WorksComputer Science
Uneven gains in protein-RNA affinity and mutation-effect prediction
Abstract
Accurate matched affinities determine mutation effects, but gains in aggregate affinity prediction may transfer unevenly to average effects, ordering and local accuracy. We evaluated eleven method families using method-specific matched populations. Controlled CoPRA reduced absolute-affinity mean absolute error from 1.2898 to 1.1431 pKd. Across 220 mutation pairs from 20 families, it generated stronger responses and improved mean-effect accuracy, with uncertain ranking gains and no decrease in mean centered local error. In an independent external test (24 quantitative pairs from three source-matched families), absolute and ordinary mutation-effect errors fell in all three families. Mean centered local error fell from 0.4198 to 0.3761 pKd, with improvements in two families and higher error in one. Error decomposition localized gains in landscape means. PRA-Pred feature tracing, fixed-weight M2-PRNet input controls and matched supervision interventions identified specific restrictions on mutation response. These results establish the need to evaluate average effects, ordering and quantitative local variation jointly.
Introduction
Single amino-acid substitutions can alter protein–RNA recognition, regulation and molecular function [2, 7]. Their effects on binding affinity follow an exact relationship: , or . Recognition combines base-specific contacts, stacking, electrostatics and conformational rearrangements [5]. Each binding free energy reflects the balance between the complex and its unbound partners. Accurate predictions at both matched endpoints therefore recover the mutation effect. Whether improvements in aggregate affinity prediction deliver that accuracy is an empirical question. Equilibrium and concentration controls affect both absolute affinities and relative specificity, making matched experimental context essential for interpreting mutation effects [13].
Kappel et al. directly tested this connection: GLM-Score’s absolute-affinity advantage did not transfer to relative prediction, motivating Rosetta-Vienna RNP- and explicit bound/unbound energetics [14]. CoPRA later evaluated protein substitutions with complex-pretrained networks and mutation-effect fine-tuning, including per-complex metrics [9]. These studies establish the problem and motivate examining where gains arise across contemporary method configurations.
The distinction extends beyond absolute versus relative error. Shared endpoint errors cancel in a WT–mutant difference, whereas unequal errors remain. Within a mutation landscape sharing one WT, a predictor can improve the average effect without resolving differences among substitutions. Ranking and response amplitude reveal whether such behavior reflects weak responses, incorrect ordering or inaccurate effect sizes Figure 1.

Figure 1. Affinity and mutation-effect evaluation. a, RNase E–ssRNA (2C4R) and the 2.80 Å Phe57–G10 contact. b, Cross-system endpoints and within-system substitutions sharing a WT and assay. c, Same-member model comparisons, PRA-Pred regression terms and M2-PRNet branches.
Using source-linked ProNAB measurements [11], we compare eleven method families spanning sequence, structure, multimodal and direct-effect prediction (Table 1). Each comparison uses common members and documented model configurations. We jointly evaluate absolute affinity, overall mutation effects and centered landscape variation; two error decompositions locate the gains. PRA-Pred feature tracing and fixed-weight M2-PRNet controls then identify computational restrictions on response, while a matched three-arm continuation tests a change in supervision. Experimental-to-input correspondence links these analyses to specific constructs and binding measurements. An independent external test of the fixed controlled models examines how these gains transfer to three source-matched families. Together these analyses identify which capabilities must be demonstrated before an affinity model is used to prioritize mutations.
System and methods
Data and evaluation units
We grouped ProNAB records by construct, study and conditions, expressing dissociation constants as pKd = and taking the median across repeated measurements of one unit. A pair linked one protein substitution to a WT under the same study and conditions. Multiple substitutions, simultaneous protein/RNA changes and unresolved WT links were excluded. Local evaluation required at least five distinct mutated positions after retaining common prediction coverage.
Source reconciliation excluded seven censored units and collapsed ten duplicate Hfq pairs to the source-consistent WT anchors. The principal scopes comprise 752 absolute units, 338 ordinary pairs and 220 local pairs at 203 positions in 21 landscapes and 20 families. Supplementary S1 documents exclusions and experimental-to-input correspondence. Training retains the original fitted membership. Source systems include Hfq, Trm5, Unkempt, Dis3l2 and Dicer ([27, 4, 18, 8, 24]).
The absolute cohort included reported dissociation constants with usable units and identified constructs. RNA variants could contribute to the absolute cohort. For mutation effects, a landscape groups substitutions sharing an experimental WT anchor. Landscapes with fewer than five positions still contribute to the ordinary mutation summary. In linked absolute analyses, each distinct WT unit is counted once even if it anchors several pairs. Source reconciliation also checks the correspondence between the assayed construct and the mapped protein and RNA chains (Supplementary S1 and Table S30).
Models, training data and evaluation design
The principal CoPRA configuration [9] used frozen ESM2 and RiNALMo encoders [15, 20], freshly initialized Co-Former and affinity-readout modules, and absolute-affinity MSE training. Its additive comparator pooled protein and RNA representations separately and summed two readouts with a learned bias. Both recipes shared data, encoder states, structural input selection and optimization settings; their task modules and capacities differed. Protein families, defined by single-linkage sequence relationships meeting 40% identity and 70% coverage, remained within one of five outer folds. Each method was fitted at three seeds. Each pair’s endpoints used the same fold-specific checkpoint.
CoPRA retained WT coordinates and placed each substitution on the first compatible matched chain. Its official parser used the first coordinate model and a 256-node interface patch. M2-PRNet used five released checkpoints with complete residue-index alignment. Table 1 lists all configurations. Supplementary S2 documents mutant-building, preprocessing and training following DOME reporting principles [25].
Both controlled models were trained on the absolute-affinity records outside the held-out fold, including records beyond the local mutation cohort. Across five folds, three seeds and two methods, the design gives 30 fits. Both methods used the final checkpoint after 30,000 Adam updates. For each cross-method comparison, we took the intersection of available predictions, including all three seeds for out-of-fold models, then reapplied the local position threshold (Supplementary S1–S2).
| Method | Configuration | Source | Abs. | Ord. | Local |
| CoPRA | Controlled family-fold OOF | [9] | 752 | 338 | 220 |
| DeePNAP | Controlled family-fold OOF | [19] | 562 | 242 | 161 |
| OmniBioTE | Official checkpoint and adapted frozen-head OOF | [3] | 736 | 328 | 220 |
| PRdeltaGPred | Structural port (WT and mutants) | [12] | 286 | 235 | 153 |
| GBNL | Author-approved reconstruction (OOF) | [29] | 391 | 149 | 82 |
| SVSBI | Author-approved reconstruction (OOF) | [23] | 391 | 149 | 82 |
| ProRB | Author-provided checkpoint | [28] | 391 | 149 | 82 |
| PANTHER | Native service score | [1] | — | 87 | 31ᵃ |
| PRA-Pred | ss+str and ds+enz applications | [10] | 121 | 106 | 78 |
| mCSM-NA | Direct mutation-effect service | [21] | — | 246 | 157 |
| M2-PRNet | Five released checkpoints with reconstructed, index-aligned inputs | [26] | 257 | 211 | 131 |
Table 1. Model configurations and matched coverage. Coverage counts absolute units, ordinary pairs and local pairs. OOF means out-of-fold. ᵃPANTHER rank covers 31 pairs in four common landscapes. See Supplementary S2 for implementation details and S5 for matched comparisons.
Metrics and statistical analysis
For mutation in landscape , the observed effect is , with the predicted effect defined identically. Negative pKd indicates weaker binding. Each position receives equal weight, shared among substitutions. Ordinary MAE is the weighted MAE of . Centered MAE (cMAE) first subtracts each landscape’s weighted mean from its observed and predicted effects. We average errors across landscapes within each family, then across families. Absolute MAE averages units within families. Out-of-fold metrics average three separate seed scores. The zero baseline predicts pKd . Position-mean Spearman measures ordering. The position-weighted ratio of predicted to observed SD measures response amplitude.
Position-mean Spearman is computed within each landscape after averaging predicted and observed effects separately over substitutions at each position. A constant predicted position profile has undefined rank. An SD ratio of one matches experimental spread, and measures amplitude mismatch. The primary local ordinary and centered errors use the same 220 pairs (Supplementary S3).
For endpoint residuals and , let and . The endpoint squared error per incidence is . Within a landscape, mutation squared error partitions into and . The centered term is predicted variance plus observed variance minus twice their covariance. Each identity uses common weights and repeated WT incidence in the endpoint decomposition.
Paired intervals resample whole families from fixed predictions. The broad ordinary and endpoint analyses use 2,000 draws. Local analyses share 10,000 draws across metrics and methods. These intervals describe between-family variation for the saved predictions. Intervals are unadjusted. Free energies use pK d at 298 K, with kcal/mol. mCSM-NA strengthening scores are divided by that denominator. PANTHER uses in native units.
Computational diagnostics and supervision intervention
PRA-Pred endpoints were reconstructed from the executed regression features and original coefficients. Contributions sum to the unrounded mutation response, distinguishing unchanged inputs, cancellation and display-rounding ties. M2-PRNet diagnostics replayed all eight WT/mutant combinations of atom, residue and image branches through fixed fusion weights. Exact Shapley enumeration [16] used WT states as the baseline and mutant states as replacements. Encoder/readout traces located response attenuation. Mask-only and complete-index alignment changed input organization while retaining physical feature values and checkpoint weights (Supplementary S6–S7).
For each of three seeds, a shared absolute-trained A0 model was cloned into three matched continuations with identical initial parameters, optimizer state, sampling and update budgets. Arm A retained absolute MSE; B added ; C added . The auxiliary terms are and : C increases difference weighting while removing the auxiliary shared-error term. This fixed-split experiment is separate from the principal five-fold evaluation. Validation includes 70 pairs, of which 54 local pairs span four held-out families; training-fit diagnostics retain the actual training records and weights (Supplementary S8).
External test. We reconciled SF3A1, SEPSECS and Npl3 constructs, WT anchors and fixed-backbone inputs with the source studies [6, 22, 17]. Before prediction we checked provenance, development use and joint sequence-family connectivity against all 771 fitted units (40% global identity, 70% coverage). We averaged five fixed fold predictions per endpoint within each seed before differencing, then averaged three separate seed scores. The external test used 24 quantitative pairs and 27 absolute units. Four censored or nonbinding pairs contributed to coverage. Each source panel forms one landscape and one family. Equal-family aggregation gives each panel the same weight (Supplementary S9).
Results
Errors use pKd. Effects are additive minus CoPRA for errors and amplitude distance, and CoPRA minus additive for Spearman and SD ratio. Intervals use paired family resampling. See Supplementary Table S36 for linked absolute errors and contrasts with zero change.
Gains differ across prediction levels
CoPRA reduced absolute and ordinary mutation errors (Figure 2; Table 2). On the 220 local pairs, ordinary MAE fell while the paired cMAE contrast spanned zero.
Figure 2. Prediction errors, response amplitude and ordering. a, Errors and paired reductions on 752 absolute units, 338 ordinary pairs and 220 local pairs. Diamonds are family-macro means and bars are 95% family-bootstrap intervals. b, Local-scope reductions across 21 landscapes and 20 families, scaled by column. c, SD ratio (log radius), position Spearman (angle) and 50%/95% bootstrap mean-density contours. See Table 2.
| Measure | CoPRA | Additive | Zero | Paired effect | 95% interval |
| A. Broad evaluation: 752 units / 338 ordinary pairs | |||||
| Absolute MAE, 752 units | 1.1431 | 1.2898 | — | — | — |
| Ordinary MAE, 338 pairs | 0.7917 | 0.8629 | 0.8816 | 0.0713 | [0.0362, 0.1068] |
| B. Local evaluation: 220 pairs / 21 landscapes / 20 families | |||||
| Ordinary MAE | 0.6841 | 0.7426 | 0.7580 | 0.0585 | [0.0006, 0.1126] |
| cMAE | 0.3991 | 0.3749 | 0.3787 | −0.0243 | [−0.0635, 0.0088] |
| Position-mean Spearman | 0.1886 | 0.1481 | — | 0.0405 | [−0.0822, 0.1645] |
| Predicted / observed SD | 0.6422 | 0.0709 | 0 | 0.5714 | [0.2808, 0.9560] |
| SD ratio | 0.7707 | 0.9291 | 1 | 0.1585 | [−0.0596, 0.3266] |
Table 2. Controlled CoPRA errors, local capabilities and paired family uncertainty.
The predicted-to-observed SD ratio showed stronger responses. Position-mean Spearman measures ordering, and distance from an SD ratio of one measures amplitude calibration. The paired intervals for ordering and calibration both span zero.
Each reduction is additive error minus method error on identical members. Positive values favor the method. Coverage counts absolute units, ordinary pairs and local pairs. Supplementary S5 gives full errors and references, including OmniBioTE’s additional sequence-method intersection (Table S3) and PANTHER diagnostics (Table S11).
Table 3 reports gains against matched additive references. Official OmniBioTE lowered absolute error on 736 units, with ordinary errors near zero change. PRA-Pred also had low absolute error, with ordinary error close to CoPRA on its scope. M2-PRNet compressed responses (SD ratio 0.0367), as did ProRB, SVSBI and GBNL (0.0515–0.2048). mCSM-NA amplified responses (SD ratio 6.197) and had local cMAE of 1.6998. Supplementary S5 gives complete same-member diagnostics.
| Method | Scope | Additive ref. | Abs. | Ord. | Local | Abs. | MAE | Ord. MAE | Local cMAE |
| DeePNAP (OOF) | S3 | Sequence | 562 | 242 | 161 | ||||
| OmniBioTE official | S4 | CoPRA | 736 | 328 | 220 | ||||
| OmniBioTE adapted (OOF) | S4 | CoPRA | 736 | 328 | 220 | ||||
| PRdeltaGPred | S6 | CoPRA | 286 | 235 | 153 | ||||
| mCSM-NA | S7 | CoPRA | — | 246 | 157 | — | |||
| GBNL reconstruction (OOF) | S8 | CoPRA | 391 | 149 | 82 | ||||
| SVSBI reconstruction (OOF) | S8 | CoPRA | 391 | 149 | 82 | ||||
| ProRB author checkpoint | S8 | CoPRA | 391 | 149 | 82 | ||||
| PRA-Pred | S9 | CoPRA | 121 | 106 | 78 | ||||
| M2-PRNet, indices aligned | S10 | CoPRA | 257 | 211 | 131 |
Table 3. Error reductions against matched additive baselines.
PANTHER (native score). The service gave zero differences for 47/87 pairs and matched the strict direction in 26.4%. Mean position Spearman was 0.2796 over the four commonly rankable landscapes (31 pairs at 29 positions). Predictions were constant in the fifth eligible landscape. See Supplementary Table S11 for matched reference values.
Error reductions concentrate in shared and landscape-mean components
Endpoint residuals shared by both members cancel when affinities are differenced. On 338 pairs, controlled CoPRA reduced pair-incidence endpoint MSE by 0.4227 : the shared component contributed 0.3686 and the difference component 0.0541 (Figure 3). The shared-component interval crossed zero, whereas the difference-component interval was positive. Thus some improvement reached mutation differences, even though the larger point-estimate contribution concerned shared endpoint placement.

Figure 3. Endpoint-error decomposition. A, Weighted residual densities and shared/difference contributions to endpoint MSE. D, Joint bootstrap density and marginal 95% intervals from 2,000 paired family draws. Reductions are additive minus CoPRA (338 pairs, 52 families). See Supplementary Figure S11 and Table S17.
The second decomposition locates the remaining gain within landscapes. On the same 220 local pairs used for centered accuracy, the landscape-mean term improved by 0.1313 pKd (95% interval, 0.0674 to 0.2103), while the centered term worsened by 0.0418 (reduction −0.0418; −0.1109 to 0.0040). Their sum yielded a total mutation-MSE reduction of 0.0895 (Figure 4A–B). Better average effects accounted for the gain; variation among substitutions had higher mean error. These are matched-member comparisons, with the broader endpoint analysis serving a separate descriptive scope.
Figure 4. Landscape-mean and local error decomposition. A, MSE reductions on 220 pairs in 21 landscapes and 20 families. B, Joint density and 95% intervals from 10,000 paired family draws. C, Local MSE change from predicted variance and covariance with experiment. D, Auxiliary losses and B-to-C responses on 54 validation pairs in four families. The inset gives B-minus-C reductions. Units are pKd. See Supplementary S11.
Variance–covariance accounting explains why stronger responses did not improve local MSE. Relative to additive, CoPRA increased predicted variance by 0.06626 pKd, while twice its covariance gain with experiment offset only 0.02443. The remaining 0.04183 is the centered-MSE increase (Figure 4C and Supplementary Table S22).
Computation can suppress responses before evaluation
All 42 exact unrounded zeros among 106 PRA-Pred pairs had unchanged effective regression feature vectors. Rounding endpoints to two decimals added 13 ties (Figure 5A). In 2C4R F57A, the mutation removes a contact whose energy was already clipped to zero and leaves the charge category unchanged, so every effective input stays the same. Supplementary S6 and Table S24 trace the fixed-backbone RNA geometry terms and the retained NO/NO1 naming behavior.

Figure 5. PRA-Pred feature contributions and response. A, Fixed-order tracks for 13 ss+str and 93 ds+enz pairs. Signed terms and summed responses share a symmetric-log scale (kcal/mol). Counts show changed features, exact zeros, two-decimal endpoint ties and added ties. B, Pairwise term contributions colored by WT feature percentile within each term. C, Unit covariance ellipses between landscapes (solid) and within landscapes (dashed, ×10 linear scale). See Supplementary Figure S13 and Section S11.
M2-PRNet input alignment
In the original reconstructed M2-PRNet path, local SD ratios were 0.0249–0.0680 across checkpoints and fell further with averaging (Figure 6A). The atom branch supplied 99.5% of summed mean absolute Shapley contributions, with little cancellation. Encoder/readout tracing found small residue displacement and weak downstream sensitivity, including weak coupling from image changes (Supplementary S7 and Figure S14).

Figure 6. Figure 6. M2-PRNet response and input alignment. A, Original-path SD ratios for five checkpoints and their ensemble (131 pairs, 12 landscapes/families). The macro row weights families equally. B, Relative SD-ratio change and centered MAE change from the original path. Counts use 1,055 pair–checkpoint records and 211 ensemble pairs. See Supplementary S7/S11 and Figure S14.
The original adapter’s indexing left 67/211 mutation slots inactive relative to the RNA-first network layout. With pairs and checkpoint weights fixed, both mask-only and complete alignment eliminated identical WT/mutant residue vectors and both exact ensemble zeros (Figure 6B). From the original to complete path, the local SD ratio rose from 0.0358 to 0.0367, while family-macro cMAE was nearly identical across original, mask-only and complete paths (0.3405, 0.3407 and 0.3407 pKd). Atom contributions stayed above 99%. This comparison applies to the reconstructed input paths documented in Supplementary S2 and S7.
Difference-weighted supervision improves average effects more than local variation
The three-arm experiment tested whether reweighting supervision could shift gains toward local variation. Relative to endpoint arm B, difference arm C reduced ordinary validation MAE from 0.7219 to 0.6963 pKd, while local cMAE changed from 0.3568 to 0.3576. On the identical 54 local pairs, the landscape-mean MSE improvement was 0.01961 pKd and the centered improvement −0.00033 (Figure 4D). One family improved cMAE at every seed; the other three showed mixed signs (Figure 7B). The gain again concentrated in average effects.
Figure 7. Supervision and local generalization. A, Three-seed means for arms A/B/C. Error colors show gains relative to A. Training uses 451 units/200 pairs (124 local), and validation uses 164 units/70 pairs (54 local). B, B-minus-C reductions on 54 local pairs, showing individual seeds, means and ranges. C, Paths join arms within landscapes on common axes. See Supplementary S8/S11.
Training fits recovered much stronger local variation (Figure 7A,C). B/C training cMAEs were 0.1133/0.1443 versus zero change at 0.4020. Their SD ratios fell from approximately 0.96 in training to 0.22–0.24 in validation, accompanied by weak held-out ranking. Thus the network fitted training-side mutation differences, but this ability transferred poorly to the four validation families. Complete arm results, including A0 and seed variation, are reported in Supplementary S8.
Family variation and experimental context
CoPRA improved cMAE in eleven families and worsened it in nine; larger losses in the latter group produced the unfavorable overall mean. This heterogeneity shows how different binding systems contribute to the aggregate result. Hfq illustrates the separation between capabilities: relatively strong mutation ordering coexisted with compressed response amplitude. Supplementary S1 connects this case and the other binding systems to their experimental constructs and the structures supplied to the controlled CoPRA model. Excluding the entire 1JBR landscape, whose mutated-copy correspondence remains unresolved, retained the local-error pattern: cMAE 0.4018 versus additive 0.3758 on 209 pairs.
Independent external test
In the independent external test, controlled CoPRA lowered absolute and ordinary mutation-effect MAE in each family. Equal-family MAE fell from 0.6932 to 0.5282 pKd for absolute affinities and from 0.8845 to 0.6995 pKd for ordinary effects. Local cMAE fell from 0.4198 to 0.3761 pKd (additive-minus-CoPRA reduction 0.0437, 95% family-resampling interval −0.0396 to 0.0891). SEPSECS and Npl3 improved and SF3A1 worsened. CoPRA produced stronger responses across all three families and had lower mean position Spearman correlation (0.1454 versus 0.2208). Supplementary Section S9 gives coverage and family results (Tables E1–E2 and Figure E1).
Discussion
The endpoint error bound, , holds for each matched pair. At population level, shared endpoint errors cancel when affinities are differenced. Within a landscape, stronger predicted responses improve centered squared error only when the gain in covariance with experiment offsets the added prediction variance. These distinctions extend earlier absolute-to-relative evaluations [14, 9] and help explain why gains can differ across constructs and prediction levels.
For mutation prioritization, evaluate how well the model orders substitutions and estimates their effect sizes on the intended binding systems. A near-constant profile can achieve low centered error while providing little basis for ranking candidates. Read ordering, response amplitude and centered error together, alongside the accuracy of average landscape effects. The external interval reflects between-family variation across three source-matched systems.
Data availability
The primary binding-affinity measurements were obtained from ProNAB and are subject to the provider’s terms. Supplementary Section S9 identifies the measurement sources for the independent external test. The authors’ original analysis and figure-generation code is licensed under the MIT License. The code and study-generated data are available on GitHub (v2026.11.17). Access to third-party model checkpoints is governed by the respective providers’ terms.
Author contributions
Shangkun Sima designed the methods, curated the data, developed the software and computational workflows, performed the analyses and validation, created the figures, and drafted and revised the manuscript. Tongfei Shen edited selected passages of the manuscript, refining their wording to improve clarity and readability. Tianchi Lu formulated the initial research question, supervised the study, and made substantive revisions to the final manuscript.
Acknowledgements
We thank the ProNAB team at the Indian Institute of Technology Madras for curating binding-affinity data, the Worldwide Protein Data Bank consortium and structure depositors for making structural data available, and the researchers and institutions that generated the experimental measurements used in the main and external evaluations. We thank the developers of the evaluated models and supporting software for making their resources available, and the M2-PRNet authors for providing three-view images and preprocessing materials.
Funding
This research received no specific funding.
AI assistance
The authors used a large language model (Claude Opus 4.8, Anthropic) solely to assist in code writing, a use permitted under the ISCB Acceptable Use of Large Language Models policy adopted by Bioinformatics. The tool helped write analysis scripts and their unit tests during data preparation in May 2026; none of that code produced a result reported in this paper. The authors specified all analyses, reviewed and edited every generated line, executed the code, and remain solely responsible for the code and for all results derived from it. No large language model was used to draft, write, paraphrase, translate or edit any part of the manuscript, and no AI-generated text, table or image appears in the paper. The machine-learning models evaluated in this study are the subject of the research and are used as described in System and methods; they are not AI tools used in the preparation of the manuscript. Details are given in Supplementary S13.
References
- [1]Aletayeb P, Biswas AD, Rocca S et al. PANTHER Score: Protein-Affinity for Nucleic Target-binding, Hybridization, and Energy Regression. RNA 2026;32:131–149. doi:10.1261/rna.080646.125.DOI
- [2]Amrane S, Rebora K, Zniber I et al. Backbone-independent nucleic acid binding by splicing factor SUP-12 reveals key aspects of molecular recognition. Nat Commun 2014;5:4595. doi:10.1038/ncomms5595.DOI
- [3]Chen SF, Steele RJ, Hocky GM et al. Large-scale multi-omic biosequence transformers for modeling protein–nucleic acid interactions. PLoS One 2026;21:e0341501. doi:10.1371/journal.pone.0341501.DOI
- [4]Christian T, Lahoud G, Liu C et al. Mechanism of N-methylation by the tRNA m1G37 methyltransferase Trm5. RNA 2010;16:2484–2492. doi:10.1261/rna.2376210.DOI
- [5]Corley M, Burns MC, Yeo GW. How RNA-Binding Proteins Interact with RNA: Molecules and Mechanisms. Mol Cell 2020;78:9–29. doi:10.1016/j.molcel.2020.03.011.DOI
- [6]de Vries T, Martelly W, Campagne S et al. Sequence-specific RNA recognition by an RGG motif connects U1 and U2 snRNP for spliceosome assembly. Proc Natl Acad Sci USA 2022;119:e2114092119. doi:10.1073/pnas.2114092119.DOI
- [7]Dominguez C, Fisette JF, Chabot B et al. Structural basis of G-tract recognition and engaging by hnRNP F quasi-RRMs. Nat Struct Mol Biol 2010;17:853–861. doi:10.1038/nsmb.1814.DOI
- [8]Faehnle CR, Walleshauser J, Joshua-Tor L. Mechanism of Dis312 substrate recognition in the Lin28-let-7 pathway. Nature 2014;514:252–256. doi:10.1038/nature13553.
- [9]Han R, Liu X, Pan T et al. CoPRA: Bridging Cross-domain Pretrained Sequence Models with Complex Structures for Protein-RNA Binding Affinity Prediction. arXiv, version 2, 2025. arXiv:2409.03773v2.arxiv.org/abs/2409.03773
- [10]Harini K, Sekijima M, Gromiha MM. PRA-Pred: Structure-based prediction of protein-RNA binding affinity. Int J Biol Macromol 2024;259:129490. doi:10.1016/j.ijbiomac.2024.129490.DOI
- [11]Harini K, Srivastava A, Kulandaisamy A et al. ProNAB: database for binding affinities of protein–nucleic acid complexes and their mutants. Nucleic Acids Res 2022;50:D1528–D1534. doi:10.1093/nar/gkab848.DOI
- [12]Hong X, Tong X, Xie J et al. An updated dataset and a structure-based prediction model for protein–RNA binding affinity. Proteins 2023;91:1245–1253. doi:10.1002/prot.26503.DOI
- [13]Jarmoskaite I, AlSadhan D, Vaidyanathan PP, Herschlag D. How to measure and evaluate binding affinities. eLife 2020;9:e57264. doi:10.7554/eLife.57264.DOI
- [14]Kappel K, Jarmoskaite I, Vaidyanathan PP et al. Blind tests of RNA–protein binding affinity prediction. Proc Natl Acad Sci USA 2019;116:8336–8341. doi:10.1073/pnas.1819047116.DOI
- [15]Lin Z, Akin H, Rao R et al. Evolutionary-scale prediction of atomic-level protein structure with a language model. Science 2023;379:1123–1130. doi:10.1126/science.ade2574.DOI
- [16]Lundberg SM, Lee SI. A Unified Approach to Interpreting Model Predictions. Adv Neural Inf Process Syst 2017;30. Conference paper.arxiv.org/abs/1705.07874
- [17]Moursy A, Cléry A, Gerhardy S et al. RNA recognition by Npl3p reveals U2 snRNA-binding compatible with a chaperone role during splicing. Nat Commun 2023;14:7166. doi:10.1038/s41467-023-42962-4.DOI
- [18]Murn J, Teplova M, Zarnack K et al. Recognition of distinct RNA motifs by the clustered CCCH zinc fingers of neuronal protein Unkempt. Nat Struct Mol Biol 2016;23:16–23. doi:10.1038/nsmb.3140.DOI
- [19]Pandey U, Behara SM, Sharma S et al. DeePNAP: A Deep Learning Method to Predict Protein–Nucleic Acid Binding Affinity from Their Sequences. J Chem Inf Model 2024;64:1806–1815. doi:10.1021/acs.jcim.3c01151.DOI
- [20]Penič RJ, Vlašić T, Huber RG et al. RiNALMo: general-purpose RNA language models can generalize well on structure prediction tasks. Nat Commun 2025;16:5671. doi:10.1038/s41467-025-60872-5.DOI
- [21]Pires DEV, Ascher DB. mCSM–NA: predicting the effects of mutations on protein–nucleic acids interactions. Nucleic Acids Res 2017;45:W241–W246. doi:10.1093/nar/gkx236.DOI
- [22]Puppala AK, Castillo Suchkow J, French RL et al. Structural basis for the tRNA-dependent activation of the terminal complex of selenocysteine synthesis in humans. Nucleic Acids Res 2023;51:4012–4026. doi:10.1093/nar/gkad182.DOI
- [23]Shen L, Feng H, Qiu Y et al. SVBSBI: sequence-based virtual screening of biomolecular interactions. Commun Biol 2023;6:536. doi:10.1038/s42003-023-04866-3.DOI
- [24]Tian Y, Simanshu DK, Ma JB et al. A phosphate-binding pocket within the platform–PAZ–connector helix cassette of human Dicer. Mol Cell 2014;53:606–616. doi:10.1016/j.molcel.2014.01.003.DOI
- [25]Walsh I, Fishman D, Garcia-Gasulla D et al. DOME: recommendations for supervised machine learning validation in biology. Nat Methods 2021;18:1122–1127. doi:10.1038/s41592-021-01205-4.DOI
- [26]Wang J, Luo G, Yang Y et al. M2-PRNet: multi-scale and multi-modal learning for protein–RNA binding affinity prediction. Bioinformatics 2026;42:btag614. doi:10.1093/bioinformatics/btag614.DOI
- [27]Wang L, Wang W, Li F et al. Structural insights into the recognition of the internal A-rich linker from OxyS sRNA by Escherichia coli Hfq. Nucleic Acids Res 2015;43:2400–2411. doi:10.1093/nar/gkv072.DOI
- [28]Xue Y, Liu X, Zhu W et al. ProRB: a structure-free unified framework for joint prediction and design of protein–RNA interactions. Nucleic Acids Res 2026;54:gkag870. doi:10.1093/nar/gkag870.DOI
- [29]Zia M, Suwayyid F, Wei GW. GBNL: Graded Betti number learning of complex biological data. Found Data Sci. Published online 18 May 2026; assigned to 2027;12:158–178. doi:10.3934/fods.2026014.DOI


