ALS-associated FUS mutations induce site-specific alterations in conformational dynamics and molecular interactions

Amyotrophic lateral sclerosis (ALS) is a rapidly progressive and always fatal disease that is associated with the selective death of upper and lower motor neurons, and for which the molecular basis of disease causation by mutations identified in affected individuals is still poorly understood. A large proportion of ALS is associated with mutations of genes that code for RNA-binding proteins, and notably mutations in the gene that codes for FUS (Fused in Sarcoma), a highly multifunctional protein that participates in RNA biology and metabolic processes that is highly characterized by its intrinsically disordered regions. As part of this dissertation research, the structural and functional implications of three different mutations associated with ALS in the FUS protein (G156E, R234L, and R521C) that map to distinct functional regions of the protein were analyzed by a computational scheme that integrates structure prediction by AlphaFold, calculation of intrinsic plasticity and binding plasticity by FuzPred, and molecular modeling simulations by HADDOCK to study crucial protein-protein and protein-RNA interactions. These studies clearly show that, while there is no large structural change in the overall protein structure, each of the mutations causes a site-directed and functionally relevant alteration in conformational dynamics and protein-binding properties. In particular, the R521C mutation enhances the difference in binding affinity between the FUS Nuclear Localization Signal (NLS) and Transportin-1, indicating that there may be a mechanistic cause for impaired transport, whereas the G156E mutation destabilizes FUS-RNA binding, indicating that there may be a change in RNA-binding affinity. The results confirm that ALS-causing mutations in FUS cause disease primarily as a result of subtle changes in intrinsic disorder, binding versatility, or molecular recognition, as opposed to classical protein misfolding and that structural biology predictions have provided insights that have helped to further our understanding of ALS.

Amyotrophic lateral sclerosis (ALS) is a rapidly progressing and always fatal neurodegenerative disorder that selectively affects upper and lower motor neurons, resulting in muscle weakness or paralysis and ultimately death from respiratory failure. While traditionally considered a motor neuron disorder, it has become increasingly evident that ALS is a disorder of multiple systems that affects cortical and subcortical areas of the brain, glial cells, and has widespread molecular abnormalities (Hardiman et al., 2017; van Es et al., 2017). In terms of its presentation, ALS has a wide range of variability in age of onset, rate of progression, and symptomatology, with an average survival from time of diagnosis of two to 5 years (Brown and Al-Chalabi, 2017). In terms of its epidemiology, ALS has an annual incidence of 1 to 2 per 100,000 population and has both sporadic and familial cases.

In terms of molecular pathogenesis, ALS has increasingly been seen as a disorder of protein and RNA homeostasis. The hallmark pathological features include cytoplasmic protein aggregates, which are dominated by RNA-binding proteins such as TDP-43 and Fused in Sarcoma (FUS) (Ling et al., 2013). These advances have led to an important extension of the conceptual framework for ALS from a purely “neuronal” disorder to one involving abnormalities in RNA biology, proteostasis, nuclear cytoplasmic transport, and cellular phase separation (Taylor et al., 2016). More than thirty genes have now been identified as contributing to ALS pathogenesis, many of which are involved in RNA binding or protein quality control, cytoskeleton function, and intracellular trafficking. Significantly, a number of ALS-related proteins are now known to be rich in intrinsically disordered regions, which are functionally versatile but also potentially aberrantly aggregating and involved in aberrant biomolecular condensate formation.

Among these proteins, FUS plays a central role because of its diverse roles in the regulation of transcription, RNA splicing and transport, and DNA damage repair. FUS has a long, unstructured region with a high propensity for low-complexity regions, multiple RGG motifs, an RNA recognition motif, and a C-terminal proline-tyrosine nuclear localization signal (PY-NLS) in its structure. This organization allows FUS to bind RNA and proteins in a dynamic and multivalent fashion, making it particularly susceptible to mutations causing alterations in its structure and functions (Dormann and Haass, 2013). FUS mutations in ALS cases are found in specific functional domains with varying pathological consequences. FUS mutations in the PY-NLS, such as the R521C mutation, compromise its ability to bind with the nuclear import receptor transportin-1, leading to cytoplasmic mislocalization and causing early-onset ALS (Dormann et al., 2010). FUS mutations in its low-complexity or RGG-containing regions, like the G156E mutation, affect its RNA-binding properties and its ability to form phases without affecting its nuclear import activity (Bosco et al., 2010).

One of the hallmarks of FUS-related ALS is that disease-causing mutations are rarely associated with bulk protein misfolding. Rather, they are associated with subtle, localized modifications of intrinsic disorder, conformational dynamics, and binding plasticity. Such fine-scale modifications can, in turn, lead to dramatic biological effects, including the disruption of nuclear transport, RNA regulatory networks, and stress granule biology (Ling et al., 2013; Dormann et al., 2010). Mutations also cause FUS to aberrantly undergo phase transitions, causing it to transition from its physiological liquid-like condensates to pathological solid-like phases (Patel et al., 2015; Boeynaems et al., 2018). Such properties are very challenging to address using standard structural biology methods, especially in the context of highly disordered proteins such as FUS and other RNA-binding proteins (RBPs).

However, recent progress in computational structural biology has led to the development of powerful approaches to overcome these difficulties. AI-based structure prediction algorithms make it possible to model full-length proteins with large disordered regions, while disorder-focused predictors estimate the multiplicity of binding modes and disorder-to-order transitions (Krokidis et al., 2024). Integrative docking engines also enable the exploration of protein-protein and protein-RNA interactions in systems with conformational flexibility (Loureiro et al., 2025). These approaches collectively allow us to explore how ALS-associated mutations modulate molecular recognition and interaction properties even in the absence of stable folded structures.

In this work, we employ an integrated computational approach that combines structure prediction, intrinsic disorder prediction, and molecular docking to investigate the structural and functional impact of three ALS-associated mutations in the FUS protein, namely, G156E, R234L, and R521C, which are positioned in three distinct functional regions of the protein. By comparing wild type (WT) and mutant FUS proteins, we seek to elucidate mutation-specific changes in conformational dynamics, binding versatility, and important molecular interactions such as nuclear import and RNA binding. Our primary hypothesis is that ALS-associated mutations in the FUS protein primarily cause pathogenicity through localized disorder and binding property alterations rather than global structural instability.

Understanding these mechanisms is critical to closing the gap between genetic mutation and cellular dysfunction in ALS. While these findings are important in the context of FUS, they also have a broader application in understanding the mechanisms of neurodegeneration caused by mutations in intrinsically disordered RBPs. Understanding these mechanisms is important for developing future experimental studies and for the development of targeted therapies to correct protein localization, interaction, and RNA homeostasis in ALS.

Structural modelling and biophysical analysis of FUS variants

Retrieval of sequence data and structure prediction

Human FUS’s coding sequence (UniProt ID: P35637) was retrieved from UniProt in FASTA format. The original sequence was copied manually and used to create separate mutant FASTA files that contained the pathogenic changes G156E, R234L, and R521C. Other than these substitutions, no modifications were made to the sequence. The four variants were sent to ColabFold, which employed the AlphaFold2 framework to predict full-length structural models. For every structure, a total of five models were created, out of which the one with the highest ranking based on pLDDT values was chosen for further investigation.

In accordance with AlphaFold’s practices, pLDDT scores are kept in the B-factor column of PDB files produced as a result. These numbers were carried on for filtering the residues later on. All the models were made up of only one 526-residue long polypeptide chain, which made it possible to compare residues one by one without needing a structural alignment preprocessing step.

Structural comparison and biophysical feature extraction

After predicting the structures, the FUS models of both types (WT and mutant) were compared thoroughly in order to get the alteration of structures caused by the mutations. This was done using a Python script which was written for this dissertation and which combined Biopython with other libraries and the Shrake–Rupley method for calculating solvent accessibility. To limit the analysis to only those structural features that had physical meaning, the computations were carried out only on residues that had AlphaFold confidence scores (pLDDT) of 70 or above, thereby ruling out areas of low structural certainty around intrinsically disordered proteins which is a feature of such proteins. The algorithm, for each mutant, pinpointed the mutation automatically by looking at the differences in residue identity with respect to the WT. Specifically, the structural alignment and superposition were driven exclusively by the C-alpha (Cα) backbone coordinates of the high-confidence residues (pLDDT ≥70). A Kabsch algorithm was applied strictly to this masked subset to derive the optimal rotation matrix. By anchoring the alignment to these structurally reliable regions, the subsequent global Cα RMSD computation specifically measures the spatial deviation of these folded domains. The local structural perturbation was measured alongside by determining RMSD in a ±10-residue window around the mutation site that was detected. If the high-confidence residue population in the window was insufficient, the local RMSD was marked as not applicable to maintain scientific transparency. The physicochemical effects of each mutation were evaluated by computing per-residue solvent-accessible surface area (SASA) for both mutant and WT proteins. The resultant ΔSASA values are indicative of the change in surface exposure which is caused by the mutation. Besides that, each residue was given a Kyte–Doolittle hydrophobicity score which made it possible to compute a hydropathy-weighted ΔSASA term (HYDRO·ΔSASA), an integrated measure that reflects the extent to which exposure of hydrophobic regions might lead to aggregation. Meanwhile, the hydrophobic character of the proteins was expressed by the Grand Average of Hydropathy (GRAVY index) which was determined by a directly hydropathy-based calculation on the primary sequence.

All the features that were computed in the process, such as global RMSD, local RMSD, ΔSASA, HYDRO·ΔSASA, GRAVY score, and residue-level annotations, were made available not only in the form of graphical plots but also as per-residue spreadsheets. Additionally, a batch summary file was created, which brought together all the numerical outputs for easy and straightforward reporting in the Results section.

Intrinsic disorder and binding mode analysis using FuzPred

To strengthen the structural evaluation that was performed using AlphaFold, a thorough investigation of the intrinsic disorder was accomplished through the use of FuzPred (Hatos et al., 2023), an online computational tool that is made to measure the conformational properties of intrinsically disordered proteins (IDPs). FUS is known to be composed mainly of low-complexity and prion-like regions, which do not acquire any stable structure under physiologically relevant conditions, hence, it was very important to determine the intrinsic disorder landscape for each ALS-associated mutation.

FuzPred was performed separately for the WT sequence and for each of the mutant sequences (G156E, R234L, and R521C). For each amino acid the technique gives three essential probability descriptors: the chance of going through a disorder-to-order transition during interaction (pDO), the chance of remaining disordered during binding (pDD), and the binding mode multiplicity (MBM), which indicates the likelihood to explore different interaction geometries. These values were outputted in a tab-separated format and they were the basis for all subsequent processing.

A Python script was created to analyze and combine the FuzPred results and to calculate the changes caused by mutations relative to the profile of the WT. After the mutation-specific data at the residue level was imported, the script built a detailed and clear picture of the intrinsic disorder characteristics of WT FUS, thus establishing a baseline disorder map throughout the whole 526-residue sequence. For every mutant, Δ-profiles were made by subtracting the WT values from the corresponding mutant values, thus producing detailed representations of mutation-specific changes in pDO, pDD, and MBM. The study took a global view, but the local environment around each mutation site was focused on. Point mutations can have their most potent effects even in a very small-time frame, and this was the case here. To do this numerically, the script took the average of changes in disorder metrics in the area of less than 10 residues of the mutation, so it became possible to compare global perturbations and perturbations with locally confined regions directly. The method allowed pinpointing the mutations that contributed to the reshaping of the global disorder subtly and at the same time identifying those mutations that caused disruption of the conformational behavior in a pronounced but localized way.

To gain a broader perspective on FUS interaction tendencies, the ratio of residues surpassing standard empirical FuzPred thresholds was measured. Specifically, we calculated the fraction of residues with pDO ≥0.6, which indicates regions that likely adopt a folded structure upon binding, and MBM ≥0.65, which indicates regions highly sensitive to cellular context that are expected to sample multiple binding modes (Hatos et al., 2023). Using these standard threshold-based measurements gave a protein’s overall disorder characteristics as a global descriptor and at the same time it indicated whether the mutations had conspicuous effects on FUS’s wider interaction tendencies. The present analysis ensured systematic documentation of qualitative disorder patterns and their quantitative effects by saving all the plots, Δ-profiles, and summary tables for dissertation incorporation.

Docking workflow for assessing the impact of ALS-Associated FUS mutations on molecular interactions

In order to find out if mutations related to ALS have influenced the key molecular actions of FUS, a unified docking workflow was used for two independent binding events evaluation: (1) the recognition of FUS NLS by Transportin-1 (TNPO1) and (2) RNA binding facilitated by the RGG1-associated region. The same methodology was applied to all the structural preparation and docking steps to ensure proper the comparison of two systems, WT and mutant, from the methodological point of view.

The structures of WT FUS were fetched from the AlphaFold Protein Structure Database without any modifications. The model of full-length WT was used as the structural base for revealing the two functional regions of interest: the NLS (residues 514–526) and the N-terminal part of the RGG1-associated region (residues 140–180). Fragments were extracted using PyMOL, where custom scripts helped to keep only the wanted residues while getting rid of all solvent molecules, hydrogen atoms, and neighboring areas. A clean PDB file was made for each fragment, allowing for further docking.

Mutant peptides were made in PyMOL to design models that were structurally controlled and that had retained the native backbone conformation. The R521C mutation was done on NLS peptide with the help of the Mutagenesis Wizard, substituting Arg521 with cysteine. The G156E mutation was performed in the same manner within the 140–180 region but by swapping glycine at position 156 with glutamate. The structures of all peptides were cleaned, evaluated for steric clashes, and then saved as separate ligand files for docking.

Even though G156E does not occur in the short RGG1-derived docking fragment (residues 173–176), its involvement in RNA binding was of sufficient biological importance to justify its inclusion in this study. The mutation is located right before the RGG1 (166–267) domain which is usually considered the RGG1 domain, in a flexible region which is rich in glycine and is responsible for RNA recognition. Earlier research has indicated alterations in RNA binding due to mutations outside the core RGG motifs, where changes in charge distribution, backbone dynamics, or even the spatial positioning of adjacent RGG residues can play a role. Full-length modeling of the intrinsically disordered RGG1 domain is not doable within HADDOCK (Honorato et al., 2024) due to computational limitations, which is why a short RGG1-derived peptide (residues 130–190) containing the central RGG motif was used as a manageable substitute for RNA binding. Therefore, A reduced RGG1-derived peptide spanning residues 130–190 was used as a chemically representative docking fragment. The use of this reductionist yet chemically representative system enabled the direct comparison of the binding behavior of WT and G156E under controlled docking conditions.

The receptor structures for the two docking systems were prepared independently. For the NLS-TNPO1 investigation, Transportin-1 was derived from the FUS-TNPO1 complex (PDB: 4FQ3) which had been experimentally validated. Only the TNPO1 chain was kept, while all water and heteroatoms were excluded for compatibility with HADDOCK formatting requirements. For RNA-binding studies, the RNA ligand was obtained from the FUS-RNA complex (PDB: 6GBM). A Biopython-based script was utilized for isolating RNA residues 1–23, thus generating a stable, crystallographically defined RNA fragment for docking.

Docking simulations were carried out via the HADDOCK 2.4 web server, where the NLS–TNPO1 system was docked using the protein–peptide protocol, while the RGG1–RNA system utilized the protein–RNA protocol. Interaction sites were identified based on either validated experimental data or sequence-based predictions. In the case of the NLS system, active residues were those communicating with the 4FQ3 crystal structure. For the RGG1-derived system, the active residues were the 130–190 region (145–150, 152–159, and 173–176) including the positions of aromatic, polar, and charged amino acids, selected for their chemical similarity to RGG-mediated RNA recognition. The region with only glycine residues (160–172) was not included as it did not have any direct RNA-binding interactions. To ensure that all potential protein-RNA contacts were explored, all RNA residues (1–23) were defined as active. The rest of the settings used default HADDOCK parameters.

The top-ranked cluster for each system was determined after docking, using HADDOCK score and Z-score as criteria. This cluster was thought to contain the most energetically favored and statistically significant conformations. The best-scoring structure from each top cluster was downloaded and visualized in PyMOL. Structural alignment based on the receptor was used for the direct comparison of WT (WT) and mutant complexes. Method of the interface metrics such as van der Waals and electrostatic contributions, buried surface area, and hydrogen bonding networks, were used to assess mutation-mediated changes in the binding modes.

This integrated computational strategy not just allowed but also gave a systematic and quantitatively comparable evaluation of the interactions between the wild type and the mutant, for both nuclear import and RNA binding. The R234L mutation was not subjected to a docking workflow due to fundamental structural constraints associated with its location in the FUS protein. Residue 234 forms part of the RGG2 domain, which is contained within the intrinsically disordered C-terminal region of FUS. Unlike either the NLS or the RGG1-derived segment examined in the case of the R521C and G156E mutations, respectively, the RGG2 domain does not adopt a stable secondary or tertiary structure, nor does it engage in binding through a well-defined, localized interface. RNA interactions within the RGG domains occur through highly dynamic, multivalent, and “fuzzy” contacts distributed across long stretches of residues rather than through discrete binding pockets (Fuxreiter, 2018). The result is that no experimentally determined FUS–RNA complex involving the RGG2 domain exists, and no validated receptor or ligand can be assigned as a docking partner for residue 234. Standard docking platforms such as HADDOCK require well-defined structural templates and identifiable interaction sites, and therefore cannot be reliably applied to this intrinsically disordered region. For these reasons, R234L was excluded from the docking analysis, and its functional implications are instead interpreted here in light of literature-based evidence concerning IDR-mediated RNA binding and phase-separation behavior.

Early structural comparison

3D structure prediction and model confidence

AlphaFold provided three-dimensional structural models for the FUS protein in its WT and mutants G156E, R234L, and R521C. The pLDDT score that measures the confidence in the prediction was extracted for each model, which correspond to a per-residue local difference distance test (pLDDT). The pLDDT plot shows that the confidence scores fluctuate quite a lot along the protein sequence (Figure 1). Although some of the regions are extremely certain with pLDDT values above 80 predominantly in the C-terminal part (residues ∼280–370, ∼420–450, and the PY-NLS region ∼515–526), the main part of the N-terminal domain (residues ∼1–280) and a linker region in the middle (residues ∼370–420) are all classified as low confidence with pLDDT values nearly always below 50. This distribution is very indicative of the respective regions being natively disordered. The locations of the mutations G156E, R234L, and R521C fall into these low-confidence regions in the mutant proteins, but R521C is located within a region of higher confidence in the WT. A pLDDT-colored graphical representation of the WT structure is shown in Supplementary Figure S1, where the small folded domains are placed amidst the large disordered areas.

Global conformational differences

To measure the conformational alterations caused by mutations, the complete structural models of WT FUS and the three mutants (G156E, R234L, R521C) were compared. The comparison was done after masking the residues that had pLDDT confidence values below 70, which made sure that RMSD and solvent accessibility metrics were applied only to the structurally reliable parts of the models. By using this filtering method, extensive volumes of the inherently disordered regions were left out, which implies that just the most reliable backbone parts played a role in the quantitative data.

The global Cα RMSD values showed significant variation among the different mutations. The G156E and R234L mutants exhibited moderate deviations in the backbone compared to the WT, with the respective values being 10.62 Å and 5.48 Å. Importantly, because the structural alignment was anchored exclusively on high-confidence residues, and these folded domains (such as the RRM and Zinc finger) retain their internal secondary structure in the predictions, the observed global RMSD values do not indicate classical intra-domain misfolding or unravelling. Rather, these values predominantly capture changes in the relative spatial orientation and packing of the intact structured domains, which are driven by mutation-induced perturbations in the highly flexible linkers connecting them. On the other hand, the R521C mutation gave rise to the greatest RMSD of 15.36 Å. Local RMSD values for the different variants were not calculated since less than five high-confidence residues were found within the ±10-residue window around each mutation site. This further indicated that all three mutations are in the structurally uncertain regions, which is in line with the inherently disordered character of FUS.

In order to assess the effects of solvent exposure, the solvent accessible surface area (SASA) was calculated on a per-residue basis for both the wild type and the mutant structures, thus enabling the calculation of total ΔSASA and, subsequently, the hydrophobicity-weighted descriptor HYDRO·ΔSASA. The three mutants showed a net positive ΔSASA, which signified that the total solvent-exposed surface area for each change was greater than that of the wild type. Interestingly, the hydrophobic contribution to this exposure varied among the mutants. G156E and R521C were the ones with the largest total ΔSASA values (104.24 Å 2 and 108.44 Å 2, respectively) and the highest magnitude of HYDRO·ΔSASA (−647.98 and −570.73), while R234L had lower overall values in both metrics. Even so, the global GRAVY score stayed constant in all variants (−0.688), which means that none of the substitutions modified the protein mutants’ overall hydrophobicity profile. A highly negative HYDRO·ΔSASA value indicates that a mutation has exposed significant hydrophobic surface area to the aqueous solvent. While traditional 3D folding stability metrics are less applicable to intrinsically disordered proteins, this pronounced exposure represents a severe decrease in solvent stability. Consequently, the protein is thermodynamically driven to minimize this unfavorable state by burying these patches through aberrant intermolecular aggregation. This forced shift away from the protein’s native, dynamic state directly connects to its impaired function; by engaging in pathological solid-phase transitions, FUS loses its necessary flexibility to properly facilitate multivalent RNA binding and nuclear transport. A full summary of structural metrics is shown in Table 1 and the structural models in Figure 2. A full summary of structural metrics is shown in Table 1 and the structural models in Figure 2.

Intrinsic disorder analysis with FuzPred

In order to delineate the inherent disordered landscape of the FUS protein and to evaluate the effect of the mutations associated with ALS on the disorder tendencies either in the specific area or globally, the FuzPred algorithm was employed on the WT and all mutant sequences. FuzPred gives three prediction types per residue: the probability of a disorder-to-order transition (pDO), the probability of disorder-only behavior (pDD), and the multiplicity of binding modes (MBM), with the latter indicating the degree of interaction versatility or conformational heterogeneity during the binding process. The results were downloaded in TSV file format and the same processing steps were applied to them in order to create disorder profiles and visual plots of differences specific to mutations.

The very first step was to create a disorder profile for the WT FUS sequence. The output trace depicted the characteristic disorder profile of FUS, where large areas of high pDD and MBM were scattered over the low-complexity regions, which is in accordance with the protein being a dynamic RNA-binding factor with high interaction tendency. This WT profile acted as the starting point for the quantification of all the deviations caused by mutations. (Figure 3). FuzPred outputs for each mutant were aligned with worldwide sequence residue-by-residue, which permitted the calculation of ΔpDO, ΔpDD and ΔMBM values for the whole protein of 526 residues. The resulting difference profiles provided a clear picture of the local impact of each substitution. The three mutants showed distinctive, sharply defined global changes, although the changes were small in magnitude, and each was in the immediate surrounding area of the mutation site. These localized differences have been highlighted in the ΔFuzPred plots (Figure 4), where the perturbations are limited to very narrow margins around residues 156, 234 and 521 respectively.

For giving numerical context, FuzPred metrics were subjected to global and local average calculations. Global averages correspond to the mean values of ΔpDD and ΔMBM calculated throughout the entire sequence, while the local metrics measure mean values within a ±10-residue window surrounding the mutation. Every single mutant exhibited global deviations that were very small, thus global ΔpDD values nearly reached zero and global ΔMBM values also remained in a slightly broader range. On the other hand, the mutation-centered local averages indicated considerable larger values which confirmed the nature of disorder perturbations being high localized.

In the case of G156E mutation, the local ΔpDD was −0.0107 on average and the local ΔMBM showed an increase of +0.0737. The R234L substitution produced a local ΔpDD of −0.0083 and a local ΔMBM of +0.0411. The R521C mutation had the strongest localized impact with a local ΔpDD of −0.0102 and a local ΔMBM of +0.0532. These values are given in Table 2 along with the corresponding global averages. Furthermore, the whole-protein means of each predictor (mean pDO, mean pDD and mean MBM) were calculated for all proteins, alongside the fraction of residues surpassing the standard FuzPred empirical thresholds (pDO ≥0.6 and MBM ≥0.65). These threshold-based summaries indicated only slight global differences between WT and mutants, which again emphasized the mostly local effect of each substitution.

Docking analysis

Docking analysis of the FUS nuclear localization signal with Transportin-1

To determine whether the structural changes detected in Phase one have an impact on the interaction of FUS with its nuclear import receptor, the proteins were docked through HADDOCK 2.4 simulations (Honorato et al., 2024). One of the main factors for the analysis was the nuclear localization signal (NLS) consisting of the residues 514–526 which is responsible for the recognition of Transportin-1 (TNPO1). For this particular step, the complete full-length WT FUS structure was obtained directly from the AlphaFold Protein Structure Database (entry AF-P35637) rather than the AlphaFold models produced in Phase one to make sure that a validated reference structure was used. Subsequently, the same procedure for extraction and preparation was performed to derive both the WT and R521C mutant peptide structures for docking.

The docking procedure was performed on human Transportin-1 structure, employing the receptor conformation confirmed by experiment from PDB entry 4FQ3. In the scenario of both the WT and R521C peptides, the HADDOCK method presented a number of clusters of potential complexes, which were assessed by applying the standard scoring function used in HADDOCK. For all instances, the highest-ranking solution was that of Cluster 2, characterized by the lowest HADDOCK score and highest Z-score. The summary of the leading cluster from each simulation is presented in Table 3.

In contrast to the complete RMSD-based structural comparison, which failed to give any indication of functional binding differences, the docking study provided unambiguous quantitative differences between the WT and mutant complexes. The docking study indicated that FUS NLS bound Transportin-1 with a HADDOCK score of −87.8 ± 3.4 for the WT and the R521C mutant formed a complex with a score of −79.5 ± 3.6, which was slightly weaker. The two complexes showed similar structural convergence in their respective top clusters (0.3–0.6 Å RMSD) which proved to be well-defined binding modes. The study of energy breakdown showed that although the mutant complex had more significant van der Waals contribution, its electrostatic interaction was drastically lower than that of the WT peptide. A comparison of buried surface area values showed the R521C mutant to have a slightly larger contact interface (1388.7 ± 46.4 Å 2) than the WT (1174.1 ± 26.2 Å 2) and still get a less favorable docking score. These observations are visualized in Figures 5A,B, which show the overlaid docking poses of WT and R521C NLS peptides in the Transportin-1 binding groove as well in Figures 5C,D with a more closeup view with visible atomic interactions.

Docking analysis of the FUS RGG1-Derived peptide with RNA

To assess the impact of the G156E mutation on the RNA-binding ability of the RGG1-derived peptide, docking simulations were applied between the FUS amino acids 130–190 and a 23-nucleotide RNA fragment using HADDOCK 2.4. The identical extraction, preparation, and mutagenesis procedure from the NLS analysis was used here, ensuring methodological consistency between WT and mutant systems. In the case of the WT peptide, the HADDOCK software generated 12 clusters, and the first cluster was picked as the best solution according to the highest favorable HADDOCK score (−74.7 ± 1.6) and the lowest Z-score (−2.0) that reflects a very strong convergence and a highly reliable binding mode. The G156E peptide led to the formation of 9 clusters, with Cluster 1 once more being the highest-ranked solution, albeit with a less favorable HADDOCK score (−61.6 ± 5.7) and a not so negative Z-score (−1.6).

Energy decomposition showed that there were distinct differences in binding between WT and G156E. The van der Waals contributions for WT were the most favorable ones (−74.4 kcal/mol) and the electrostatic profile was the strongest one (−62.5 kcal/mol), which can be related to better packing and higher charge complementarity. On the contrary, the change caused by the G156E mutation brings a newly added negatively charged glutamate closer to the RNA phosphate backbone, thus raising the electrostatic repulsion and weakening the interaction. The mutant was also corroborated by interface analysis to have impaired RNA binding, and the G156E complex showed a significant reduction in buried surface area (1400.6 Å 2 versus 1735.6 Å 2 in WT), which is a sign of a smaller, and less stable, interface. The RMSD values were consistently low in every top cluster, which verified that the two simulations had converged reliably. These outcomes are visualized in Supplementary Figures S2, S3.

FUS mutations constitute one of the most fascinating molecular pathways for exploring the primary cause of amyotrophic lateral sclerosis, ALS, at the cellular level. FUS is a universal RNA-binding protein that requires a flexible structure rather than a rigid one to display its characteristic activities. The current study has investigated the effects of three clinically important mutations, G156E, R234L, and R521C, on FUS protein using a combination of various computational techniques such as AlphaFold-based structural modeling, solvent-accessibility and hydrophobicity metrics, FuzPred intrinsic disorder analysis, and two types of docking simulations (protein–protein and protein–RNA). Overall, the presented methods offered a comprehensive mechanistic description of how different mutations, situated at various spots of the FUS sequence, eventually lead to the same impact on RNA interactions, nuclear transport, and phase behavior, the latter being the pathological hallmarks that are almost always pointed out in ALS research.

A common result from all the studies is that minimal global structural disturbances are present. The pLDDT distributions and RMSD measurements of AlphaFold corroborated that the mutations do not bring about any large-scale conformational alterations. This is not at all surprising since FUS is a member of the family of intrinsically disordered proteins (IDPs), where transitions in structure are usually local, temporary, and dependent on the context rather than global or stable. The absence of global perturbation corresponds well with biochemical and NMR data which indicate that the N-terminal low-complexity (LC) region and all three RGG domains remain disordered even when exposed to conditions that favor the ordering of the proteins (Kato et al., 2012; Burke et al., 2015). This finding is of great significance as it allows for the functional effects of the mutations associated with ALS to be traced back to changes in local flexibility, binding propensities or electrostatic environments, rather than to misfolding in a classical structural sense.

Definitely, when paying attention to local structural effects, the findings showed that every mutation has a unique pattern of disruption that is very much related to its sequence context and the biological function that has been discovered. The ΔSASA and HYDRO·ΔSASA measurements—though small in absolute value—unveiled specific modulations of solvent exposure and hydrophobic contributions in particular regions. These alterations are indicative of local interaction perturbations instead of large conformational changes. The case of glycine-to-glutamate substitution at position 156 (G156E) is an example, where the charged and bulkier side chain is introduced into a very flexible RGG1-adjacent region and thus, solvent exposure is increased and hydrophobic interactions altered, all of these changes becoming measurable. Past investigations of LC domains have indicated that even single charged substitutions can completely change multivalent interactions, lower condensate stability or cause phage transitions that are not going to be expected (Qamar et al., 2018; Patel et al., 2015). Therefore, the physical behavior that is observed in the current study is perfectly aligned with the previously known models of LC-driven RNA granule dynamics.

The conclusion is further corroborated by the intrinsic disorder analysis conducted using FuzPred. The probability of global disorder-to-order transition (pDO) remained slightly the same, however, the mutations of all three types caused local disturbances that were very clearly visible within the range of ten residues on either side of the mutation site. The things that happened were: the drops in pDD and the rises in MBM (binding-mode multiplicity), suggesting a loss of the protein’s native, highly flexible ‘fuzzy’ state (lower pDD). This indicates a transition toward aberrant and uncoordinated interaction behavior, where the protein improperly explores multiple interaction geometries (higher MBM) without maintaining its functional disorder. IDPs depend on the sequences that are able to sample the ensembles of states and at the same time, they become locally more ordered on binding; thus, when this delicate balance is disrupted, the protein either fails to form productive complexes or it takes part in undesired interactions. The behavior shown for G156E and R521C perfectly matches up with the experimental studies which have revealed that ALS-associated charge-altering mutations reduce interaction specificity, alter phase behavior, and promote pathological aggregation (Wang et al., 2018; Hofweber and Dormann, 2019). Even R234L, which puts a hydrophobic leucine into the positively charged RGG2 motif, showed local pDD reductions. This indicates a loss of its healthy, dynamic disordered state, which would normally be expected from the removal of an arginine critical for the electrostatic interactions with RNA.

The outcomes of docking studies give direct structural-functional elucidations of the changes in disorder. For the R521C nuclear localization signal (NLS) mutation, the protein-protein docking with Transportin-1 (TNPO1) indicated a significant reduction in the nuclear import capacity. The WT PY-NLS peptide accommodated the typical extended pose in the TNPO1 groove, agreeing with the crystallographic data from Zhang and Chook (2012). However, the R521C mutant showed even worse energetic scores, less burial of interface area, and weakened electrostatic interactions-precisely corresponding to the cellular and patient biochemical phenotype, where R521C dramatically reduces TNPO1 affinity causing cytoplasmic mislocalization of FUS (Dormann et al., 2010; Higelin et al., 2016). FUS cytoplasmic accumulation is a major pathogenic factor that drives the cell to undergo toxic phase transitions, irregularly incorporate into stress granules, and cause the loss of essential RNAs and proteins (Taylor et al., 2016). While previous molecular dynamics studies have established that the R521C mutation disrupts FUS-Transportin-1 binding (Swetha et al., 2017), our integrative computational pipeline adds novel insights by quantifying the specific energetic penalties—namely, the dramatic loss of electrostatic favorability—while simultaneously mapping these local binding deficits to sequence-wide alterations in intrinsic disorder. Thus, the docking analysis not only repeats the already known disease mechanisms but also gives atomic-level explanations for the disruptive effect of a single cysteine substitution on nuclear import.

Another key aspect of FUS biology is the cross-talk with RNA, especially through the LC/RGG regions. The protein-RNA docking study for G156E indeed focused on this functional aspect. The docking of the WT 140–180 fragment to RNA showed the strong binding that was anticipated from the RGG1-RNA interactions, which are known to be multivalent. The top cluster of the WT showed good van der Waals and electrostatic energies along with a large buried surface area of ∼1736 Å 2, which is not far from the experimentally defined multivalent contact patterns in FUS-RNA complexes (Loughlin et al., 2019). In contrast, the G156E mutant showed a significant drop in binding strength, less densely packed interface and a much smaller buried surface. These results obviously have pathogenic importance: the FUS LC and RGG domains are necessary for RNA-dependent phase separation, self-regulation of FUS localization, and formation of ribonucleoprotein granules (Patel et al., 2015). If these interactions are weakened as in the case of G156E, it may lead to the destabilization of normal condensates, promotion of misregulation of RNA metabolism, and eventually contribute to the formation of cytoplasmic aggregates. The docking data thus lend support to the widely held view that the mutations associated with ALS disrupt the fine balance between the sequence-encoded disorder, multivalent RNA interactions, and liquid-liquid phase separation.

R234L mutation, although not extensively featured in experimental studies like R521C, harmoniously fits the scenario of mechanism. The RGG domains have arginines that are mainly involved in the RNA recognition over a range of both electrostatic attraction and cation–π interactions (Wang et al., 2018). Their removal leads to a huge drop in RNA recognition and a consequent reversal of phase (Hofweber and Dormann, 2019). The structural docking is not useful for the study of intrinsically disordered regions (Wright and Dyson, 2015), but the disorder evaluation in this thesis showed the pDD to be reduced, indicating a decrease in native, functional flexibility around R234L; this loss of healthy disorder is exactly in line with the lower RNA-binding capability. Thus, even in the absence of a fixed docking complex, the silico data have an excellent justification for why R234L has been implicated in the disruption of RNA metabolism and the uncoupling of condensate behavior.

The outcomes of this research, taken together, point to the similarity of pathogenic effects among the mutations that are situated in very different FUS domains. Every mutation prunes a unique functional modality: R521C has the most influence on nuclear import, G156E makes RNA binding weak and changes LC-driven disorder, and R234L has the least interaction in the RGG2 region. However, all of them together compromise FUS’s capacity to keep its regular distribution between the nucleus and the cytoplasm, RNA binding, and phase behavior—these three properties are known to be indispensable for neuron survival. If any of these three properties get disturbed, then the normal situation of cytoplasmic retention, stress granule association, and pathological aggregation happening will get more frequent and these processes have been repeatedly related to ALS (Sharma et al., 2016).

We acknowledge several limitations associated with the use of computationally predicted structural models. AlphaFold, which was developed to predict protein folding and three-dimensional protein structures, generally shows reduced predictive accuracy for intrinsically disordered proteins (IDPs), intrinsically disordered regions (IDRs) and highly flexible loop regions, as these segments do not adopt a single stable three-dimensional structure (Krokidis et al., 2024). Furthermore, although AlphaFold2 predicts protein structures from amino acid sequences with high accuracy, it cannot fully capture protein dynamics and conformational flexibility (Jumper et al., 2021). Therefore, structural differences inferred from AlphaFold-derived models, particularly within IDRs, should be interpreted as plausible conformational hypotheses rather than unique structural states. Although IDPs and IDRs lack well-defined conformations, they play essential roles in numerous biological processes, including those associated with neurodegenerative diseases. Consequently, structural differences inferred from AlphaFold models within these regions should be interpreted cautiously and regarded as plausible conformational hypotheses rather than unique or experimentally validated structural states. In addition, AlphaFold-derived structures have limitations in applications such as reverse docking, where structural uncertainty in flexible regions may affect prediction accuracy (Wong et al., 2022). Complementary deep learning-based disorder prediction methods can improve the identification and annotation of IDRs and provide a more reliable characterization of protein disorder (Yang et al., 2023). Future methodological advances are expected to improve the prediction of protein–protein interactions involving proteins implicated in neurodegenerative disorders, enhance the modeling of intrinsically disordered and membrane proteins, and facilitate the integration of experimental structural data from X-ray crystallography, nuclear magnetic resonance and cryo-electron microscopy (Maiti et al., 2024). Accordingly, the computational analyses presented here should be viewed as hypothesis-generating and complementary to experimental approaches rather than as definitive structural evidence.

The current study, by the combination of structural models, disorder metrics, and docking simulations, provides a clear mechanistic framework that connects single amino acid changes to broader cellular dysfunction. Moreover, the results highlight the value of using disorder-recognizing computational techniques in the research of diseases related to IDPs. Traditional structure-based modeling can deal with disordered proteins but cannot uncover their fine features, and tools like FuzPred and flexible docking methods allow a better and more precise portrayal of these proteins’ interaction landscapes. This combined methodology which was employed with FUS can be taken to other disordered ALS proteins like TDP-43 or hnRNPA1, thus providing a robust avenue for unraveling the underlying mechanisms and in developing novel therapies.

In summary, this study provides an atomistic perspective on the convergence of different mutations linked to ALS in FUS that leads to a disruption of nuclear transport, RNA interactions, and a disordered functional mechanism. The findings corroborate, and even expand upon, the current body of literature, thus offering a unified computational interpretation of mutation-induced dysfunction in FUS-related ALS as a whole. One of the major findings arising from this research is that mutations in FUS causing ALS can be generally understood in terms of local rearrangements in disorder/molecular recognition interactions as opposed to global misfolding. While R521C primarily affects frontiers in PY-NLS recognition interactions in agreement with reduced transportin-1-mediated transport and subsequent cytoplasmic accumulation, G156E primarily affects those in RNA interactions by shifting local binding plasticity properties and lowering binding stability/weight for equal docking conditions as a peptide-RNA interface. R234L, found in a strongly disordered RGG2 environment, does not exhibit classical docking properties, but rather a distinct loss of native disorder and binding plasticity. This structural rigidification is in perfect agreement with reduced multivalency in RNA interactions as found in RGG-rich domains.

While the present study gave a comparative, structure, and disorder-based framework for interpreting ALS-associated FUS mutations, many avenues remain to be pursued in the future. In this context, a logical next step would represent using ensemble-based methodologies that better capture the conformational heterogeneity of IDRs beyond single-state static structural representations. Molecular dynamics simulations of selected systems, in particular the PY-NLS-Transportin-1 complex, may help provide more detailed insights into the mutation-dependent changes in interaction stability and recognition dynamics on longer time scales. Another complementary direction involves further extending the RNA-binding study from a single to multiple RNA sequence contexts and longer RNA fragments to address whether the observed trends are general across different binding motifs and multivalent interaction scenarios. Finally, integration with computational predictions based on available experimental results, such as NMR, cross-linking, or cellular localization assays reported in the literature, would add real value to the validation and help bridge the gap between in silico observation and functional consequence. All these directions will be crucial to advance the mechanistic understanding of how sequence-encoded disorder and binding plasticity underpin FUS-mediated neurodegenerative disease.