Predicting molecular recognition features in protein sequences with MoRFchibi 2.0 Authors: Nawar Malhis, Jörg Gsponer Published: 2026 ABSTRACT Molecular Recognition Features (MoRFs) are segments within disordered protein regions (IDRs) that undergo a disorder-to-order transition upon binding to their partners. Identifying MoRFs remains a significant challenge. This paper introduces MoRFchibi 2.0, a specialized prediction tool designed to identify the locations of MoRFs within protein sequences. Our results show that MoRFchibi 2.0 outperforms all existing MoRF and general predictors of protein-binding sites within IDRs, including the top-performing models from the Critical Assessment of protein Intrinsic Disorder (CAID) rounds 1, 2, and 3. Remarkably, MoRFchibi 2.0 surpasses predictors that utilize AlphaFold data and state-of-the-art protein language models, achieving superior ROC and Precision-Recall curves and higher success rates. MoRFchibi 2.0 generates output scores using an ensemble of logistic regression convolutional neural network models normalized for the priors in the training data, making them individually interpretable and compatible with other tools utilizing the same scoring framework. Introduction Intrinsically disordered regions (IDRs) in proteins are segments that lack a unique tertiary structure in isolation under physiological conditions [1, 2]. Despite this, IDRs are functionally significant and play key roles in numerous cellular processes, often by interacting with other proteins, DNA, RNA, or small molecules [3–6]. Their dynamic nature allows IDRs to mold themselves around their interaction partners [7–9]. As of the latest release of the DisProt database [10] in June 2024, 641 out of 3,022 proteins annotated with IDRs include one or more experimentally validated binding segments. However, this is a small number, especially considering that approximately 32% of the human proteome consists of proteins with more than 30% disordered residues [11]. Consequently, the computational identification of binding regions within IDRs has emerged as a significant challenge. Of particular interest are regions within IDRs that bind to other proteins, and numerous computational tools have been developed to aid in identifying these protein-binding segments [12–52]. A subfamily of protein-binding segments in IDRs is known as molecular recognition features (MoRFs) [53–57]. MoRFs fold upon binding, i.e. they transition from a disordered to an ordered state upon interacting with their partners. Typically, MoRFs are short, though they can sometimes span 50 or more residues. The functional significance of MoRFs, along with the challenges posed by their experimental identification and characterization, has led to the development of several computational tools that predict the locations of MoRFs in IDRs. Notable examples include Retro-MoRFs [15], MoRFpred [20], fMoRFpred [25], MoRFchibi_light [27], MoRFchibi_web [23], OPAL [33], MoRFPred-plus [31], MoRFcnn [46]. MoRF_ESM [49], MoRFMPM [52], and Trans-MoRF [51]. Since MoRFs represent only a small fraction of all residues in protein sequences, i.e., have low “prior” probabilities, high-quality predictors are required to achieve acceptable posterior probabilities [58]. The MoRFchibi predictors, including light and web, are among the most accurate MoRF predictors available today. They are ranked among the leading IDR-protein-binding predictors in the first two rounds of the Critical Assessment of protein Intrinsic Disorder (CAID) competition [59, 60] and are available in the CAID Portal set of predictors [61]. However, prediction accuracy depends heavily on the quality and size of the training datasets, and MoRFchibi predictors were developed using training data collected in 2008 [20]. Since then, several databases have been created that contain curated and annotated IDR protein-binding segments, including MoRFs. The availability of more extensive and accurate training data, along with recent improvements in machine learning models, enabled us to redesign MoRFchibi and improve its predictive performance. Thus, we introduce MoRFchibi 2.0, a redesigned MoRF predictor that encompasses an ensemble of four logistic regression convolutional neural network (LR-CNN) models. We also provide newly assembled datasets with MoRF annotations for the community to optimize and test other models. Methods Datasets We assembled a dataset of protein sequences annotated with disordered regions that undergo disorder-to-order transitions upon binding to other proteins, drawing from four manually curated databases: DisProt [10], DIBS [62], MFIB [63], and IDEAL [64]. Two key factors influencing the quality of MoRF annotations are: (1) the availability of experimental evidence confirming that a sequence segment is disordered before binding, and (2) evidence demonstrating that it folds upon binding to a protein target. The degree to which these essential details are documented varies among the four databases. DisProt, a database for disordered protein sequences, provides disorder annotations and details on IDR properties, including protein and nucleotide binding, as well as disorder-to-order (DtO) transitions. DIBS and IDEAL are databases that focus on interactions between binding segments in IDRs and their ordered protein partners. In these databases, the disordered state of binding segments is supported either by experimental data or inferred from homology. While all interacting IDR segments in IDEAL undergo a disorder-to-order transition upon binding, a small fraction of these segments in DIBS do not fold upon binding. The MFIB database, on the other hand, includes IDRs that bind to other IDRs and fold. The majority of interactions in MFIB are supported by experimental evidence, while some are based on homology. Using the DisProt release (24_06), we annotated all disordered protein-binding residues that overlap with DtO transitions as MoRF residues; Since DtO annotations are likely incomplete, we masked out annotated protein-binding residues not associated with DtO, treating them neither as MoRFs nor as non-MoRFs. All other residues in DisProt are considered non-MoRF residues. We included IDEAL and DIBS binding segments validated by experimental data and restricted the DIBS segments to those associated with DtO annotations in DisProt. Additionally, we incorporated all MFIB binding annotations. This process yielded 899 high-quality MoRF segments (HQ MoRFs) across 769 sequences, totalling 46,694 residues. Alongside these HQ MoRF annotations, we included low-quality MoRF annotations (LQ MoRFs) for training. These LQ MoRFs are binding segments from IDEAL and DIBS, supported by evidence from homology, and DIBS binding segments not corroborated by DisProt DtO annotations. This approach yielded 1217 MoRF segments (HQ and LQ) across 1030 sequences, totalling 52,560 residues. Thus, 769 sequences are annotated by HQ MoRFs and possibly LQ MoRFs, and the remaining 261 sequences are annotated by LQ MoRFs only. A high degree of redundancy in the testing data can lead to biased evaluations, and in the training data, it can lead the optimization process to overweight these redundant sequences at the expense of others. Thus, we used CD-Hit [65] to cluster the identified sequences at an 80% identity. To avoid unnecessarily losing some training MoRF annotations, within each cluster, MoRF annotations for sequences with > 95% identity to the cluster centers, or > 92% identity and equal length to the cluster centers, are transferred to the cluster centers as LQ annotations, i.e., to be used for training only. Then, we used only cluster centers, resulting in 907 sequences with 1,081 MoRFs, totalling 46,504 HQ and LQ MoRF residues. This included 674 sequences with 790 HQ MoRF segments, totalling 40,657 MoRF residues. As sequences sourced from the DisProt database are likely to be more completely annotated than those from IDEAL, DIBS, and MFIB, the non-MoRF residues in DisProt sequences are less likely to include false negatives. Thus, we masked out non-MoRF residues in sequences not sourced from the DisProt. We divided the total of 907 MoRF sequences into five subsets, ensuring that sequences in each subset have less than 40% identity to those in other subsets. For that, we clustered the sequences at 40% identity using CD-Hit. We sorted these clusters in descending order based on the total number of HQ MoRF residues in each cluster. Then, starting with five empty subsets, we iteratively moved clusters to these subsets, such that at each iteration we moved the cluster with the highest number of HQ MoRF residues to the subset with the fewest total HQ MoRF residues. We used four subsets containing both LQ & HQ MoRF annotations to train four cross-validated LR-CNN models and the remaining fifth subset for testing. We further filtered the test dataset to remove sequences with an identity greater than 30% relative to the four training datasets, resulting in a test dataset with 143 sequences (Test-143). Although sequences in the test dataset are less than 30% identical to those used in the training set, short test regions can still exhibit local homology with training sequences. Thus, we used HAM [66] to identify and mask out regions with 10 or more residues in the test sequences that are homologs (with an identity greater than 80%) to the training sequences. HAM identified three regions (totalling 33 residues) in the testing dataset as homologs of the training data, one of which is a homolog of a masked training region of 11 residues. Therefore, HAM only masked the remaining 22 residues. Thus, the test set (Test-143) contains 143 sequences with 148 MoRFs, totalling 7,757 MoRF residues. To avoid potential bias to the negative class by masking ambiguous residues and limiting the negatives to “hard” negatives, we also created Test-143 (unmasked), for which all previously masked residues, except those masked by HAM, were unmasked. In addition, we created the Test-128 (AF2) subset, which is limited to Test-143 sequences with AlphaFold2 structures in the AlphaFoldDB database [67]. We also created the Test-106 (SR) subset, which only contains Test-143 sequences with both MoRF and non-MoRF residues. Specifically, we excluded all sequences from Test-143 that do not have at least five residues in each class, i.e., all sequences must have five or more MoRF residues and five or more non-MoRF residues. Moreover, as SR (success rate) is not sensitive to under-annotated sequences (Evaluation measures Section) and to avoid significantly reducing the test data size, we unmasked non-MoRF sequences outside DisProt. Finally, we assembled the Test-new-9 dataset from the new sequences in DisProt 2025_12 relative to 2024_06. MoRF sites are disordered protein-binding residues that overlap with DtO transitions; we masked out protein-binding residues not associated with DtO. Table 1 provides details on all four training datasets, the Test-143 and its three Test-143 variations, and the Test-new-9 datasets. Table 1Training and testing datasets: Four subsets, cv1, cv2, cv3, and cv4, utilize LQ annotations for training and HQ annotations for validationTraining (CV1)MoRFs qualitySequences withMoRFs/TotalMoRFsNon-MoRF ResiduesMoRF ResiduesMasked ResiduesLQ & HQ141/14117228,8128,90247,440Training (CV2)LQ & HQ182/18220857,2129,26743,553Training (CV3)LQ & HQ192/19223251,2319,34745,729Training (CV4)LQ & HQ198/19823752,9089,53258,255Test-143HQ126/14314846,7447,75720,037Test-143 (unmasked)HQ126/14314866,7597,75722Test-128 (AF2)HQ113/12813233,9237,24616,449Test-106 (SR)HQ106/10612849,6765,8231,163Test-new-9HQ9/9113,686387245The Test-143, its variations, and the Test-new-9 have only HQ annotations for testing Table 1Training and testing datasets: Four subsets, cv1, cv2, cv3, and cv4, utilize LQ annotations for training and HQ annotations for validationTraining (CV1)MoRFs qualitySequences withMoRFs/TotalMoRFsNon-MoRF ResiduesMoRF ResiduesMasked ResiduesLQ & HQ141/14117228,8128,90247,440Training (CV2)LQ & HQ182/18220857,2129,26743,553Training (CV3)LQ & HQ192/19223251,2319,34745,729Training (CV4)LQ & HQ198/19823752,9089,53258,255Test-143HQ126/14314846,7447,75720,037Test-143 (unmasked)HQ126/14314866,7597,75722Test-128 (AF2)HQ113/12813233,9237,24616,449Test-106 (SR)HQ106/10612849,6765,8231,163Test-new-9HQ9/9113,686387245The Test-143, its variations, and the Test-new-9 have only HQ annotations for testing Training and testing datasets: Four subsets, cv1, cv2, cv3, and cv4, utilize LQ annotations for training and HQ annotations for validation The Test-143, its variations, and the Test-new-9 have only HQ annotations for testing Input features We used the seven features introduced in [47] to encode amino acid sequences. The first two features estimate the energy associated with intrachain interactions of each residue by using the previously derived 20 × 20 energy predictor matrix [68]. Specifically, the first interaction energy feature is the sum of the pairwise energies between the residue at position k and all residues within a window of w1 = 10 on either side of k. The second interaction energy feature is the sum of this energy for all residue pairs between w1 and w2 = 25 residues on either side of k. The additional five features were obtained from the F5 amino acid enrichment matrix. F5 is a 20 × 5 matrix encoding the enrichment of each amino acid within IDRs, linker regions, IDR nucleic binding sites, IDR protein binding sites, and folded domains, as described in [47]. Model selection and optimization We utilized the ensemble window-in/window-out architecture originally introduced in IPA [47], which consists of four logistic regression convolutional neural network (LR-CNN) models (Fig. 1). Each LR-CNN comprises three convolutional layers with a kernel size of seven, separated by two average pooling layers (also with a kernel size of seven), followed by two fully connected (FC) layers. The output size of the first FC layer is set to 150. The input features a sliding window of size 10, flanked by 250 residues on each side, for a total of 510 window-in residues. When an input sequence lacks sufficient flanking residues, we pad the sequence with blank residues—hypothetical residues with all features set to zero. Each of the four LR-CNN models was trained on three cross-validation (CV) subsets, with the remaining fourth subset reserved for validation (Table 1). All architectural and input hyperparameters were selected via a grid search to minimize loss on the validation sets. We employed binary cross-entropy with logits (BCEWithLogitsLoss) as the training loss and applied a sigmoid function to the evaluation output of each model. To factor out training priors—specifically, the imbalances in MoRF residue percentages across the training subsets—we applied reverse Bayes to each model output as introduced in [47]. The four resulting scores were averaged, and a visually selected, equally weighted smoothing window of size 5 was applied to generate the final prediction score. Because independent training runs converged to distinct local minima with varying validation performance, we implemented a selection heuristic to mitigate the risk of choosing a poor local minimum. Specifically, we trained 15 independent instances of each of the four LR-CNN models and selected the instance with the third-best validation Area Under the Curve (AUC). This approach ensures the selection of robust optimization points while avoiding overfitting the validation data. Ultimately, two crucial design choices underpin this architecture: Large Receptive Field: We use a substantial flanking size of 250 residues on each side to provide the model with a broad receptive field, allowing it to incorporate a larger structural environment during evaluation.Ensemble Strategy: We employ an ensemble of four LR-CNNs trained on overlapping datasets to improve prediction consistency and mitigate potential biases arising from our limited training data size. Large Receptive Field: We use a substantial flanking size of 250 residues on each side to provide the model with a broad receptive field, allowing it to incorporate a larger structural environment during evaluation. Large Receptive Field: We use a substantial flanking size of 250 residues on each side to provide the model with a broad receptive field, allowing it to incorporate a larger structural environment during evaluation. Ensemble Strategy: We employ an ensemble of four LR-CNNs trained on overlapping datasets to improve prediction consistency and mitigate potential biases arising from our limited training data size. Ensemble Strategy: We employ an ensemble of four LR-CNNs trained on overlapping datasets to improve prediction consistency and mitigate potential biases arising from our limited training data size. Fig. 1Overall architecture of one CNN model that was used here, with results of convolutional kernels in blue, average pooling kernels in green, and the fully connected layers in red. Four of these CNN models were used in the ensemble predictor Overall architecture of one CNN model that was used here, with results of convolutional kernels in blue, average pooling kernels in green, and the fully connected layers in red. Four of these CNN models were used in the ensemble predictor Evaluation measures We calculated Receiver Operating Characteristic (ROC) curves to evaluate the performance of tools independent of the binary threshold cut-off values. ROC curves chart the relationship between the percentage of the positive class that passes any given score (true-positive rate, TPR) and the percentage of the negative class that passes that score (false-positive rate, FPR) across all possible classification thresholds. The formal definitions of TPR and FPR are: \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:TPR=\frac{TP}{TP+FN} \& \:FPR=\frac{FP}{FP+TN}$$\end{document} TP, FP, TN, and FN represent the counts of True Positives, False Positives, True Negatives, and False Negatives, respectively. Since both TPR and FPR are normalized: TPR by the total number of actual positives and FPR by the total number of actual negatives, ROC curves are insensitive to class imbalance. We use the area under the ROC curve (AUC) to evaluate a predictor’s overall performance in a single value. The diagonal line is the ROC curve for a random (naïve) classifier. The AUC for a random classifier is 0.5, and the better the classifier is, the higher its AUC value will be. Since ROC curves are insensitive to dataset imbalance, so are AUC values. For a randomly selected positive and a randomly selected negative class instance, the AUC is the probability that the positive instance is ranked higher than the negative one [69]. We used the Precision-Recall (PR) curves to chart the relationship between precision and recall, where: \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:Precision=\frac{TP}{\left(TP+FP\right)}\:\&\:Recall=\frac{TP}{\left(TP+FN\right)}$$\end{document} We included the average precision score (APS) [72] for each MoRF predictor in the legends of the PR curves. We also included a horizontal line with precision equal to the percentage of MoRF residues in the test data to represent the precision for a random classifier. Note that, unlike the ROC curves, the Precision values, and thus the PR curves and the APS, are highly sensitive to class imbalance. We also used the success rate (SR) to evaluate all predictors. SR measures the percentage of protein sequences with the average score of the positive class residues in the sequence (MoRF residues) higher than the average score of the negative class (non-MoRF residues). SR scores are essential as they are insensitive to under-annotation in the data and factor out the contribution of protein-level features that adjust the scores of residues for entire proteins based on the overall protein activity. Using protein-level features, such as protein-protein interaction data, can boost the AUC value of predictors by up-scoring entire sequences of hub proteins. However, outputs from predictors that rely on protein-level features are less precise at identifying the locations of MoRF residues within each sequence. Predictors with higher SR scores are more suitable for determining the position of the MoRF(s) in a protein sequence. Results Prediction assessment We evaluated MoRFchibi 2.0 (MC2) against both MoRF-specific predictors and general predictors of protein-binding regions within intrinsically disordered regions (IDRs). Using the Test-143 dataset, we compared MC2 against the original MoRFchibi (web and light versions) and other established predictors with accessible software or HTML servers, fMoRFpred, ANCHOR-2, and OPAL. Furthermore, we extended the comparison to twelve additional methods designed to identify protein-binding segments in IDRs, regardless of whether they undergo Disorder-to-Order transitions upon binding. As shown in Fig. 2; Table 2, and Table 3, MoRFchibi 2.0 outperforms all 17 tested predictors in both Area Under the ROC Curve (AUC) and Average Precision Score (APS) across Test-143 and its variations. Comparison with AlphaFold-2 Based Methods: Recent studies suggest that AlphaFold-2 often assigns higher pLDDT scores to MoRF residues within IDRs, potentially allowing it to identify MoRFs with high precision. Consequently, we included two AlphaFold-2-reliant predictors—AlphaFold-Bind and IPA-AF2_protein—which leverage pLDDT scores for their predictions. The results in Tables 2 and 3 demonstrate that MoRFchibi 2.0 consistently outperforms these AlphaFold-2-based methods as well. Success Rate Analysis: We also assessed the predictors using the Test-106 (SR) dataset. MoRFchibi 2.0 achieved the highest AUC and Success Rate (SR) among all tested methods. While DisoFLAG-PB (indicated by the dotted black line in Fig. 2) yields an AUC nearly equal to MC2, a closer inspection of the ROC curve reveals a performance trade-off: DisoFLAG-PB achieves its AUC primarily through superior identification of non-MoRF residues (top right corner of the ROC curve), while MoRFchibi 2.0 exhibits superior performance in identifying actual MoRF residues (lower left corner of the ROC curve), making it more effective for MoRF discovery. Fig. 2ROC and PR curves generated with predictions: Top) MoRFchibi 2.0 performance (left: ROC and right: Precision-Recall) against the Test-143 dataset compared to a set of MoRF predictors: OPAL, MoRFchibi_web, MoRFchibi_light, ANCHOR-2, and fMoRFpred, and twelve general predictors of protein-binding segments in IDR (in gray lines). Bottom) The performance of these predictors against the SR subset, where the general binding predictor DisoFLAG-PB is in the dotted black line to highlight MoRFchibi 2.0 higher performance at the most critical lower left corner of the ROC curve and the left of the PR curve. In the PR curves (Top and Bottom), to improve figure clarity, we excluded points with recall values below 0.05. The dotted gray lines represent the performance of a naïve predictor ROC and PR curves generated with predictions: Top) MoRFchibi 2.0 performance (left: ROC and right: Precision-Recall) against the Test-143 dataset compared to a set of MoRF predictors: OPAL, MoRFchibi_web, MoRFchibi_light, ANCHOR-2, and fMoRFpred, and twelve general predictors of protein-binding segments in IDR (in gray lines). Bottom) The performance of these predictors against the SR subset, where the general binding predictor DisoFLAG-PB is in the dotted black line to highlight MoRFchibi 2.0 higher performance at the most critical lower left corner of the ROC curve and the left of the PR curve. In the PR curves (Top and Bottom), to improve figure clarity, we excluded points with recall values below 0.05. The dotted gray lines represent the performance of a naïve predictor Table 2Comparison of the prediction AUC and Success Rate performances against the Test-143 dataset and its variations: The AUC values for MoRFchibi 2.0 and other predictors using the complete test dataset, Test-143, the complete unmasked dataset, Test-143(unmask), the subset of the test dataset with AlphaFold structures, Test-128 (AF2), and the subset of sequences with at least 5 MoRF and 5 non-MoRF residues, Test-106 (SR)MeasureAUCAUCAUCAUCSuccess RateTest DataCompleteUnmaskedSubset (AF2)Subset (SR)Subset (SR)MoRFchibi 2.0 0.813 0.803 0.806 0.745 0.736 OPAL 0.759 0.752 0.744 0.683 0.660 MoRFchibi_web 0.703 0.702 0.710 0.653 0.717 MoRFchibi_light 0.698 0.694 0.700 0.636 0.689 fMoRFpred 0.597 0.591 0.597 0.579 0.670 ANCHOR2 0.575 0.556 0.566 0.600 0.613 IPA_all0.7500.7220.7290.7030.708IPA_protein0.7410.7160.7200.7050.717DisoFLAG-PB0.7270.7150.7260.7410.726ProBiPred_protein0.6750.6760.6730.6270.575AIUPred-Binding0.6290.6100.6230.6450.632ENSHROUD_all0.6280.6060.6450.6440.632ENSHROUD_protein0.6010.5770.6140.6170.642DeepDISObind_all0.6080.5910.5910.5880.557DeepDISObind_protein0.5980.5820.5810.5810.632bindEmbed21IDR_idrGeneral0.6090.6180.6040.5760.481bindEmbed21IDR_rawGeneral0.6010.6110.5950.5690.377DisoRDPbind0.5800.5630.5780.6090.670AlphaFold_binding0.657IPA-AF2_protein0.712We also used the Test-106 (SR) subset to compare the success rates of these predictors. MoRF predictors are in bold, and the highest AUC and Success Rate values are underlined Table 2Comparison of the prediction AUC and Success Rate performances against the Test-143 dataset and its variations: The AUC values for MoRFchibi 2.0 and other predictors using the complete test dataset, Test-143, the complete unmasked dataset, Test-143(unmask), the subset of the test dataset with AlphaFold structures, Test-128 (AF2), and the subset of sequences with at least 5 MoRF and 5 non-MoRF residues, Test-106 (SR)MeasureAUCAUCAUCAUCSuccess RateTest DataCompleteUnmaskedSubset (AF2)Subset (SR)Subset (SR)MoRFchibi 2.0 0.813 0.803 0.806 0.745 0.736 OPAL 0.759 0.752 0.744 0.683 0.660 MoRFchibi_web 0.703 0.702 0.710 0.653 0.717 MoRFchibi_light 0.698 0.694 0.700 0.636 0.689 fMoRFpred 0.597 0.591 0.597 0.579 0.670 ANCHOR2 0.575 0.556 0.566 0.600 0.613 IPA_all0.7500.7220.7290.7030.708IPA_protein0.7410.7160.7200.7050.717DisoFLAG-PB0.7270.7150.7260.7410.726ProBiPred_protein0.6750.6760.6730.6270.575AIUPred-Binding0.6290.6100.6230.6450.632ENSHROUD_all0.6280.6060.6450.6440.632ENSHROUD_protein0.6010.5770.6140.6170.642DeepDISObind_all0.6080.5910.5910.5880.557DeepDISObind_protein0.5980.5820.5810.5810.632bindEmbed21IDR_idrGeneral0.6090.6180.6040.5760.481bindEmbed21IDR_rawGeneral0.6010.6110.5950.5690.377DisoRDPbind0.5800.5630.5780.6090.670AlphaFold_binding0.657IPA-AF2_protein0.712We also used the Test-106 (SR) subset to compare the success rates of these predictors. MoRF predictors are in bold, and the highest AUC and Success Rate values are underlined Comparison of the prediction AUC and Success Rate performances against the Test-143 dataset and its variations: The AUC values for MoRFchibi 2.0 and other predictors using the complete test dataset, Test-143, the complete unmasked dataset, Test-143(unmask), the subset of the test dataset with AlphaFold structures, Test-128 (AF2), and the subset of sequences with at least 5 MoRF and 5 non-MoRF residues, Test-106 (SR) We also used the Test-106 (SR) subset to compare the success rates of these predictors. MoRF predictors are in bold, and the highest AUC and Success Rate values are underlined Table 3Comparison of the prediction average precision score (APS) performances against the Test-143 dataset and its variations: The APS values for MoRFchibi 2.0 and other predictors using the complete test dataset, Test-143, the complete unmasked dataset, Test-143(unmask), and the subset of the test dataset with AlphaFold structures, Test-128(AF2)MeasureAPSAPSAPSAPSTest DataCompleteUnmaskedSubset (AF2)Subset (SR)Naive0.1420.1040.1760.105MoRFchibi 2.0 0.312 0.239 0.365 0.205 OPAL 0.280 0.211 0.327 0.172 MoRFchibi_web 0.255 0.194 0.310 0.161 MoRFchibi_light 0.254 0.191 0.306 0.153 fMoRFpred 0.191 0.140 0.230 0.135 ANCHOR2 0.163 0.114 0.199 0.123 IPA_all0.2590.1850.2950.174IPA_protein0.2530.1800.2870.175DisoFLAG-PB0.2410.1740.2840.186ProBiPred_protein0.2380.1790.2870.164AIUPred-Binding0.1830.1300.2160.144ENSHROUD_all0.1860.1310.2430.139ENSHROUD_protein0.1800.1260.2290.135DeepDISObind_all0.1640.1160.2070.109DeepDISObind_protein0.1630.1150.2080.109bindEmbed21IDR_idrGeneral0.1890.1440.2290.132bindEmbed21IDR_rawGeneral0.1780.1360.2140.123DisoRDPbind0.1680.1220.2240.127AlphaFold_binding0.271IPA-AF2_protein0.284 Table 3Comparison of the prediction average precision score (APS) performances against the Test-143 dataset and its variations: The APS values for MoRFchibi 2.0 and other predictors using the complete test dataset, Test-143, the complete unmasked dataset, Test-143(unmask), and the subset of the test dataset with AlphaFold structures, Test-128(AF2)MeasureAPSAPSAPSAPSTest DataCompleteUnmaskedSubset (AF2)Subset (SR)Naive0.1420.1040.1760.105MoRFchibi 2.0 0.312 0.239 0.365 0.205 OPAL 0.280 0.211 0.327 0.172 MoRFchibi_web 0.255 0.194 0.310 0.161 MoRFchibi_light 0.254 0.191 0.306 0.153 fMoRFpred 0.191 0.140 0.230 0.135 ANCHOR2 0.163 0.114 0.199 0.123 IPA_all0.2590.1850.2950.174IPA_protein0.2530.1800.2870.175DisoFLAG-PB0.2410.1740.2840.186ProBiPred_protein0.2380.1790.2870.164AIUPred-Binding0.1830.1300.2160.144ENSHROUD_all0.1860.1310.2430.139ENSHROUD_protein0.1800.1260.2290.135DeepDISObind_all0.1640.1160.2070.109DeepDISObind_protein0.1630.1150.2080.109bindEmbed21IDR_idrGeneral0.1890.1440.2290.132bindEmbed21IDR_rawGeneral0.1780.1360.2140.123DisoRDPbind0.1680.1220.2240.127AlphaFold_binding0.271IPA-AF2_protein0.284 Comparison of the prediction average precision score (APS) performances against the Test-143 dataset and its variations: The APS values for MoRFchibi 2.0 and other predictors using the complete test dataset, Test-143, the complete unmasked dataset, Test-143(unmask), and the subset of the test dataset with AlphaFold structures, Test-128(AF2) Finally, we evaluated MC2 and other predictors using the post-release dataset Test-new-9. Table 4 shows that MoRFchibi 2.0 outperforms all tested predictors in both AUC and APS. Table 4Comparison of the prediction AUC and APS performances against the Test-new-9 dataset: The AUC and APS values for MoRFchibi 2.0 and other predictors using the post-release Test-new-9 dataset. The APS of a naive predictor for the Test-new-9 dataset is 0.095MeasureAPSAPSMoRFchibi 2.0 0.878 0.246 OPAL 0.807 0.213 MoRFchibi_web 0.809 0.206 MoRFchibi_light 0.833 0.208 fMoRFpred 0.637 0.140 ANCHOR2 0.731 0.177 IPA_all0.8730.215IPA_protein0.8620.210ProBiPred_protein0.6730.157AIUPred-Binding0.8100.189ENSHROUD_all0.8100.226ENSHROUD_protein0.8320.238DeepDISObind_all0.8220.210DeepDISObind_protein0.8150.209bindEmbed21IDR_idrGeneral0.6210.140bindEmbed21IDR_rawGeneral0.5970.117DisoRDPbind0.7800.225 Table 4Comparison of the prediction AUC and APS performances against the Test-new-9 dataset: The AUC and APS values for MoRFchibi 2.0 and other predictors using the post-release Test-new-9 dataset. The APS of a naive predictor for the Test-new-9 dataset is 0.095MeasureAPSAPSMoRFchibi 2.0 0.878 0.246 OPAL 0.807 0.213 MoRFchibi_web 0.809 0.206 MoRFchibi_light 0.833 0.208 fMoRFpred 0.637 0.140 ANCHOR2 0.731 0.177 IPA_all0.8730.215IPA_protein0.8620.210ProBiPred_protein0.6730.157AIUPred-Binding0.8100.189ENSHROUD_all0.8100.226ENSHROUD_protein0.8320.238DeepDISObind_all0.8220.210DeepDISObind_protein0.8150.209bindEmbed21IDR_idrGeneral0.6210.140bindEmbed21IDR_rawGeneral0.5970.117DisoRDPbind0.7800.225 Comparison of the prediction AUC and APS performances against the Test-new-9 dataset: The AUC and APS values for MoRFchibi 2.0 and other predictors using the post-release Test-new-9 dataset. The APS of a naive predictor for the Test-new-9 dataset is 0.095 Case studies To demonstrate MC2’s output and help understand its predictions, we present examples of predictions for six of the Test-143 proteins in Fig. 3. In the example shown in Fig. 3A, MC2 correctly identifies the 157 residues annotated by DisProt as a MoRF. It is clear from the ROC and PR curves in Fig. 2 that such “perfect” predictions are not always possible, and there are scenarios where MoRFchibi 2.0 is not able to identify MoRF sites, as illustrated in Fig. 3B, where MC2 only identifies residues 20–30 that are annotated in the DIBS database as MoRF and fails to identify the extended site 142–485 annotated as MoRF in DisProt. Not all cases are clear-cut; in Fig. 3C, MC2 identifies only half of the MoRF annotated in DisProt (the residues 1-105 case). However, the original paper cited by DisProt, from which the annotation stems, states that ‘the minimal AID construct σ54(16–41) is sufficient to form the activator complex’ [70], i.e., indicating that the first half of this site is more important than the second, notably consistent with MC2’s score distribution for this site. A similar case is shown in Fig. 3D. While the segment 412–490 is annotated as MoRF in DisProt, a structure of the interacting MoRF is only available for the segment 465–490, where the MC2 score is the highest. Perhaps MoRF annotations should not be binary. In Fig. 3E, MC2 scores indicate a MoRF spanning residues 378–439. However, DisProt only annotated residues 378–393 and IDEAL residues 435–439 as MoRFs, leaving the region 394–434 as either an over-prediction by MC2 or an under-annotated region. Interestingly, MC2 appears to predict regions that undergo DtO upon binding to proteins and other molecules, which may be considered overpredictions. Figure 3F shows a case where MC2 scores identify the site around the 41-residue-long segment 294–334 as a MoRF. However, this 41-residue site, which binds to a protein partner, constitutes a zinc finger not stabilized by binding to the protein partner but by zinc ions [71]. Fig. 3Case studies: MoRFchibi 2.0 Normalized Probability scores in blue compared to the DisProt MoRF annotations in red, IDEAL and DIBS annotations in green, and those derived from PDB structures are in black. In (C), we also included a subset of DisProt annotations provided by the source publication, which were deemed sufficient to form the protein complex in Orange Case studies: MoRFchibi 2.0 Normalized Probability scores in blue compared to the DisProt MoRF annotations in red, IDEAL and DIBS annotations in green, and those derived from PDB structures are in black. In (C), we also included a subset of DisProt annotations provided by the source publication, which were deemed sufficient to form the protein complex in Orange Discussion We introduce MoRFchibi 2.0 (MC2), a specialized prediction tool designed to identify the locations of Molecular Recognition Features (MoRFs) in protein sequences. MoRFs constitute a subset of all binding regions within IDRs that undergo a disorder-to-order transition upon binding to protein partners. Our results demonstrate that MC2 outperforms existing MoRF predictors and predictors that identify protein-binding segments more broadly within IDRs. The latter group of predictors includes those ranked highest in the community-driven critical assessments of protein intrinsic disorder predictions (CAID) rounds 1, 2, and 3 [72]. Importantly, MC2’s prediction performance surpasses even methods that leverage AlphaFold output and employ advanced protein language models, delivering superior ROC and Precision-Recall curves and higher success rates. This result is notable, given recent studies showing that segments within IDRs that undergo DtO transitions often have higher pLDDT scores than the surrounding segments or IDRs that do not generally have MoRFs [73]. Regardless, neither MoRFchibi 2.0 predictions nor MoRF annotations are perfect. We provide examples that demonstrate MC2 predictions across various scenarios, including accurate predictions, underpredictions, overpredictions, and predictions for regions with ambiguous or incomplete annotations. Most importantly, MC2, relying solely on amino acid sequences, predicts regions with a low disorder ‘signature’ that bind to proteins and are likely to fold upon binding as MoRFs. However, we highlighted in Fig. 2F that protein binding and folding can be influenced by environmental factors, such as ions or post-translational modifications, thereby further complicating predictions. MoRFchibi 2.0 generates output scores using an ensemble of four LR-CNN models followed by a reverse Bayes Rule. Consequently, these four scores, along with their average, represent probabilities normalized for the priors in the training data. As a result, MC2 scores are individually interpretable, enabling direct comparison and integration with other scores following the same output framework, such as IPA [47]. We provide an HTML server and downloadable code for the broader community to use this new method and make the newly assembled dataset with MoRF annotations accessible to all developers: https://gsponerlab.msl.ubc.ca/software/morf_chibi/mc2.