LAMP-PRo

Nucleic-acid-binding proteins (NABPs) fall into two broad functional classes — DNA-binding proteins (DBPs) and RNA-binding proteins (RBPs) — that are notoriously hard to tell apart computationally. The two classes share structural motifs, evolutionary history, and physicochemical character, so sequence-based predictors routinely commit cross-prediction errors: calling a DBP an RBP and vice versa. Harder still is the subset of proteins that bind both, the dual DNA- and RNA-binding proteins (DRBPs), which most tools either ignore or handle as an afterthought of two independent binary classifiers.

LAMP-PRo attacks both problems by treating the task explicitly as multi-label learning. Protein sequences are embedded with a pretrained protein language model — ESM-2 — and passed through a convolutional encoder and multi-head self-attention, then through two attention modules built for the label structure itself: a label-aware attention layer that computes a separate, class-specific summary of the sequence for DBP, RBP, and non-NABP, and a cross-label attention layer that lets the DBP and RBP label representations attend to one another. DRBP is never predicted as its own output; it is inferred from the co-activation of the DBP and RBP probabilities, which keeps the label space consistent between training and inference.

The model was developed by Nimisha Ghosh and colleagues in the Department of Computer Science and Engineering at Shiv Nadar University Chennai and posted to arXiv in September 2025. It is a preprint and has not yet completed peer review.

#Key Features

• Multi-label rather than multi-class: DBP, RBP, and non-NABP are three independent sigmoid outputs, so a protein can carry more than one label and dual binders emerge naturally instead of needing a fourth class.
• Label-aware attention: Instead of mean- or max-pooling the sequence into a single vector, learned label embeddings act as queries, producing a distinct attended representation per label and letting each class attend to the residues that matter for it.
• Cross-label attention: A masked attention module operating over label representations models DBP–RBP dependency directly, which is what lifts DRBP recall well above the independent-binary-task baselines.
• Invalid-label penalty: The loss adds a term that discourages logically impossible combinations, such as a protein predicted as both DBP and non-NABP.
• Residue-level interpretability: Attention weights project back onto the sequence, highlighting lysine and arginine residues in correctly predicted DNA binders — consistent with known DNA-binding chemistry.

#Technical Details

The backbone is the 150M-parameter ESM-2 checkpoint, chosen over the 8M and 35M variants after comparison and used as a frozen feature extractor — embeddings are precomputed in a separate step, giving an L x 640 matrix per sequence. Downstream sit a single 1D convolution (256 filters, with batch norm, GELU, and dropout), a 4-head self-attention block fused to the CNN output through a gated residual, the label-aware attention layer, and a 2-head cross-label attention layer with its own gated residual, followed by a linear layer and sigmoid. Training used binary cross-entropy at a learning rate of 1e-4, batch size 32, and early stopping on AUC-ROC within a 15-epoch budget, on two NVIDIA A40 GPUs.

Training and evaluation reuse the benchmark splits from iDRBP_MMC: 10,966 training proteins (3,846 DBPs, 2,616 RBPs, 329 DRBPs, 4,175 non-NABPs) with four independent test sets. On TEST474 the model reaches an AUC of 0.98 and 1-AURC of 0.97 for DNA binding and 1-AURC of 0.95 for RNA binding; on the DRBP206 set it reports AUC 0.96, accuracy 0.88, and MCC 0.79, against 0.74/0.69/0.38 for the strongest prior method. On TEST474 it makes seven DRBP calls and six are correct, out of eight true dual binders — against iDRBP_MMC's two correct calls from twenty. Results are mixed rather than uniformly best: RNA-binding AUC on TEST474 (0.90) and DNA-binding AUC on PDB255 (0.80) trail the strongest baselines.

#Applications

The model annotates nucleic-acid-binding function directly from amino acid sequence, useful for prioritizing uncharacterized proteins from genome annotation before committing to CLIP-seq, ChIP-seq, or EMSA validation. Its distinguishing capability is the dual-binder call — DRBPs are enriched in gene-regulatory and disease-associated roles, and the authors' analysis of predicted hits recovers proteins such as FUS, NONO, DDX3X, and EP300 that are implicated in leukemias, ALS, and intellectual disability.

#Impact

LAMP-PRo is a narrow, single-task head rather than a general-purpose foundation model, and most of its representational power comes from the frozen ESM-2 backbone; its contribution is the label-structured attention stack on top. That contribution is measurable where it was designed to matter — dual-binder recognition, where prior multi-label methods reported F1 scores below 0.25. The code, the training and testing scripts, and a trained checkpoint (Full.pt) are on GitHub, which the authors note is more than the closest comparison methods provide, though the repository carries no license file and evaluation to date is entirely in silico on third-party benchmark splits.