude, by this definition, increased by 7 to 14%, depending on the latitude and data set, from 1981 or 1982 to the end of 1990 (Fig. Ic). Because NDVI is a measure of photosynthetic activity of vegetation as noted above, this increase indicates a substantial change in photosynthetic activity of plants at higher northern latitudes. A similar increase (14%) is indicated in the amplitude of the seasonal cycle of atmospheric CO; measured at Point Barrow, Alaska' (Fig. 1d). This CO2 cycle, although observed in the Arctic (71°N), registers changes in CO, gas exchanges, and hence in the biotic activity of plants and soil over all northern temperate and polar latitudes". Together, the NDVI and CO2 data indicate increased biospheric activity north of about 35° N. Two recent studies have also reported increased photosynthetic activity in the northern high latitudes as increased biomass from deposition in European forests" and from tree-ring analysis in Mongolia", respectively.

Timing of the seasonal rise and fall in NDVI suggests possible changes in the length of the active growing season, that is, the period during which photosynthesis actually occurs (as opposed to the concept of growing season, measured for example in degree days. As shown in Fig. 2a in Pathfinder data, the rise in NDVI, spatially averaged from 45° N to the northern limit of the data, came progressively earlier in the season between 1982 and 1990, as

shown by successive 10-day averages, where each plot shows an average over two years for clarity. Because spatially averaged NDVI rose each year at nearly a constant rate from early April (about day 110) to late June (about day 170) the advance in the active growing season is apparent, notwithstanding the relatively coarse time resolution (10 day) afforded by the NDVI data. From six estimates of the time advance at six successive thresholds of NDVI, we estimate an advance of 8 ± 3 days (Fig. 2a).

An advance of about 7 days in the seasonal cycle was previously inferred from atmospheric CO, data as having taken place between the 1960s and early 1990s, with most of the increase occurring after 1980 (Fig. 1 of ref. 1). The NDVI data suggest that this increase occurred over an extensive region of the extratropical Northern Hemisphere. The NDVi data in Fig. 2a further indicate a prolongation of the declining phase of the active growing season. estimated at 4 2 days between 1982-3 and 1989-90. Therefore, the active growing season north of 45° N appears to have lengthened by 12 ± 4 days over the 1980s. These estimates must be interpreted as suggestive of a longer active growing season, rather than in an absolute sense, in view of the coarse temporal resolution (10 days) and residual atmospheric effects in NDVI data. The associated standard errors given here are not rigorous, for low-frequency variations in NDVI data invalidate the assumption of statistical independence required of the successive threshold values.

Variations in the amplitude and timing of the seasonal cycle of atmospheric CO2 have shown an association with surface air ten.perature consistent with the hypothesis that warater temperatures have promoted increases in biospheric activity outside the tropics. A likely cause is an increase in the length of the active growing season brought about by warmer temperatures'. As shown in Fig. 2b, a pronounced increase in late-winter and early-spring temperatures took place over the period of NDVI changes, especially during March.

Because of their high spatial resolution (relative to ground-based meteorological measurements), NDVi data provide spatial detail of where the average changes in amplitude and timing of the active growing season occurred. To address regional variations in NDVI, we show in Fig. 3 a map related to the time plots shown in Fig. 1 together with a map of the 9-year average of NDVI for comparison. The linear rate of change in NDVI, averaged over the 9 years of