Assessment of relationships between grape chemical composition and grape allocation grade. Dr Paul Smith Research Manager - Chemistry

Transcription

1 Assessment of relationships between grape chemical composition and grape allocation grade. Dr Paul Smith Research Manager - Chemistry

2

3 Context Quality Measures Many individual compounds reported to have relationships with grades of grapes, wines, relationships with sensory. See ASVO Proceedings Colour historically used in Australia for grapes. Colour and tannin relationships with wine grade. Companies around the world using varying degrees of these for various applications; Gallo fruit streaming, index of multiple compounds. Germany released aroma from glycosyl glucose assay. Portugal - tannin, colour for fruit streaming, winemaking. US, NZ, Italy, Portugal, Sth Africa tannin, colour for winemaking decisions etc.

4 Project Questions Regardless of the grade, can chemical measurements be used to cluster fruit to identify similar patches or highlight differences? Assumes no knowledge of grade. Could be used to stream fruit. Can existing grading allocations be predicted using some of the previously identified (and some new) chemical measurements? Assumes the current grading system is accurate and seeks to objectify the process.

5 2013 Vintage Accolade as partner. Needed one standardised approach for grading of all fruit. Grades fruit and pays independent of wine outcome. Cabernet Sauvignon grapes only in samples across grades 2-7 (2 = higher value, 7 = lower value) 9 regions; Swan Valley Western Australia (SW) Riverland McLaren Vale Langhorne Creek Clare Valley Padthaway Coonawarra Wrattonbully

6 Spectroscopy

7 PCA: UV-Vis, marked with grade/region Marked with grade PCA analysis of spectra, nm 1 st two PC s accounted for 99 % of the variation between samples Clustering related to grade but some signs of non-linearity Marked with region Same data but marked with region Clustering related to region Riverland and Swan Hill very distinct from other regions

8 PCA using MIR spectra (homogenate) Marked with grade Marked with region Indistinct clustering with first 2 PC s (80% of the variation) (note that PLS performed well but required 6 factors and had problems with regions)

9 PCA using NIR spectra (homogenate) Marked with grade Marked with region Indistinct clustering with first 2 PC s (92% of the variation) (note that PLS performed reasonably well but required 7 factors and had nonlinearity problems with regions)

10 Take home message Principal Component Analysis (PCA) with UV-Vis spectra of grape homogenate extracts shows some separation by grade or region, but not distinct. NIR and MIR show very little separation with the first PC s so data structure is more complex.

11 Grade prediction: Discriminant analysis Perform discriminant analysis (DA) with spectral and chemical data using PCA scores. Chemical data must be standardised to remove scaling effects. Grading data is categorical so this type of analysis is best. Down-side is that it is hard to identify the drivers of the discrimination.

12 Discriminant analysis UV-Vis spectra of grape extracts: Use discriminant analysis - first 5 PCA factors of spectra, nm 40 out of 46 samples correctly predicted Incorrect predictions were all in adjacent classes MIR Spectra: 35 out of 41 samples correctly predicted 5 incorrect predictions were in adjacent classes, 1 was 2 classes away NIR spectra: 38 out of 41 samples correctly predicted All incorrect predictions were in adjacent classes

13 Grade prediction: PLS regression Treat grades as numerical values and predict using PLS regression Advantage is that significant variables will be identified Cross-validation can be used to test the model Again, chemical data must be standardised All data and figures shown are with cross-validation

14 PLS correlation: UV-Vis ( nm) of grape homogenate extract and grape grade R 2 = 0.51 SECV= 0.82 grade points (but consider that perfect prediction would be 0.5 as there is no fractional grading in reference data) Grade 7 is tight but in combination with all the data, Riverland grades 6 and 7 are not discriminated well WA grade 5 s are not predicted well compared with others in that class Grade 4 is relatively tight

15 PLS correlation between MIR spectra of grape homogenate and grape grade R 2 = 0.80 SECV= 0.49 grade points i.e. best possible error considering no incremental scores in reference data Stats look good on the face of it, non-linearity seen within regions (e.g. Riverland grades 6 and 7; McLaren Vale grades 4 and 5)

16 PLS correlation between NIR spectra of grape homogenate and grape grade R 2 = 0.42 SECV= 0.80 grade points Non-linearity, with grades 4 and 5 not predicted well

17 Take home message Overall, the results suggest that mid infrared (MIR) spectra of grape homogenate, with PLS regression modelling is a promising technique to assist in assessing/predicting grade. A relatively low cost and rapid analytical method.

18 Chemical analysis Berry basics: avg. Berry weight (g) ph TA7 (g/l) & TA8.2 (g/l) Brix Moisture (%) Malic acid (g/l) Alpha Amino Nitrogen (mg/l) Ammonia(mg/l) YAN (mg/l) Possible negative markers: Laccase activity (units/ml) Chloride (mg/kg) UV-Vis Spectral: Total phenolics A280 (AU) Colour A520 (AU) A420 (AU) MCP tannin (mg/l epicat. eq.) Aroma: C-6 compounds (µg/l) Methoxypyrazines (µg/l) GG in homogenate (µmol/kg) Free β-damascenone (µg/l)

19 PCA with chemical analysis, marked with grade/region PCA analysis of standardised chemical data First two PC s accounted for 47 % of the variation between samples With only the first 2 PC s clustering related to grade is less distinct than with UV-Vis spectral data (more PC s needed to discriminate grades) Regional clustering stronger than grade Compared with UV-Vis spectral data, stronger overlap of other regions with Riverland

20 PCA with chemical analysis: scores and loadings biplot 0.5 Wrat Clare Chloride Swan 0.4 MVale Clare Wrat 0.3 MVale WA b-damascenone Methionine Tyrosine Threonine Asparagine Glycine Proline 0.2 Serine Swan MVale Padtha Histidine Z-3-hexeno Brix Ammonia Coona RivL MVale 0.1 Lysine hexano WA Padtha MVale Aspartate RivL MVale MVale 0 LangHC Coona Glutamine LangHC RivL WA -0.1 alpha2amin LangHC RivL RivL Mesityl Ox -0.2 Arginine WA RivL Cysteine RivL -0.3 Alanine WA RivL WA RivL RivL WA WA -0.4 E-2-hexena RivL -0.5 Moisture -0.6 Glutathion Glutamate RivL Berry wt PC-2 (20%) RivL Clare Amino acids Isoleucine Leucine Phenylalan Valine Laccase Tryptophan ph YAN TA7 TA8.2 Bi-plot GG Clare Alpha AAE-2-hexenol MVale Malic Phenolics Total phen MCP tannin A420 A520 RivL WA Like samples are near each other Correlating analytes are near each other or opposite if negatively correlated Distance from 0 indicates strength of loading PC1 driven mainly by phenolics and colour in positive direction, aa s in negative direction Phenolics and colour loadings are stronger for PC2 aa s are split over PC2 Riverland and WA tend to be separated from others by PC2 and are lower in phenolics/colour and some aa s but higher in others PC-1 (27%) Acids Amino acids

21 Discriminant analysis with chemical data Discrimination Linear Discriminant Analysis Accuracy: 84.78% Categories: 4 Method Used: Quadratic Projected: 5 Components 4 predicted as 6 7 predicted as 6 39 out of 46 samples correctly predicted predicted as 6 Almost half of incorrect predictions were 2 classes away

22 PLS to predict grade with chemical analysis (standardised) R 2 = 0.49 SECV= 0.79 grade points Riverland grade 6 is predicted better than with spectral data Grade 4 has more spread than with spectral data and appears to be regional. Uncertainty test was used to identify the significant analytes (see next slide)

23 PLS predicting grade with chemical loadings Lower amounts in better grades Higher amounts in better grades In this case only 1 factor used for prediction so loadings for that factor can be examined directly Only significant variables shown Coefficients can be positive or negative

24 Take home message Moderate ability to predict grade using chemical data and PLS regression. Suggests some lead targets to keep an eye on for 2014 sample set; YAN, Total Phenolics, A420 and A520 and β-damascenone are positively associated (i.e. higher amounts) in higher value grades TA, cysteine, glutamate and glutathione are negatively associated (i.e. lower amounts) in higher value grades.

25 Commercial wine allocation grading CAS SHZ Tannin (epicat. eq. g/l) Tannin (epicat. eq. g/l) Higher Lower Higher Lower Quality Grading Quality Grading Analysis of Variance F Ratio = 12.1 p Value = <.0001 N = 349 Analysis of Variance F Ratio = 42.2 p Value = <.0001 N = 634 One-Way ANOVA - plot shows mean and mean error bars and Std dev ( )

26 New metrics: Wine extractable tannin Whole grape extracts, crushed berries in acidified 15% ethanol Harsh grape extractions (50% ethanol) following homogenisation 1 kg fermentations HYPOTHESIS: Extract dilute ethanol Extract with high solvent:water

27 Wine-extractable grape tannin A strong correlation was found with wine-like extraction and wine tannin Small differences were found between cultivars Wine tannin (g/l) y = x R² = Wine tannin (g/l) Shiraz Cabernet S. y = x R² = y = x R² = Wine like Grape Tannin (g/kg F. wt) Wine like Grape Tannin (g/kg F. wt)

28 Regionality affects wine-extractable grape tannin Large differences were found between regions Riverland = lowest tannin; McLaren Vale = highest tannin Highest tannin grapes/wines were also the ripest samples y = x R² = Wine tannin (g/l) Shiraz Cabernet S. 1 Langhorne Creek Riverland McLaren Vale 0.5 Clare Padthaway Linear (Shiraz) Wine like tannin extract (mg/g)

29 Homogenate extracts still work OK... Wine tannin (g/l) Shiraz Cabernet S. Linear (Shiraz) Linear (Cabernet S.) y = 0.386x R² = y = x R² = Homogenate Grape Tannin (g/kg F. wt) Correlation grape wine for homogenate extracts was not as clear-cut as for wine-like extracts BUT when the data was separated for cultivar, homogenate tannin strongly correlated with wine tannin

30 Take home messages Principal Component Analysis (PCA) with UV-Vis spectra of grape homogenate extracts shows some separation by grade or region, but not distinct. NIR and MIR show very little separation with the first PC s so data structure is more complex. Overall, the results suggest that mid infrared (MIR) spectra of grape homogenate, with PLS regression modelling is a promising technique to assist in assessing/predicting grade. Using chemical data and PLS regression gives moderate ability to predict grade. Suggests some lead targets to keep an eye on for 2014 sample set; YAN, Total Phenolics, A420 and A520 and β-damascenone are positively associated (i.e. higher amounts) in higher value grades TA, cysteine, glutamate and glutathione are negatively associated (i.e. lower amounts) in higher value grades. A new metric is available wine extractable tannin to complement total grape tannin.

31 Acknowledgements: Keren Bindon Jacqui McRae Alex Schulkin Bob Dambergs (Chemometrics) Sheridan Barter Mark Solomon Ruchi Shah Leigh Francis Stella Kassara Wies Cynkar Ella Robinson Neil Scrimgeour Peter Godden Eric Wilkes Accolade Wines (Chris Bevin, Warren Birchmore, Alex Sas)

32 Acknowledgements, a member of the Wine Innovation Cluster in Adelaide, is supported by Australian grapegrowers and winemakers through their investment body, the Australian Grape and Wine Authority with matching funds from the Australian Government.

33 EXTRA S Alanine (ug/l) alpha2amino (ug/l) Arginine (ug/l) Asparagine (ug/l) Aspartate (ug/l)cysteine (ug/l) Glutamate (ug/l) Glutamine (ug/l) Glutathione (ug/l) Glycine (ug/l) Histidine (ug/l) Isoleucine (ug/l) Leucine (ug/l) Lysine (ug/l) Methionine (ug/l) Phenylalanine (ug/l) Proline (ug/l) Serine (ug/l) Threonine (ug/l) Tryptophan (ug/l) Tyrosine (ug/l) Valine (ug/l)