Artificial intelligence in neurodegenerative disease research: use of IBM Watson to identify additional RNA-binding proteins altered in amyotrophic lateral sclerosis

ALS and non-neurologic disease control post-mortem tissue samples were obtained from the University of Pittsburgh ALS Tissue Bank, the Barrow Neurological Institute ALS Tissue Bank, and the Target ALS Human Postmortem Tissue Core. All tissues samples were collected after informed consent from the subjects or by the subjects’ next of kin, complying with all relevant ethical regulations. The protocol and consent process were approved by the University of Pittsburgh Institutional Review Board (IRB) and the Dignity Health Institutional Review Board. Clinical diagnoses were made by board certified neuropathologists according to consensus criteria for ALS. Subject demographics are listed in Suppl. Table 5.

Paraffin-embedded post-mortem tissue sections from spinal cords and cerebellum were used for this study. All sections were deparaffinized, rehydrated and antigen retrieval performed using Target Antigen Retrieval Solution, pH 9.0 (DAKO) or a citrate buffer (pH 6) for 20 min in a steamer. After cooling to room temperature, non-specific binding sites were blocked using Super Block (Scytek), supplemented with Avidin (Vector Labs). Primary antibodies used for immunohistochemistry were incubated overnight in Super Block with biotin (antibodies listed in Suppl. Table 3). Slides were then washed and incubated for 1 h in the appropriate biotinylated IgG secondary antibodies (1:200; Vector Labs) in Super Block. Slides were washed in PBS and immunostaining visualized using the Vectastain Elite ABC reagent (Vector Labs) and Vector NovaRED peroxidase substrate kit (Vector Labs). Slides were counterstained with hematoxylin (Sigma Aldrich). Sections were visualized using a Leica AperioScope microscope, and analyzed using the Aperio eSlide manager image analysis.

For color intensity analysis, regions of interest (ROI) were delineated by a blinded user (motor neuron nuclei or Purkinje nuclei), slides were deconvolved for RGB of hematoxylin (blue channel) and antibody staining color (red channel) using the Leica Aperio ImageScope color deconvolution algorithm and the intensity value measured for each pixel within the ROI. These values were used to set intensity scales for each color from 0 to 255 (0 = black and 255 = white) prior to the analysis, and the same intensity thresholds were used across each antibody analysis. For hnRNPU, the negative threshold was set to be for intensities ranging from 210 to 255; weak positive staining intensity ranges were from 145 to 210; medium positive staining from 90 to 145 and strong staining was set to ranges from 0 to 90. All neurons were selected for each spinal cord section (numbers of neurons per section ranged from 20 to 50), and ROIs were defined. For Syncrip, the negative threshold was set to be for intensities ranging from 180 to 255; weak positive staining intensity ranges were from 155 to 180; medium positive staining from 95 to 155 and strong staining was set to ranges from 0 to 95. We selected 50 Purkinje cells from different areas of each section.

Lumbar spinal cord and cerebellum total RNA were prepared from frozen tissue from control and ALS cases. Samples were homogenized in Trizol (Invitrogen), and RNA was extracted using the Ambion PureLink™ RNA Mini Kit. RNA quality was determined by RIN (RNA integrity number) using a Tapestation and all samples showed RIN values of > 5. cDNA was synthesized using Superscript VILO (Invitrogen) and real-time RT-PCR was performed using the FastStart Universal SyberGreen master mix (Roche). Primer sequences used are listed in Suppl. Table 7.

For laser-capture microscopy, fresh-frozen cerebellum were sectioned at 20 μm, slides were fixed for 2 min in 70% ethanol (in nuclease-free water), washed and stained for 6 min with the RNA/DNA stain Methyl Green Pyronin (Abcam, ab150676) supplemented with SUPERase In RNAse inhibitor (AM2694, ThermoFisher). Slides were consecutively dipped in nuclease-free water, 100% ethanol, and air-dried for 2 min before capture. We used the Zeiss Axiovert Zoom, fitted with a PALM system to capture at least 120 Purkinje cells per slide. Capture time was limited to 1 h to minimize RNA degradation. Two slides from each sample were used for a total of 250 neurons, and the cells were combined for subsequent processing. RNA was extracted using the RNAqueous micro total RNA isolation kit from Ambion (AM1931), cDNA was synthesized using Superscript VILO and real-time RT-PCR was performed.

All data generated or analyzed during this study are included in the published article and its supplementary information files (Suppl. Tables 2, 3 and 4). The pseudo-code used to generate our analysis by IBM Watson is included in the supplementary information files.

A detailed description of the analytical methods used by Watson to predict RBPs associated with ALS is provided in the Supplemental Materials and Methods section. Briefly, IBM Watson extracts domain-specific text features from published literature to identify new connections between entities of interest, such as genes, proteins, drugs, and diseases. From these annotated documents, Watson creates a semantic model of the known set of RBPs previously linked to ALS, and then applies that model to a candidate set of all other RBPs, in order to rank all the candidates by similarity to the known set using a graph diffusion algorithm [58]. To test the model generation by Watson, we first performed a leave-one-out (LOO) cross-validation to demonstrate the predictive power of Watson using the 11 known ALS-linked RBPs. To do so, the graph diffusion algorithm was applied 11 times based on the same distance matrix, but each time a different known RBP was taken out of the positive set and placed into the candidate set. If the overall model is accurate, then the positive RBPs placed into the candidate set should rank high based on the model built from the other ten known positive RBPs. Indeed, in our experiment, the LOO cross-validation results were strong, with 5 of the positive RBPs ranking in the top-15 out of 1478 RBPs, and 8 in the top 4.1% of all RBPs (p = 3.17 × 10⁻⁷, one-sided Wilcoxon rank sum test to assess whether the scores of the knowns are greater than those of the candidates, Table 1).

To measure accuracy of the model based on the LOO cross-validation, we used the receiver operating characteristic (ROC) curve. The area under the ROC curve (AUC, computed following the trapezoid rule) is a measure of model accuracy, where a value of 0.5 corresponds to a random model and a value of 1 corresponds to a perfectly predictive model. The AUC for our model was 0.935.

An added value of the LOO cross-validation is that it provides a point of reference for where to expect true ALS-related RBPs from the candidate set to be listed within the overall ranking. In the cross-validation performed by Watson, 10 of the 11 positive RBPs (90%) ranked within the top 8% of all RBPs. ANG was ranked number 713 by this analysis, suggesting that this gene is dissimilar to known ALS-associated RBPs. Extrapolating these results to all known and yet-to-be discovered RBPs, we can expect that approximately 90% of all true positive ALS-linked RBPs should fall within the top 8% of the ranked list.

We next used a retrospective study to test IBM Watson’s analytics for predicting RBPs in ALS. We restricted the corpus of data analyzed by Watson to literature published prior to 2013, and used as a positive known set the eight RBPs with ALS disease causing mutations that were published through 2012 (Table 2). We used all 1542 RBPs in the genome listed in Gertsberger et al. [21], as candidate RBPs. 1439 of these RBPs had at least one mention in Medline^® abstracts up to the end of 2012 and these were thus chosen to be our candidate set. IBM Watson built a distance matrix relating each RBP with all others and used a graph diffusion algorithm to rank all RBPs based on the known set of eight RBPs [58]. The RBPs found to be mutated in familial ALS and published between 2013 and 2017 were ARHGEF28, Matrin 3, GLE1, and TIA1 [16, 27, 28, 36]. We asked if Watson could predict these RBPs as high-ranking candidates. The results from this retrospective analysis identified Matrin 3 as the top candidate, ARHGEF28 and TIA1 ranked within the top 5%, and GLE1 ranked within the top 11% of all known RBPs (Table 2), thus demonstrating the performance capabilities of the model. Another RBP linked to ALS in 2014 with no known mutations, hnRNPA3, was also ranked within the top ten in this retrospective study.

After having established that IBM Watson methodology is valid and capable of identifying RBPs involved in ALS, we next analyzed all 1478 RBPs from the Gertsberger et al. [21] list that had at least one mention in Medline^® abstracts up to the end of 2015 as our candidate set. The known set included all 11 known RBPs mutated in ALS identified prior to 2016 (TDP-43, FUS, Ataxin-2, hnRNPA1, hnRNPA2B1, Senataxin, Angiogenin, TAF15, GLE1, Matrin-3 and ARHGEF28). We excluded from the known set RBPs such as RBM45, hnRNPA3, or MTHFSD that have been shown to be altered in ALS tissue samples, but without any mutations described to date [6, 10, 37, 40]. IBM Watson rank ordered RBPs by similarity to the known set, and assigned a score from 0 to 1, corresponding to how closely related each RBP is to the overall set of 11 positive RBPs (see “Materials and methods” for more details). Results of the ranked proteins along with their graph diffusion (GD) scores are shown in Table 3 and Supplemental Table 1. Among the top-ten-ranked RBPs was RBM45, which our group previously reported to localize to cytoplasmic inclusions in ALS cases [10]. In addition, MTHFSD, an RBP reported by the Robertson group in 2016 to be a novel component of stress granules and altered in ALS was ranked at number 10 by Watson [37]. Other RBPs previously linked to ALS and ranked in the top 5% included hnRNPA3, SMN2 and EWSR1 [13, 14, 40] (Table 3).

We focused validation studies on the top-ten-ranked RBPs and asked whether they were altered in ALS. Validation methods included protein subcellular distribution using IHC, measures of protein levels by immunoblot, RNA levels by total tissue extracts and laser-captured microdissection, and RNA analysis of motor neurons generated from patient-derived iPS cells. RBPs had to exhibit statistically significant differences between ALS and controls by at least two methods to be termed validated in our study. Our top-ten proteins included two previously shown to be altered in ALS (RBM45, and MTHFSD), so we did not pursue these further. Unsurprisingly, many of the IBM Watson top-ten-ranked RBPs are involved in RNA processing and export, and four out of the eight proteins we validated were contained within supplemental tables of proteins that potentially interact with TDP-43 and/or FUS in recent proteomic studies (Table 4). It is noteworthy that these interactions were listed in supplemental materials within these publications, not within the published abstracts, and as such, the putative protein interactions were not made available to IBM Watson for its analysis. The Watson top-ranked RBP, hnRNPU, possesses both RNA and DNA binding domains and potentially interacts with TDP-43, FUS, ubiquilin2 and the G4C2 repeat of C9orf72 [4, 19, 23, 24, 34, 49]. In addition, hnRNPU was recently shown to modulate nuclear TDP-43 toxicity in cultured cells [49], but was not directly linked to ALS. The second-ranked protein, Syncrip, is an RBP that resides in the cytoplasm and has been identified as an SMN-interacting protein found in RNA granules [34, 46, 53]. It was identified by mass spectrometry-based proteomics as a potential interacting protein with TDP-43, Ataxin-2, FUS, optineurin and ubiquilin, but again was never studied in ALS [4]. RBMS3 (ranked 4), hnRNPH2 (ranked 6) and RBM6 (ranked 9) all function in RNA metabolism or processing and have no prior links to ALS. NUPL2 (ranked 7), a nucleoporin-like protein interacts with Gle1, functions in CRM1-mediated RNA export and is a risk locus for Parkinson’s disease [15, 30, 41]. SRSF2 (also known as SC-35, ranked 5) is a nuclear speckle component that potentially interacts with TDP-43, FUS and the G4C2 repeat expansion and has been shown to co-localize with 34% of C9 antisense RNA foci in cerebellar Purkinje cells of C9-ALS patients [12]. Caprin-1 (ranked 8) is an RBP involved in neuronal RNA transport that interacts with FMRP and G3BP [17, 47]. Recently, a comprehensive proteomic study investigating common interactors for TDP-43, FUS and Ataxin-2 identified Caprin-1 as interacting with all 3 proteins [4]. In addition, they demonstrated that Caprin-1 co-localized with TDP43 and FUS inclusions in spinal cord motor neurons from 3 patients with TDP-43 inclusions and 2 patients with FUS-R521C mutations, respectively [4]. These results were published after our Watson analysis of the literature (Caprin-1 also does not appear in the abstract) and confirms our IBM Watson prediction that Caprin-1 is altered in ALS.

We validated the top-ten-ranked RBPs using immunohistochemistry (IHC) of lumbar spinal cord and cerebellum from SALS, C9-ALS and non-neurologic disease controls. RBM45 and MTHFSD were previously shown to be altered in ALS spinal cord using a similar IHC approach [10, 37]. The recent discovery of G4C2 repeat foci in the cerebellum [2], along with global splicing changes in both SALS and C9-ALS cerebellum [44] prompted us to examine potential RBP changes in the cerebellum. To test the specificity of IBM Watson results, we also performed immunohistochemistry for three RBPs from the bottom of the list (QTRT1, NARS and WARS). Our IHC results are summarized in Table 5.

IHC for hnRNPU was performed and nuclear staining pixel intensity was quantified as described in the “Materials and methods”, and reported as negative, weak, medium or strong immunoreactivity. hnRNPU exhibited negative or weak immunoreactivity in 70–80% of spinal cord motor neuron nuclei of control cases (Figs. 1a, 3a). Conversely, spinal motor neurons in SALS cases displayed strong hnRNPU nuclear staining in a majority (50–85%) of neurons per case. C9-ALS cases exhibited medium hnRNPU signal intensity that was significantly different from SALS neurons but failed to show significance when compared to controls, likely due to one of the of four C9-ALS cases being negative for hnRNPU. In addition, we observed increased glial staining in ALS compared to controls (Fig. 1a), and multiple cytoplasmic as well as nuclear hnRNPU inclusions in both SALS and C9-ALS motor neurons that co-localized with TDP-43 (Fig. 4a). In the cerebellum, hnRNPU showed weak immunostaining in nuclei of Purkinje and granule cells of control subjects, while three out of five C9-ALS cases displayed medium-to-strong IHC patterns (Suppl. Figure 2c). Purkinje cells in SALS had variable staining intensities, ranging from negative to strong.

Syncrip and RBMS3 both showed variable IHC staining patterns in spinal motor neurons of control as well as SALS and C9-ALS, with no significant differences detected between the different subject groups (Suppl. Figure 1a–b). In the cerebellum, Syncrip immunoreactivity was significantly increased in ALS versus controls (Fig. 2a). Purkinje cells displayed weak Syncrip immunostaining in control subjects (89% of neurons displayed negative-to-weak immunoreactivity; Fig. 3b), while neuronal staining was significantly increased in both the C9-ALS and SALS groups (58 and 64% of neurons were associated with medium-to-strong Syncrip staining in each subject group, respectively). Syncrip staining in SALS tend to be nuclear, while many C9-ALS cases showed more diffuse cytoplasmic immunoreactivity.

We detected no differences of RBMS3 staining in cerebellar Purkinje or granule cells in ALS (Fig. 2c). However, there was increased RBMS3 staining in cerebellar interneurons in the molecular layer as well as the granular and Purkinje layers of ALS cases. RBMS3-positive interneurons were found in SALS and C9-ALS cases co-labeled for the interneuron marker calretinin (Fig. 5a). Based on location and cellular morphology of these interneurons, these were identified as Lugaro cells, characterized by dendrites running parallel to the Purkinje cell layer [22], as well as unipolar brush cells, (calretinin-positive parvalbumin-negative, located in the granular layer), Golgi cells (calretinin and parvalbumin-positive cells) and basket or stellate cells (calretinin negative, parvalbumin positive; Fig. 5a; [18]).

The nuclear speckle protein SRSF2/SC-35 displayed a nuclear punctate staining pattern in all subjects, with some ALS cases showing strong and larger speckles, while one SALS case also exhibited cytoplasmic SC-35 inclusions that did not co-localize with TDP-43 (Figs. 1b, 4b, white arrowheads). Occasional neuropil staining for SC-35 was also observed for ALS, as well as increased glial immunoreactivity. In addition, SC-35 positive cytoplasmic tangle-like inclusions and neuropil staining were detected in the frontal cortex of C9-ALS and frontotemporal lobar degeneration (FTLD) (FTLD-tau and FTLD-TDP) cases, but not SALS or controls (Suppl. Figure 3a). SC-35 IHC levels were reduced in the frontal cortex of three SALS compared to three controls, though further studies are required to validate these findings. In the cerebellum, strong SC-35 immunostaining was associated with large nuclear speckles in Purkinje cells of ALS versus controls (Suppl. Figure 2b).

No significant differences in hnRNPH2 localization or staining intensity were observed in either SALS or C9-ALS cerebellum and spinal cords when compared to controls (Suppl. Figures 1c and 2a).

The nucleoporin-like protein NUPL2, ranked 7 by Watson, showed variable immunoreactivity in spinal cord motor neurons (Suppl. Figure 1d), with cytoplasmic puncta detected in most ALS cases and prominent nucleolar staining in two ALS cases. Strong astrocytic NUPL2 staining was also detected in four out of five ALS cases, and one out of two C9-ALS cases. However, no consistent neuronal IHC pattern was noted that differentiated ALS from controls. In the cerebellum, NUPL2 was also localized to astrocytes in SALS but not C9-ALS or control cases (Fig. 2d and Suppl. Figure 3b). In four out of five SALS cerebellum, but only one out of four C9-ALS cases, moderate-to-strong NUPL2 staining was observed in the granular layer and white matter, as well as in fiber tracts of the molecular layer (Suppl. Figure 3b). NUPL2 co-localized with GFAP in the cerebellum, indicating that NUPL2-positive cells were indeed cerebellar astrocytes (Fig. 5b). Purkinje, granule neurons and interneurons were typically negative for NUPL2.

Caprin-1 localizes to cytoplasmic granules in control motor neurons and Purkinje cells as previously described (Figs. 1c, 2b; [47]). ALS motor neurons and Purkinje cells exhibited larger and strong Caprin-1-positive cytoplasmic granules (Figs. 1c, 2b). These large cytoplasmic granules occasionally co-localized with p62 (Fig. 4c) indicating that some of these granules were cytoplasmic inclusions. In addition, most SALS cases displayed Caprin-1 redistribution to the nucleolus, while none of the control or C9-ALS cases had any nuclear or nucleolar Caprin-1 immunostaining.

We detected weak RBM6 immunoreactivity in spinal cord motor neurons, and none in the nucleus of control spinal motor neurons or Purkinje cells (Fig. 1d and Suppl. Figure 2d). ALS cases displayed increased RBM6 spinal cord motor neuron nucleolar immunoreactivity. Similar to Caprin-1, this phenotype was exclusive to the SALS group and not observed in C9-ALS. No differences across the subject groups were seen for RBM6 in the cerebellum (Suppl. Figure 2d).

We performed a similar IHC analysis in the spinal cord and cerebellum with three RBPs from the bottom of the IBM Watson-ranked candidate list: QTRT1 (ranked at position 1467), WARS (position 1463) and NARS (ranked at 1453), each chosen for commercial availability of specific antibodies. No differences were seen for these proteins in either the cerebellum or spinal cord (Suppl. Figure 4), indicating that RBPs near the bottom of the Watson ranking are not altered in ALS.

To further confirm IBM Watson’s top-ranked RBPs, we measured their transcriptional levels in total RNA from spinal cord and cerebellum of ALS patients and non-neurologic disease controls (Fig. 6). Of the eight top-ten RBPs examined, only RBMS3 showed significant decreases in transcriptional levels in ALS spinal cord when compared to controls (Fig. 6a). When we performed a similar analysis in cerebellar tissues, four out of these eight RBPs, hnRNPU, Syncrip, hnRNPH2 and NUPL2 were significantly downregulated in ALS compared to controls (Fig. 6b). None of the bottom-ranked proteins showed significant changes in gene expression in spinal cord or cerebellum (Fig. 6a, b).

Since RNA extracted from total spinal cord or cerebellum tissues includes various cell types, we also investigated transcriptional levels of these same RBPs in patient-derived pluripotent stem cells (iPSC) differentiated into motor neurons. We isolated RNA from motor neurons generated from five C9-ALS iPSC lines, two SALS iPSC cell lines, and three control iPSC lines (Fig. 7a). Both Caprin-1 and Syncrip showed significant upregulation in C9-ALS iPSC-MN compared to controls, while RBMS3 showed decreased expression in the two SALS lines compared to controls, recapitulating results obtained in total spinal cord, although a larger sample size is needed to confirm these findings.

To quantify gene expression alterations in cerebellar Purkinje cells, we used laser-capture microscopy (LCM) to isolate individual Purkinje cells from frozen cerebellum sections of ALS and neurologic disease controls (Fig. 7b). Approximately 250 Purkinje cells were isolated per sample to examine transcriptional levels for each candidate RBP. Syncrip was upregulated in four out of seven SALS cases compared to three controls, while hnRNPU was increased in three out of seven SALS cases. None of the other RBPs examined exhibited statistically significant transcriptional changes in ALS-Purkinje cells when compared to controls (data not shown).

We next investigated protein levels of IBM Watson-ranked RBPs by western blot analysis in cerebellum and spinal cord tissues. Protein levels of hnRNPU were increased in many ALS samples compared to non-neurologic disease controls in both cerebellum and spinal cord, mirroring increases observed by IHC in spinal cord (Fig. 8a, b and Table 5). However, results failed to reach statistical significance due to the large sample-to-sample variability within the ALS group. Syncrip levels were significantly increased in ALS cerebellum, but not spinal cord, again reflecting changes seen by IHC (Fig. 8a, b and Table 5). RBMS3 had significantly increased protein expression levels in ALS cerebellum and spinal cords compared to non-neurological disease controls, recapitulating observed increased cerebellar interneuronal staining (Fig. 8a, b and Table 5). NUPL2 was significantly decreased in ALS spinal cords, in apparent contradiction to its increased IHC in ALS spinal cord and cerebellum. No significant changes were seen for negative controls NARS or WARS (data not shown).