Your concern is valid and common when working with subcellular localization prediction tools, especially for chloroplast transit peptides (cTPs), which can be highly variable and tool-dependent.

Here’s a comprehensive and accurate explanation of what’s happening and what you can do:

1. Why Do Different Tools Give Different Results?

Each tool you mentioned—WoLF PSORT, TargetP 2.0, PredSL, iPSORT, DeepLoC—uses different algorithms, training datasets, and definitions of targeting signals. This leads to inconsistencies, especially for borderline or non-canonical sequences. For example:

• TargetP and PredSL are more specific to plant chloroplast signals.
• WoLF PSORT is more general and less accurate for plants.
• DeepLoc uses deep learning but is not specifically trained for plant chloroplasts.
• iPSORT is older and less reliable for complex targeting.

Thus, it's not uncommon to get contradictory predictions, particularly for multi-targeted proteins or when transit peptides are not very clear.

2. Input Sequence: Full-Length vs. N-terminal (e.g., first 70 AA)

This is a critical point:

• Transit peptides (such as chloroplast or mitochondrial signals) are located at the N-terminus.
• Most prediction tools (like TargetP, PredSL, iPSORT) use only the N-terminal segment, typically the first 50–100 amino acids, because the signal is encoded there.

Therefore:

Use the full-length sequence for tools that automatically trim to N-terminal (like TargetP 2.0). If a tool asks for just the N-terminal region, you can manually input the first ~70 AA. Do not input only the N-terminal unless the tool is designed to work that way. Otherwise, predictions might be skewed or invalid.

3. Discrepancy with Literature (e.g., PgSDH3_2 vs. PgSDH3_1)

In your case, the literature (Habashi et al., 2019) experimentally characterized PgSDH3_2 as plastid-localized, while your prediction tools suggest PgSDH3_1 based on signal peptide analysis.

Possible explanations:

• There might be alternative isoforms or splice variants.
• The experimental evidence in the paper (e.g., GFP fusions) should take precedence over in silico predictions.
• Your input sequences may differ (e.g., due to annotation errors or incorrect start codons).
• Some isoforms might be dual-targeted or have ambiguous transit signals.

Recheck the sequence alignment and annotation of your PgSDH isoforms. It’s crucial that you are using the correct translation start site—even a few residues off can change the targeting signal.

4. What Should You Do? Recommendations:

• Cross-validate: Use multiple tools, but give greater weight to plant-specific tools like TargetP 2.0 and PredSL.
• Manual curation: Compare the N-terminal region of your sequences with known chloroplast-targeted proteins in UniProt or TAIR.
• Experimental evidence trumps predictions: If the literature says PgSDH3_2 is plastidial and has experimental support (e.g., GFP localization), prioritize that.
• Consider signal alignment: Run a multiple sequence alignment of PgSDH isoforms (including from related species) to look for conserved N-terminal motifs.
• Experimental validation (if possible): GFP-tagging or chloroplast import assays provide definitive answers.

Summary Answer

The discrepancy in your predictions arises from differences in algorithmic approaches among prediction tools and the variability of chloroplast transit peptides. It's generally appropriate to use the full-length amino acid sequence for prediction, as most tools automatically assess the N-terminal region. However, some tools may require just the first 70 amino acids if they only process the N-terminal part. Discrepancies between predictions and published results (such as for PgSDH3_2) may be due to annotation issues, isoform confusion, or limitations of in silico tools. Ultimately, experimental evidence should guide interpretation, and careful sequence validation is recommended.

Let me know if you want help aligning your sequences or if you'd like a table comparing your tool results for each PgSDH isoform.