We analysed 33 formalin fixed paraffin embedded (FFPE) specimens (Table 1) from the following sites: lung (4), tongue (3), vocal cord (2), bladder (4), soft tissue (5), gallbladder (1), seminal vesicles (1), stomach (1), pleura (1), skin (8), conjunctiva (2), colon (1). Respective controls were also included. There were 18 cases of AL (in 4 of 18 subtyping permitted exclusion of AA type deposit but could not demonstrate AL which is reported to be the case in ~20% of AL deposits, these cases were called non-AA subtype), 6 cases of transthyretin (TTR), 1 β-2 microglobulin, 1 case of Insulin (Ins) deposition, 3 cases where subtyping was negative on all proteins tested and positive only for Congo red stain, 1 case where even Congo red on repeated sections turned out to be negative, and 3 cases where subtyping was not possible for paucity of material. Figure 1 shows the staining in light micrographs of some of these specimens, indicating the patchy deposition of aggregated proteins.
The defining feature of AL amyloid deposits is their characteristic intermolecular β-sheet structure11. Most tissues contain a large portion of α-helical proteins, so the build-up of amyloid deposits with their characteristic β-sheet absorption should change the shapes of the amide bands in the infrared spectra from the affected regions compared to those from healthy tissue. Amide A at ~ 3300 cm-1, Amide I at ~ 1650 cm-1, Amide II at ~ 1550 cm-1 and Amide III at around 1300-1200 cm-1 are all sensitive to the protein secondary structures12,13.
Figure 2 compares the ratio of Amide I band intensities at 1630 cm-1 and 1655 cm-1 in the IR spectra of each specimen. There were some differences between controls from different tissues: the ratio was 0.71 in CtrlG3, 0.72 in CtrlR3, and 0.76 in CtrlST4. The spectra from these sections showed a relatively small variation of the ratio, and the amide I bands had peak positions around 1650 -1655 cm-1 characteristic for α-helical protein structures12. However, two control specimens – CtrlR1 and CtrlST3 - had ratios around 1, due to more β-sheet content, and the spectra of these specimens also had a considerably larger variability, indicating that these samples were much less homogeneous. The average ratio in the remaining controls was typically around 0.8.
In the amyloid case samples, the AL lambda subtype tended to have a very strong effect in most tissues except in the GI tract specimens, with a ratio above 1, and a shift of the amide I band maximum down to 1634-1638 cm-1. AL kappa on the other hand only had such strong shifts in the skin specimens. In the other tissues the change was much less pronounced or not visible at all. The Ins ST/fat case specimen with amyloidosis due to insulin deposition had a value of 1. TTR showed a strong shift in the soft tissue specimen only, but not in the GI tract or Urinary system specimens. NonAA subtype showed a pronounced signal from amyloid structures in the respiratory tract specimens, but not in the soft tissue or the GI tract. Generally, the GI tract seemed to be less affected than the other tissues, and the respiratory system the most.
Thus, in our data set the band shapes varied widely both with amyloid type/subtype and tissue. While many specimens showed the expected increase in the Amide band I component at 1630 cm-1 and the associated 1531 cm-1 component in the Amide II band 11, in many other case specimens the IR spectra looked almost identical to those of the controls. Obviously, there was no consistent correlation between the amounts of β-sheet structures and the disease state.
Tissue-type specific amyloidosis models
To compare the impact of amyloidosis in different organs, the specimens were grouped into five tissue systems: Respiratory tract (resp), containing lung, pleura and vocal cord tissue, GI tract (GI), containing tongue, stomach, gallbladder and colonic mucosa, Soft tissue (ST), containing conjunctival soft tissue, abdominal fat, retinaculorum and synovium, Skin (skin), and Urinary system (Uro), containing bladder and seminal vesicles.
Comparison of the original IR spectra (Supplementary figure SF1 and supplementary table ST1) hinted at the variation within specimens in the tissues. Principal Component Analysis (PCA) (Supplementary figure SF2) mostly did not separate cases from controls and did not consistently distinguish between amyloid subtypes. PCA picks out variance in the spectra, which can be due to any number of factors. The PCA loadings indicated a strong contribution from the amide bands, but the spectra did contain other variability, in the baseline, the paraffin bands, and also different noise levels. Section thickness, treatment, sample mounting, can have an effect as well as different protein structures. In contrast, Partial Least Squares (PLS) regression extracts variance that is correlated with observations / sample values. In this particular model setup, the controls had been assigned the value 0 and cases the value 1, and this model was not designed to distinguish between subtypes. Using PLS in this way has the advantage of giving a numeric value related to similarity rather than simply group membership. The PLS model loadings can also give an idea which spectral features were correlated with the transition from healthy to the disease state. The models were made with the first derivative of the two spectra regions 3800-3100 cm-1 and 1900-1100 cm-1. Consequently, the model loadings and coefficients contained sinusoidal first-derivative features. To aid with the interpretation, the loadings were integrated to re-gain recognisable IR bands in the figures.
The PLS predictions for all specimens with the different models are listed in Table 2. Figure 3 shows the integrated first three PLS loadings for the control vs disease PLS models in the five tissue systems.
Gastrointestinal tract: The GI tract cases vs GI controls PLS model performed very well. The GI calibration sample fits were close to the nominal values - 0 for controls, 1 for cases. In GI tract, we have a specimen from colonic mucosa, gallbladder, tongue and so called Gastrointestinal stromal tumour (GIST) of the stomach. The latter shows amyloidosis only in the wall of blood vessels, so the shifts found here may not be organ specific. This was also the case in colonic mucosa where the deposits were exclusively vascular.
When the model was applied to the other tissues, all controls were assigned low values, and cases were assigned high ones. The only exception was U3 (Urinary/Bladder ALSKappa), which was assigned a similarly low value, and also had the same amide I band ratio, as a control. All other case predictions were clearly higher than any control, but there was a group of specimens in the test cases with values hovering around 0.5. However, at this stage the values could not be attributed unambiguously to the subtypes or locations of the amyloid deposits because the samples were mostly from single specimens and differences between the case specimens could be due to sample variation.
The first loading of the GI PLS model relating to the largest correlated compound, showed positive bands at 3268 cm-1 in the amide A region, 1649 (shoulder) and 1610 cm-1 in the amide I, at 1516 cm-1 in the amide II and at 1222 cm-1 in the amide III region. It also had negative bands at 3336 cm-1, 1669 cm-1, 1554 cm-1, and 1245 cm-1. The changes in the Amide I and II band regions are very similar to the pattern identified by Ami et al. 11. Therefore, this loading would represent loss of native α-helix, random, and turn structure and increase in the amyloid β-sheet. Other changes in the tissue are indicated by the higher loadings. A feature to note is a change in the Amide A band region around 3300 cm-1. All five tissue models had a positive band at 3280 cm-1 associated with the amyloid β-structure, and there was a negative band around 3330 cm-1 associated with native (helical etc.) structure. Overall, there was a strong indication that the Amide A band experienced a characteristic shift towards lower wavenumbers, whose extent varied between the tissue systems. The Amide A band is generally not thought to be influenced by the backbone structure, only by hydrogen bonding strength. However, Krimm and Bandekar reported shifts as large as 40 cm-1 between different sheet structures13. Moreover, the bands observed in the PLS models are well-defined and quite distinct from the wide baseline variations. Therefore, it would be reasonable to consider them as specific markers in addition to the patterns seen in the Amide I and II band region.
Respiratory system: The Resp cases vs Resp controls model performed almost as good. Controls and case specimens were clearly distinguished. The three Resp control specimens, which originated in different organs - lung, pleurum, and vocal cord - were close, but had slightly different average values. The Resp calibration specimens had all average values close to 1 as expected. The spectra from R4 had a remarkably wider spread of values than those from the other lung samples. This could be patient specific, but without having more samples to confirm this, variations in the averages and in the spread of the values were most likely due to the sampling and preparation of individual sections. When this Resp model was applied to the other tissues, the average assigned values for the ‘unknown’ controls was somewhat higher than the Resp controls, and one Uro/SemVes control came out as suspect with a value of 0.77. The case sample values were mostly between 0.5 and 1, lower than the Resp calibration samples, but still higher than the controls. However, four specimens (G4, ST2, ST3, and U4), had low values that could have them classed as false negatives if a case diagnosis were made based on the output of this PLS model.
The first loading of the Resp PLS model showed strong positive peaks at 3280, 1615, 1540-1510, and 1220 cm-1, indicating an increase in amyloid β-sheet structure. Only the third loading showed clear negative bands at 3315, 1648, 1540 and 1240 cm-1 relating to the loss of native helical structures. This strong emphasis on the amyloid deposition could explain why specimens with low amide I ratio were generally fitted with lower PLS values.
Skin: The Skin cases vs Skin control model was probably skewed by the fact that there was only a single control specimen in the calibration. This PLS model had a big gap between the skin control and the skin case values. Perhaps that was not surprising, because skin specimens had some of the most extreme shifts in the amide bands. When applied to the other tissues, this model gave rather high values for all controls other than skin. However, the case sample values were also higher throughout. There were only three case samples (LN1, R4 and U4) with values similar to the controls. On the whole, it would seem that this model separated reasonably well between controls and cases, albeit with a shifted boundary.
The Skin PLS model basis appeared quite different from the others – the first loading showed negative peaks at 3340, 1660, 1563, 1460, 1238 and 1205 cm-1, and positive bands only at 3275 and 1620 cm-1 (small). The second loading had positive bands at 3335 cm-1, negative at 1652, 1610 and 1545 cm-1. The third loading had negative bands at 3289, 1635, 1564 cm-1, but positive at 1230 cm-1. The skin specimens were mostly ill defined (cr, neg) and had lower amide I ratios. These band changes could therefore relate to tissue damage in those cases, rather than a defined deposition of amyloid.
Soft tissue: The four calibration controls from conjunctival soft tissue, connective tissue, abdominal fat and synovium varied slightly, but again, those were single specimens, so it could not be attributed to organ origin. The ST2 specimen (conjunctival soft tissue ALLkappa) had a markedly lower average than the other cases and overlapped with the controls. All other calibration specimens were well separated from the controls. However, while the calibration controls and samples were assigned well, the distinction between controls and cases was less clear when this model was applied to the other tissues. The averages for controls were around 0.5 and the cases around 0.8, with considerable overlaps when taking into consideration the distribution of values within each specimen.
The first loading of the ST PLS model showed similarly strong positive bands at 3285, 1620, 1515, and 1225 cm-1 related to deposition of amyloid as the Resp model. The second loading had negative bands at 3315, 1658(sh), 1615, 1520, and 1245 cm-1. This model seemed to detect changes in inter and intramolecular β-sheet structures but was less accurate with differentiating between controls and cases.
Urinary tract: The two control specimens behaved quite differently, and the PLS model indicated that the Urinary/bladder specimen (CtrlU1) spectra showed a particularly wide spread of values. All three Uro TTR specimens were tightly grouped around similar average values, and there was no suggestion of one behaving differently. ALLambda and ALSkappa Uro specimens had more variety. Again, the PLS model did clearly separate all calibration cases from the calibration controls, but it did not work as well when applied to other tissues. While case samples did show broadly higher values than controls, there was a large overlap; many case specimens would count as false negatives, and even some controls as false positives.
The first loading of the Uro PLS model had a positive peak at 3267 cm-1, and a sharp peak at 1627 cm-1 and broad 1516 cm-1 and weak 1231 cm-1 peaks. The second and third loadings also showed sharp peaks in the Amide I/II band regions. Like the soft tissue model, the Urinary tract model assigned the calibration controls and samples well but made a poor distinction between unknown controls and cases.
Taken together, these results showed a strong dependence of the model performance on the calibration samples. While the first two tissue-based PLS models correctly assigned unknown samples from other tissues, the latter three models had problems. The case values were generally higher than controls, but the gap between controls and cases was smaller and the variability much higher.
Type-specific amyloidosis models
PLS models for AL (both lambda and kappa), nonAA, TTR and Ins cases versus all controls were made to find characteristics of these amyloidosis types across the tissues. The model predictions are listed in Table 2. Figure 4 shows the integrated first three PLS loadings for the type vs control PLS models.
AL: The AL calibration samples had fitted PLS between 0.57 and 1.4. However, the other case samples were also assigned relatively high values, higher than the controls (-0.13 to 0.24) and spanning a similar range than the AL samples.
The first loading in AL / ctrl model had strong positive bands at 3285, 1618, 1520 and 1227 cm-1 and indicated the deposition of β-sheet structures. The second loading had neg/pos feature at 3334/3270 cm-1, broad negative bands at 1655-1610 cm-1 and 1530-10 cm-1. The third loading showed negative bands at 3310, 1655, 1610, 1545 and 1242 cm-1. Both these related to the loss of native, predominantly helical structures, and seemed to be more important with the control/case distinction than the variable amounts of β-sheet. Thus, this model did clearly separate amyloid cases and controls, but could not distinguish AL from other types.
NonAA: The nonAA vs controls PLS model performed similar to AL. The controls had values around 0, but all case specimens except U4 had high values, so there was no separation at all between the nonAA and other types. The loadings were similar to the ones form the AL model:
The first loading Non-AA / ctrl model showed strong positive bands at 3273 (neg 3338), 1631, 1515 and 1222 cm-1. The second loading had positive amide A region band at 3285 cm-1, broad negative amide bands at 1620, 1535-1510, and 1236 cm-1. The third loading was indistinct in the Amide A region, but had positive bands at 1670, 1627 and 1555 cm-1. The positions were slightly different as AL, but the pattern was the same: the first loading related to β-sheet structures, whereas the other two indicated the loss of native structure.
TTR:
The TTR vs controls model assigned a few more case specimens (LN1, R4, S7, U4) similarly low values as the controls, but most test specimens fell in same range as the TTR specimens, so there was again no clear separation of TTR from other types.
The first loading of the TTR / ctrl model had negative peaks at 3324, 1628 (1660 sh), 1545 and 1240 cm-1. The second loading showed positive peaks at 3270, 1618, 1509 and 1233 cm-1. The third loading had positive bands at 3345, and 1700-1670 cm-1, negative at 3266, 1615, 1510 and 1223 cm-1. This inversion reflected the observation that the amide I ratios of the TTR specimens were as low or even lower than the controls. Nevertheless, the model did not separate TTR from other types with higher amide I ratios, which again indicated that the amount of amyloid deposit was irrelevant for this kind of model.
Insulin:
The Insulin vs controls model was the only PLS model which did separate the Ins type from the other amyloid types. Two case specimens (S8 and U4) were close to 0, in the same range as the controls, but the majority ranged from 0.17 to 0.55, markedly lower than the 0.87 of the Ins. The problem with this model was that there was only this single Ins specimen for calibration, and therefore the model cannot be considered reliable. The bands in the Ins / ctrl model were sharper than in the previous models, perhaps indicating a better defined deposit: The first loading showed strong positive peaks at 3275 cm-1, 1630 cm-1, 1511 cm-1 and 1221 cm-1. Loading 2 had positive peak at 3329 but negative peaks at 1622, 1510 and 1226 cm-1. The third loading was pos / neg at 3350 / 3286 cm-1. It had a positive band at 1660 cm-1.
Direct comparison of types in amyloidosis cases
The previously described AL, nonAA and TTR models with controls as reference all tended to assign unknown specimens as cases. The model loadings exhibited similar spectral features than the tissue type models. Therefore, these models were influenced by common changes from healthy to disease state, rather than by amyloid type. To establish more specific discriminators, PLS models were made for the AL (both lambda and kappa), nonAA, TTR and Ins types against all characterised cases. The model predictions are listed in Table 2. Figure 5 shows the integrated first three PLS loadings for the type versus all cases PLS models.
AL cases: The fitted values from this model clearly distinguished between AL type and the other type cases. The fitted values of NonAA, TTR and Ins specimens were lower. The control samples, which had not been included in the model calibration, returned also low values, with the exception of CtrlU2. When this AL cases model was applied to the poorly characterised specimens (cr, nd etc.), three (GI2, R5, S2) fell in the same range as the AL cases, the other four were much lower. The first two loadings in the AL / cases model indicated changes in the inter/intramolecular β-sheet structures. Loading 1 had positive peaks at 3290, 1616, 1511, and 1222 cm-1. Loading 2 was positive at 3278 cm-1, negative at 1649, 1504, 1236 cm-1. Only loading 3 had negative peaks at 3321, 1656, 1616, 1560, 1238 cm-1 which related to α-helical structures.
Non-AA cases:
The nonAA model performed erratically. One of the nonAA calibration samples (R7) was fitted a low value consistent with the other amyloid types. From the unassigned test cases, S3 and S5 and some of the controls had similar values as the non-AA specimens. The erratic behaviour may be due to the fact that different proteins may fall into this very vague nonAA category, we know very little about nonAA and what really makes this group of exclusion. In the nonAA / cases model loading 1 showed negative peaks at 3294, 1619, 1515, 1227 cm-1 due to changes in β-sheet. nonAA specimens had mostly lower amide I ratios than the AL. Loading 2 had positive bands at 3276 and 1627 (sharp) cm-1. Loading 3 had positive peaks at 3291, 1617, and 1545-1515 cm-1.
TTR cases:
The TTR model also separated TTR cases from the other type cases, with consistently higher PLS values than the other specimens. It fitted all test cases except S6 similarly low values. However, the control sample values ranged rather widely, some had equally high values as the TTR specimens.
Loading 1 in the TTR / cases model had negative peaks at 3287, 1622, 1517 and 1226 cm-1. As with nonAA, TTR had mostly lower amide I ratios than AL. Loading 2 was negative at 3278 cm-1, and positive at 1649, 1605, 1510 and 1236 cm-1. Loading 3 had positive bands at 3317, 1650, 1608, 1540, 1515, and 1227 cm-1.
Ins type:
In the Ins model only the Ins specimen had a large value (0.43), all other case and control sample values were around 0. Loading 1 of the Ins / cases model had strong negative peaks at 3295, 1616, 1515, and 1220 cm-1, showing comparatively less β-sheet than the other case specimens in the calibration set. Loading 2 had positive peaks at 3283, 1626, 1522, and 1230 cm-1. Loading 3 had strong negative peaks at 3285, 1621, 1541 and 1249 cm-1. Both were quite distinct. However, as mentioned before, there was only a single Ins specimen for calibration, and therefore the model cannot be considered reliable.
Poorly characterised samples:
The sample set contained six samples (R5, S2, S3, S4, S5, S6) whose subtype was not defined. Only R5 and S4 showed elevated amide I band ratios, but all six specimens were denoted as amyloid cases by the GI tract case model. R5 and S2 scored high in the AL type model, S3 and S5 scored high in the nonAA type model, S4 and in particular S6 scored high with the TTR type model. None showed any similarity with Ins.