# Protein Sequence Analysis Tool Mistakes: Causes and Fixes The single most common cause of [protein sequence analysis](/blog/guides/protein-sequence-analysis-a-practical-guide-to-domains-motifs-and-homology) tool mistakes is a bad input. A sequence that is in the wrong format, contains hidden characters, comes from the wrong isoform, or was copied with extra residues will produce plausible-looking but wrong results from almost any tool. The first thing to check, before questioning the software itself, is the exact sequence you submitted and the exact record the tool reports back. If the tool echoes a sequence length, a protein name, or a first and last residue that do not match your expectation, stop there. Everything downstream inherits that error. This article walks through the causes of [protein sequence](/blog/guides/protein-sequence) analysis tool mistakes in order of how often they occur, explains the mechanism behind each failure, and gives concrete fixes. The patterns below apply to BLAST searches, multiple sequence alignments, domain and motif scans, subcellular localization predictors, transmembrane helix finders, and comparative modeling pipelines alike. ## Why Protein Sequence Analysis Tools Produce Wrong Results Prediction tools do not read your mind. Each one compares your query against a model, a database, or a set of learned patterns, and the answer is only as good as the match between your protein and that model. Bigelow and Rost make this point explicitly for membrane protein prediction: accuracy depends strongly on how closely the query fits the model, so the practical approach is to submit the protein to several methods and weigh the results by confidence score [1]. The same logic extends across the field. Mooney and colleagues note that because user queries vary in size, scope, and character, trade-offs in speed, accuracy, and scale must be weighed when choosing which method to adopt [2]. Tool failure therefore has a short list of root causes: 1. The input does not represent the protein you think it does. 2. The tool was built for a different kind of protein or a different question. 3. The tool's model does not cover your protein's family or region. 4. The output was misread, often because a score, threshold, or coordinate system was misunderstood. 5. Errors already present in public databases propagated into your analysis. Each of these has a diagnostic signature and a fix, covered in the sections that follow. ## Troubleshooting Table: Symptom, Cause, Confirmation, Fix | Symptom | Likely cause | How to confirm | Fix | |--|--|--|--| | Tool returns a sequence shorter or longer than expected | Truncated FASTA, stop codons left in, wrong isoform, or non-ASCII characters pasted in | Compare the echoed sequence length and first and last residues against the UniProt entry for the correct isoform | Re-export the sequence from the canonical source, strip whitespace and asterisks, and resubmit | | BLAST finds no significant hits for a protein you know has homologs | Remote homologs below sequence-search sensitivity | Try fold-recognition or profile-based methods, which detect homologs at much lower identity | Use iterative profile searches or fold-recognition servers, as demonstrated for sialyltransferases where BLAST and PSI-BLAST failed but fold recognition succeeded [3] | | Two domain scanners disagree on domain boundaries | Different hidden Markov model libraries or thresholds behind each tool | Check which library (for example Pfam versus SUPERFAMILY) each result came from | Run a combined pipeline such as InterProScan so all libraries are reported together, then compare architectures consistently [4] | | Localization predictors disagree wildly | Each tool emphasizes different sequence features | Note which features each tool reports, such as [signal peptide](/knowledge/molecular-biology/signal-peptide) versus transmembrane segments | Integrate outputs with a structured strategy, since integration raised mitochondrial prediction accuracy to 92 percent, far above individual tools [5] | | Alignment looks wrong at the ends or in low-similarity blocks | The alignment tool was a poor match for the sequence family's similarity level | Compute pairwise identity, then compare against benchmarks showing where tools differ | Select the alignment tool based on sequence similarity, and consider realignment pipelines trained for low-similarity families [6] | | Model structure looks distorted in one loop | Template mismatch in that region, often the CDRs in antibodies | Compare per-region RMSD or MolProbity statistics against the experimental structure if one exists | Recognize that no single modeling program is uniformly superior, and evaluate several models before trusting one [7] | | Residue numbering does not match the literature | PDB structures carry their own numbering that differs from the UniProt sequence | Check whether the structure record is aligned to the UniProt canonical sequence | Use a tool that reports coverage aligned to the UniProt sequence and resolves numbering differences [8] | | A "ligand" appears in a structure where none was expected | Overinterpretation of ambiguous electron density | Inspect the electron density maps directly | Validate ligand models with dedicated checks, since non-experts routinely misinterpret weak density [9] | ## Cause 1: Bad Input, the Most Common Mistake Most protein sequence analysis tool mistakes begin before the tool runs. A FASTA file pasted from a word processor can carry curly quotes, zero-width characters, or line-wrap artifacts. A sequence copied from a paper's figure may contain typographical substitutions. A gene may have multiple isoforms, and the one you pasted may lack a domain present in the canonical form. The mechanism is simple. These tools parse text, not intent. A single character that is not a valid amino acid code can shift the reading of an entire record, or a tool may silently discard it and misalign everything downstream. When a genome-scale annotation pipeline feeds predicted proteins into sequence analysis, mistakes in the upstream annotation propagate, which is exactly the class of problem that MICheck was built to detect in bacterial genomes, where sequences were sometimes annotated as genes when they were not protein-coding [10]. **How to fix it:** - Retrieve the sequence directly from a curated source, ideally the UniProt canonical entry, rather than from a PDF, a supplementary table, or a colleague's email. - Paste the sequence into a plain-text editor first and confirm that every character is a valid amino acid letter. - Record the accession number and isoform identifier in your [lab notebook](/knowledge/molecular-biology/lab-notebook) or analysis log. A sequence without provenance cannot be verified later. - After submission, read the tool's echo of your input. Most web servers report the sequence length and identifier they received. Check it every time. ## Cause 2: Relying on a Single Tool When Tools Disagree A close second among protein sequence analysis tool mistakes is treating one tool's output as ground truth. Dozens of tools exist for predicting protein subcellular location in eukaryotes, and their predictions frequently disagree for the same protein, because each tool performs well on certain datasets and emphasizes different sequence features [5]. Bigelow and Rost describe the same phenomenon for membrane protein prediction and recommend an opportunistic strategy, submitting the protein of interest to all available methods and choosing results with the highest confidence scores [1]. The mechanism behind the disagreement is that each tool encodes a different model of what a mitochondrial targeting sequence, a signal peptide, or a transmembrane helix looks like. A protein with an unusual feature, such as a signal anchor that also functions as a transmembrane segment, will satisfy one model and violate another. Shen and Burger showed that integrating the outputs of multiple localization tools with a two-layer [decision tree](/blog/data-analysis/what-is-a-decision-tree), one that emphasized biologically relevant features such as the targeting peptide and transmembrane domains, reached up to 92 percent accuracy for mitochondrial proteins, far above any individual tool [5]. **How to fix it:** - Run at least two independent tools for any categorical prediction, and record the confidence value from each. - When tools disagree, look at the sequence features each tool reports. Disagreement usually traces to one interpretable region, such as an N-terminal extension or a hydrophobic stretch. - Prefer consensus or integration approaches where they exist. Pánek demonstrated for neurotransmitter transporters that combining standard [multiple sequence alignment](/blog/guides/multiple-sequence-alignment) with an alternative method for hydropathy distribution analysis, then validating against literature and knowledgebase annotations, produced a consensus classification that identified previously uncharacterized protein families [11]. ## Cause 3: Choosing the Wrong Tool for the Similarity Regime Not every tool works at every evolutionary distance. BLAST and PSI-BLAST excel at detecting homologs with recognizable sequence similarity, but when similarity drops below roughly 20 percent, profile-based searches and fold-recognition methods become necessary. In a study of human sialyltransferases, sequence similarity servers could not detect any homolog, while fold-recognition servers identified a bacterial sialyltransferase as a remote homolog at only 15 to 20 percent sequence similarity [3]. Even then, the fold-recognition servers could not find templates for all regions, and the authors had to perform template identification, modeling, and evaluation iteratively, using multiple structural templates for the uncovered regions [3]. The same regime-dependence applies to multiple sequence alignment. MLProbs addressed this by training simple classification models to choose a better alignment tool and to determine whether and how much realignment helps. On families with sequence similarity at or below 50 percent, MLProbs achieved a TC score of 56.93, while the next three best tools ranged from 55.41 to 55.91, an improvement of more than 1.8 percent on precisely the families where alignment is hardest [6]. **How to fix it:** - Estimate pairwise identity across your sequence set before choosing an alignment method. If identity is low, default settings of general-purpose tools are not the best choice. - For remote homolog detection, move from pairwise search to profile or fold-recognition methods rather than concluding that no homolog exists. - Accept that some regions may have no reliable template. Modeling them separately with structural templates, then assembling, is a legitimate strategy [3]. ## Cause 4: Domain and Motif Misannotation Domain architecture, the order and combination of domains along a protein, drives function, interaction, and evolutionary interpretation. Mistakes here are common because different domain libraries use different hidden Markov models and different significance thresholds. The d-Omix platform was built on the observation that domain combination provides important clues to [protein function](/blog/guides/protein-function), and it lets users select among domain search tools such as hmmpfam and SUPERFAMILY while unifying the downstream comparison of domain architectures [4]. Pattern-based mistakes also occur in specialized domains. The canonical structure model for antibody hypervariable loops depends on precisely defined sequence patterns, and the authors of a reanalysis found that the enormous growth in antibody structural data had introduced mistakes and confusion into the model's patterns for loops L1 and L3. They proposed corrections that made the patterns unambiguous and improved forecasting of canonical structures [12]. The general lesson applies to every motif or pattern resource: the pattern definitions themselves can contain errors, so a prediction that contradicts strong experimental evidence deserves a second look at the pattern source. **How to fix it:** - Run domains through a combined resource such as InterProScan, which reports matches from multiple member databases under one output, and feed that unified output into downstream comparison tools such as d-Omix [4]. - Compare domain architectures across a protein family, not within a single protein. A domain call that appears in one protein but breaks the family's architecture pattern is a candidate for a false positive. - For antibody and repeat-protein work, verify loop classifications against the current pattern definitions, since older pattern sets contained documented errors [12]. ## Cause 5: Structure Selection and Template Errors ![Decision path weighing coverage of the region, bound ligands, model quality and numbering before picking a structure.](/article-images/illustrations/protein-sequence-analysis-tool-mistakes-structure-choice.webp) Computational methods that rely on [protein structure](/blog/guides/protein-structure) depend heavily on which structure was selected for the analysis. Degn and colleagues frame this directly in PDBminer, noting that experimental PDB entries and AlphaFold Database models each carry significant advantages and drawbacks, and that exploring the wealth of available structures to find suitable ones is a daunting task prone to error [8]. Two failure modes dominate. The first is picking a structure with poor coverage of the region you care about. PDBminer addresses this by aligning every available structure to the UniProt sequence and reporting coverage, model quality, complex partners, and bound ligands in a ranked table, which overcomes the numbering differences between PDB records and the reference sequence [8]. The second is template error in comparative modeling. When a template shares only marginal similarity with the target, errors concentrate in loops and solvent-exposed regions. The sialyltransferase study showed that template identification must be treated as iterative, evaluating each candidate template and rebuilding weak regions from alternative structural templates [3]. Antibody modeling illustrates the ceiling of the approach. Zhao and colleagues solved the crystal structure of an anti-ricin Fab at 1.9 angstrom resolution and compared it against models from PIGS, RosettaAntibody, and WAM. Differences between the crystal structure and the predicted models concentrated in the complementarity-determining regions, and no one modeling program was clearly superior to the others [7]. A related trap is the self-reference problem. Modeling programs are validated against structures already deposited in the [Protein Data Bank](/blog/guides/protein-data-bank), so their apparent accuracy can be inflated when the benchmark structure influenced tool development. Zhao and colleagues note that accuracy can only be tested fairly when a new structure has not yet been deposited [7]. For your own work, this means a model validated on deposited structures may still fail on your novel protein. **How to fix it:** - Use a structure-selection tool that reports coverage against the UniProt reference before you commit to a PDB entry [8]. - Match the resolution and quality requirements to the question. A structure with a bound ligand at moderate resolution may serve a binding-site analysis better than a higher-resolution unliganded structure. - Treat template choice as iterative. Identify, model, evaluate, and repeat for weak regions rather than accepting the first hit [3]. - When several modeling tools are available, build with more than one and compare, since no single program is uniformly best [7]. ## Cause 6: Misreading Scores, Thresholds, and Output Many protein sequence analysis tool mistakes happen after the tool finishes correctly. Common variants include reading an E-value as a probability, treating a transmembrane helix prediction with a borderline score as certain, or comparing TM-scores and Z-scores across different modeling runs as though they shared a scale. The antibody study cited above evaluated models on RMSD, TM-score, Z-score, and MolProbity metrics, and the small differences among predicted models fell within the range that could reflect method noise rather than real differences in quality [7]. Another frequent misreading involves residue numbering. PDB structures use the numbering of the crystallized construct, which often differs from the UniProt canonical sequence because of tags, truncations, or species differences. Tools such as iMolTalk exist precisely to map residues from sequence to structure and to extract the sequence implied by the three-dimensional coordinates, which lets you verify that residue 45 in the structure is the residue you mean [13]. PDBminer extends the same principle across a whole table of candidate structures by aligning each one to the UniProt sequence [8]. **How to fix it:** - Read the tool's documentation for the meaning and cutoff of every score you report. Write the definition into your methods section. - Map any residue number you cite back to the canonical sequence with a sequence-to-structure mapping tool [13]. - Report the range of model quality metrics rather than a single favorable one when comparing modeled structures [7]. ## Cause 7: Overinterpreting Marginal Evidence Sequence and structure tools can encourage overinterpretation because they always return something. The crystallography literature offers the sharpest warning. Pozharski, Weichenberger, and Rupp analyzed deposited ligand-protein complex structures and found that ambiguous ligand electron density is often difficult to interpret, that modeled ligands are generally harder to validate properly, and that some deposited structures described ligands where no ligands were present. They classified the common categories of building errors and overinterpretation and built a tool that lets non-expert users inspect ligand density directly [9]. The analogous mistake in sequence analysis is assigning a function based on a weak motif hit or a low-confidence localization score and then building an experiment on that assignment. In silico protein design tools carry the same risk, which is why Dahiyat emphasizes that the workflow requires deliberate selection of the protein structure, definition of the portion to be designed, and careful choice of input parameters, rather than accepting software defaults [14]. **How to fix it:** - Attach a confidence level to every computational claim before you act on it. - For structural claims, inspect the underlying evidence, such as electron density or alignment coverage, not just the summary output [9]. - Treat tool defaults as a starting point. For design work, choose the input parameters deliberately for your protein and goal [14]. ## Cause 8: Functional Annotation and Gene Ontology Misuse Gene Ontology (GO) terms enrich protein analyses, but comparing proteins by GO terms requires choosing an appropriate semantic similarity measure, and different measures weight the ontology differently. DaGO-Fun was developed to integrate many information-content-based GO similarity measures, including topology-based and annotation-based approaches, so users can select the measure most relevant to their application [15]. The mistake to avoid is running a GO comparison with whatever default measure a tool exposes and reporting the result as though the choice of measure did not matter. Annotation-based measures depend on the curated annotations available, which means they shift as databases update, so a similarity score computed today may not reproduce after the next annotation release. **How to fix it:** - Name the GO similarity measure in your methods, not just the tool. - When possible, compare results under more than one measure class, since topology-based and annotation-based measures can rank protein pairs differently [15]. ## Cause 9: Format Conversion and Pipeline Fragmentation Multi-tool pipelines break at format boundaries. A tree built from an alignment exported in one format may silently drop sequences or gaps when imported into a package expecting another. DAMBE handles conversion among more than 20 standard sequence formats, including MEGA, NEXUS, PHYLIP, GenBank, and NeXML, precisely because interoperability failures are routine [16]. The mechanism of failure is usually a parser assuming fields or line widths that the writer did not produce. Results range from truncated alignments to phylogenies computed on a subset of the data you intended. **How to fix it:** - After any conversion, verify sequence counts, lengths, and gap positions in the new file. - Prefer formats designed for interoperability, such as NeXML, when chaining tools [16]. - Keep the original downloaded file immutable, and do conversions on copies so you can always re-derive a result. ## Cause 10: Physicochemical and Comparative Workflows Applied Too Narrowly Physicochemical properties of primary sequences, such as isoelectric point and hydrophobicity profiles, inform structure and function, and GUI-based tools now make these computations accessible to experimentalists as well as bioinformaticians [17]. The mistake is not using them but using them in isolation. A physicochemical profile that flags a protein as unusually acidic means little without family context. The Rab5 comparative study shows the stronger pattern, combining physicochemical profiling, modeling, and docking across pathogenic species to arrive at structural and functional insight [18]. Similarly, the VAR2CSA vaccine study combined in silico tools with peptide arrays and structural modeling, so that computational predictions about sequence variation and antibody targeting were tested experimentally rather than asserted [19]. **How to fix it:** - Anchor every physicochemical observation to a family-level or literature context. - Pair in silico predictions with an experimental or orthogonal computational check before treating them as findings [19]. ## A Short Diagnostic Sequence ![Flowchart of seven checks from confirming the input sequence through running a control protein with known answers.](/article-images/illustrations/protein-sequence-analysis-tool-mistakes-diagnostic-sequence.webp) When a protein sequence analysis tool gives a result that looks wrong, work through this list in order. Most errors are caught by step 3. 1. Confirm the input. Check the sequence length, the first and last residues, the isoform, and the absence of non-ASCII characters, against the canonical database entry. 2. Confirm the tool's echo. Compare what the server reports receiving with what you intended to send. 3. Run a second, independent tool for the same question and compare predictions with their confidence scores. Expect disagreement on borderline proteins and investigate the disagreeing region. 4. Check the similarity regime. For searches, ask whether the homology might be below pairwise-detection sensitivity, and consider fold recognition. For alignments, ask whether the family's low similarity calls for a specialized method [6]. 5. Verify identifiers and numbering. Map structure residue numbers to the UniProt sequence before quoting positions [13] [8]. 6. Inspect the raw evidence behind any structural claim, including coverage, model quality, and, for ligands, the electron density itself [8] [9]. 7. Re-run the analysis on a known control protein with established answers. If the control also fails, the problem is the tool, the parameters, or the pipeline, not your protein. ## Bioinformatics Tools Comparison as a Corrective Practice A recurring theme in the literature is that no single tool wins, and that structured comparison of tools is itself a quality-control method. This is why the best bioinformatics software for a given task is defined by fit rather than by a universal ranking. For localization, integration of specialized tools outperformed every individual predictor [5]. For membrane prediction, an opportunistic multi-method strategy is recommended explicitly [1]. For alignment, pipelines that choose the tool based on measured sequence properties outperform any fixed choice on difficult families [6]. For domain analysis, platforms that unify multiple libraries under one architecture comparison prevent library-specific artifacts [4]. For structure selection, ranking all available experimental and predicted structures against your protein beats picking the first hit [8]. Build the comparison habit into your standard workflow. Keep a small validation set of proteins with experimentally established answers for each analysis type, run new tools against it, and record which tools performed well on proteins resembling yours. ## Common Mistakes and Limitations - Trusting a single tool for a categorical prediction. Localization and membrane tools disagree routinely on borderline proteins, and integration outperforms individual predictors [5] [1]. - Assuming no homolog exists because BLAST found none. Remote homologs require profile and fold-recognition methods [3]. - Comparing residue numbers between a PDB structure and a UniProt sequence without mapping. Numbering differences are documented and systematic [8]. - Quoting a model as if it were a structure. Comparative models, especially in CDR loops, can differ measurably from experimental structures, and no modeling program is uniformly superior [7]. - Accepting ligand placements without checking density. Overinterpretation of ambiguous density is a documented, recurring error class [9]. - Reporting GO similarity without naming the measure. Different measures can rank the same protein pairs differently [15]. - Ignoring isoforms. Tools cannot know which isoform you meant, and domain architecture often differs between isoforms. - Running pattern-based predictions against outdated pattern definitions. Documented corrections to established pattern models exist, which means older definitions may misclassify [12]. - Letting annotation errors propagate. Errors in upstream gene annotation reach public databanks and then flow into every downstream sequence analysis built on them [10]. - Skipping provenance. A sequence analysis that cannot be traced to an accession number, a version, and a date cannot be reproduced or audited. One limitation deserves a plain statement. Computational predictions guide experiments, they do not replace them. Every high-impact claim built on sequence analysis should carry an experimental or orthogonal validation, as the strongest published workflows in this area consistently do [3] [19]. ## Prevention Practices **Version and provenance discipline.** Record the tool name, version, database release, and date for every run. Annotation-based resources change, and a result tied to a dated release remains interpretable. **Fixed validation sets.** For each tool class you use, maintain a handful of proteins with known answers. Run them whenever you adopt a new tool or update an existing one. **Two-tool minimum for categorical calls.** Any yes-or-no prediction, such as mitochondrial localization or the presence of a transmembrane helix, gets at least two independent tools, with disagreement treated as a flag for manual inspection rather than a tie to break arbitrarily [5] [1]. **Write the parameters down.** For design and modeling work, the choice of structure, the region defined for design, and the input parameters determine the outcome, so document all three [14]. **Check the echo, always.** The few seconds spent confirming the tool received the intended sequence prevent the most common class of protein sequence analysis tool mistakes. **Convert with verification.** After every format conversion, verify sequence counts and lengths before continuing [16]. **Match the method to the similarity regime.** Estimate identity first, then choose search and alignment methods built for that regime, and expect iterative template work when similarity is marginal [6] [3]. ## Frequently Asked Questions ### What is the most common protein sequence analysis tool mistake? The most common mistake is submitting an incorrect or corrupted input sequence, such as the wrong isoform, a truncated FASTA record, or text pasted with hidden characters. Check the sequence the tool echoes back against the canonical UniProt entry before trusting any output. ### Why do two bioinformatics tools give different answers for the same protein? Each tool encodes a different model of the feature it predicts, and accuracy depends on how closely your protein fits that model. Disagreement is expected on borderline proteins, and integrating several tools' outputs with their confidence scores outperforms any single predictor [5] [1]. ### How do I fix a BLAST search that finds no homologs? Confirm the input first, then try profile-based and fold-recognition methods, which detect homologs at much lower sequence identity than standard pairwise searches. In published work, fold recognition identified homologs that BLAST and PSI-BLAST missed entirely [3]. ### Should I use an experimental PDB structure or an AlphaFold model? It depends on your question, since both options have significant advantages and drawbacks. Use a selection tool that ranks all available structures by coverage against the UniProt sequence, model quality, ligands, and binding partners, then choose based on fit [8]. ### Can I trust a modeled protein structure? Treat a model as a hypothesis with measurable uncertainty. Errors concentrate in loops and low-similarity regions, comparisons across modeling programs show no single tool is uniformly superior, and validation against truly novel structures is complicated by the self-reference problem [7]. ### How do I avoid mistakes when comparing protein domains? Run domains through a combined resource so multiple libraries are reported consistently, then compare domain architectures across a protein family rather than within one protein. Unified platforms exist specifically to standardize this comparison [4]. ### What confidence level should I require before acting on a prediction? Require enough confidence that a false positive would cost you less than the follow-up experiment, and treat borderline scores as prompts for a second method rather than as answers. Document which thresholds you applied so others can reproduce your reasoning. ### Do I need to validate computational predictions experimentally? For claims that will drive experiments, publications, or decisions, yes. The strongest published workflows pair in silico analysis with peptide arrays, structural determination, or literature and knowledgebase validation before drawing biological conclusions [11] [19]. ## Sources 1. [Online tools for predicting integral membrane proteins.](https://pubmed.ncbi.nlm.nih.gov/19153681/) 2. [In silico protein motif discovery and structural analysis.](https://pubmed.ncbi.nlm.nih.gov/21780007/) 3. [Fold-recognition and comparative modeling of human alpha2,3-sialyltransferases reveal their sequence and structural similarities to CstII from Campylobacter jejuni.](https://pubmed.ncbi.nlm.nih.gov/16620397/) 4. [d-Omix: a mixer of generic protein domain analysis tools.](https://pubmed.ncbi.nlm.nih.gov/19465389/) 5. ['Unite and conquer': enhanced prediction of protein subcellular localization by integrating multiple specialized tools.](https://pubmed.ncbi.nlm.nih.gov/17967180/) 6. [MLProbs: A Data-Centric Pipeline for Better Multiple Sequence Alignment.](https://pubmed.ncbi.nlm.nih.gov/35120007/) 7. [Crystal structure and computational modeling of the fab fragment from a protective anti-ricin monoclonal antibody.](https://pubmed.ncbi.nlm.nih.gov/23285112/) 8. [PDBminer to Find and Annotate Protein Structures for Computational Analysis.](https://pubmed.ncbi.nlm.nih.gov/37977136/) 9. [Techniques, tools and best practices for ligand electron-density analysis and results from their application to deposited crystal structures.](https://pubmed.ncbi.nlm.nih.gov/23385452/) 10. [MICheck: a web tool for fast checking of syntactic annotations of bacterial genomes.](https://pubmed.ncbi.nlm.nih.gov/15980515/) 11. [A comparative computational analysis of protein sequences and literature mining classify 'orphan' neurotransmitter transporters.](https://pubmed.ncbi.nlm.nih.gov/18621060/) 12. [Modifications to canonical structure sequence patterns: analysis for L1 and L3.](https://pubmed.ncbi.nlm.nih.gov/11933071/) 13. [iMolTalk: an interactive, internet-based protein structure analysis server.](https://pubmed.ncbi.nlm.nih.gov/15215439/) 14. [In silico protein design: fitting sequence onto structure.](https://pubmed.ncbi.nlm.nih.gov/16671409/) 15. [DaGO-Fun: tool for Gene Ontology-based functional analysis using term information content measures.](https://pubmed.ncbi.nlm.nih.gov/24067102/) 16. [DAMBE5: a comprehensive software package for data analysis in molecular biology and evolution.](https://pubmed.ncbi.nlm.nih.gov/23564938/) 17. [GUD-VE visualization tool for physicochemical properties of proteins.](https://pubmed.ncbi.nlm.nih.gov/37424755/) 18. [A comparative in silico analysis of Rab5 proteins from pathogenic species to find its role in the pathogenesis.](https://pubmed.ncbi.nlm.nih.gov/31432591/) 19. [Epitope mapping and topographic analysis of VAR2CSA DBL3X involved in P. falciparum placental sequestration.](https://pubmed.ncbi.nlm.nih.gov/17112315/)