Abstract: F-block metal recognition is important for rare-earth separation, nuclear-waste treatment, bioinspired materials, and medical technologies, yet sequence-level identification of f-block metal-binding peptide motifs remains difficult because experimentally characterized examples are scarce, structurally heterogeneous, and unevenly distributed across different elements. Here, we developed a protein-language-model-assisted framework for first-pass screening of sequence patterns associated with f-block and rare-earth metal coordination environments. Protein structures containing target f-block elements or yttrium were collected from the RCSB Protein Data Bank and processed through structural parsing, distance-based identification of coordination-associated residues, hierarchical sequence clustering, representative selection, and cluster-aware five-fold partitioning. A 4.5 Å(1 Å=0.1 nm) metal-centered distance window was used as an operational criterion for extracting local metal-associated environments rather than as a strict definition of direct metal-donor bonding or the first coordination sphere. Continuous peptide fragments spanning the first and last coordination-associated residues were treated as positive samples, whereas non-overlapping fragments containing no coordination-associated residues were used as structure-derived negative samples relative to the annotated metal site. After structural filtering, 648 protein-chain objects were retained. A two-level clustering strategy was then applied to reduce sequence redundancy and minimize information leakage. Complete protein chains were first clustered, and coordination-associated fragments within the resulting chain-level clusters were subsequently clustered again. Within each second-level cluster, fragments were ranked primarily by the number of metal-proximal residues and secondarily by the mean metal-residue distance, and no more than three representative coordination-associated fragments were retained. Non-coordinating fragments originating from the same primary chain-level clusters were assigned together with the corresponding cluster units during fold construction. The final curated dataset contained 609 coordination-associated sequences and 1167 non-coordinating sequences. Residue-level contextual embeddings extracted from the pretrained ESM-2 model were mean-pooled into fixed-dimensional fragment representations and used to train logistic regression and multilayer perceptron classifiers for binary screening of sequence patterns associated with f-block metal coordination. The models were trained and evaluated using the cluster-aware five-fold partition. Metrics reported as five-fold averages were calculated separately within each held-out fold, whereas pooled out-of-fold metrics were calculated after combining predictions from all held-out folds. The multilayer perceptron model achieved a pooled out-of-fold area under the receiver operating characteristic curve of 0.9096 and an area under the precision-recall curve of 0.8846, outperforming the logistic regression baseline. The classification threshold was selected by maximizing the Matthews correlation coefficient on the pooled out-of-fold predictions. Consequently, threshold-dependent performance values should be interpreted as out-of-fold-optimized estimates rather than as results obtained using a threshold fixed independently beforehand. An independent negative-only dataset derived from M-Ionic was further used to evaluate negative-sequence rejection and false-positive tendency rather than complete external two-class generalization. At the selected thresholds, the multilayer perceptron model correctly rejected 3080 of 3145 external negative sequences, whereas logistic regression rejected 2704. The improved rejection performance of the multilayer perceptron model therefore supported its use as the primary classifier for subsequent feature-ablation and designed-motif analyses. Feature-ablation experiments compared ESM-2 embeddings with position-specific scoring matrix features, DSSP-derived structural descriptors, and combined PSSM+DSSP representations under the same training and evaluation protocol. ESM-2 produced the strongest overall performance, indicating that protein-language-model-derived contextual representations captured sequence information that was not fully reproduced by the explicit evolutionary and structural descriptors used in this study. To investigate model behavior from a coordination-chemistry perspective, EF-hand-inspired 12-residue motif libraries were constructed as chemically constrained probes. These libraries contained 4500 theoretical positive motifs and 5000 chemically perturbed negative motifs representing key-site disruption, acidic-residue mispositioning, acidic-residue overloading, inappropriate donor-type replacement, and pseudo-EF-hand variants. Because the structure-derived training fragments were embedded within their complete protein-chain context whereas the designed motifs were embedded as standalone 12-residue sequences, the designed-motif analysis was treated as a qualitative probe under a shifted sequence context rather than as an independent validation of peptide-binding performance. The model recognizes 98.0% of the designed positive motifs but rejects only 55.9% of the designed negative motifs. Donor-type replacement and disruption of key coordination positions produced strong negative responses, whereas acid-overloaded and positionally misarranged acidic motifs generated substantial false-positive predictions. Position-wise residue-frequency analysis further revealed pronounced Asp and Glu enrichment among false-positive motifs. These results are chemically consistent with the preference of lanthanide-like hard metal ions for oxygen-rich donor environments, while also showing that mean-pooled ESM-2 representations capture overall acidic-residue composition and donor-type compatibility more effectively than precise position-specific donor arrangements. The extracted fragments originated from folded proteins, and their coordination-associated patterns may depend on tertiary structure, flanking secondary-structure elements, long-range interactions, and conformational preorganization. Moreover, the available structural examples were strongly imbalanced across different elements, with far fewer actinide-containing structures than lanthanide- or yttrium-containing structures. Accordingly, the resulting scores should be interpreted as first-pass indicators of coordination-associated sequence patterns rather than as direct measures of binding affinity, metal selectivity, independent peptide binding, or detailed coordination geometry. Instead, the scores are best used to rank or enrich candidate sequences for subsequent structural prediction, metal-coordination modeling, molecular simulation, and experimental validation. This study establishes an initial sequence-based screening route for data-scarce f-block coordination problems and defines both the practical value and current limitations of protein-language-model-derived representations for chemically informed peptide screening.