Loading…
Parameter-efficient peptide representation model whose BiLSTM autoencoder keeps latent axes aligned with physicochemical and DFT-derived descriptors.
No providers recorded yet. Browse all providers
A protein language model can tell you that one peptide is more likely to be antimicrobial than another, but it cannot tell you that the reason is a run of cationic residues and a hydrophobic face. Its embedding dimensions are learned coordinates with no physical referent, and recovering meaning from them requires a separate round of probing or sparse-autoencoder analysis. dcBiLSTM-AE approaches the problem from the other end: it never treats the sequence as text. Each peptide enters the model as a matrix whose rows are 46 per-residue descriptors — 43 RDKit physicochemical properties plus three density functional theory (DFT) derived interaction energies with Ca²⁺, Mg²⁺ and Ba²⁺ cations — and whose columns are positions along the chain. Every input cell carries units, and the training objective is arranged so that the latent dimensions keep that meaning.
Two additions to an ordinary bidirectional LSTM autoencoder do the work. A soft weight matrix non-diagonality penalty, paired with forced diagonal initialization, discourages the encoder from blending input features into one another, so a latent coordinate stays tethered to the descriptor it began as; Jacobian analysis confirms that alignment survives training. An InfoNCE contrastive loss, whose positive and negative pairs come from MiniBatchKMeans clusters of one-hot-encoded sequences, then organizes the latent space semantically, so sequence-similar peptides sit together. Both ablation stages are reported alongside the full model — BiLSTM-AE with neither term, cBiLSTM-AE with the contrastive loss alone — and each addition raises benchmark performance rather than trading accuracy away for interpretability.
The model was built at ITMO University's Center for AI in Chemistry in Saint Petersburg by Evgeniy Nam, Yevgeniya Din and Nikita Serov, first released as a September 2025 conference submission and published in the Journal of Chemical Information and Modeling in 2026. The group's codebase and released embedding files carry the project name SeQuant.
Pretraining uses 155,920 unique unlabeled peptide sequences sampled from NCBI by cluster-stratified sampling, with a separate scaling experiment on a 467,792-sequence corpus. Evaluation is deliberately disjoint from that corpus: the published version reports a suite of eight public benchmarks — antimicrobial, anti-inflammatory, antidiabetic, antioxidative and hemolytic activity, nonfouling behavior and solubility as classification tasks, plus regression on minimum inhibitory concentration against E. coli — scored by 5-fold cross-validation over frozen embeddings. The comparison set is five larger protein and peptide language models: ProtT5, ProtBERT, Ankh, ESM Cambrian and PeptideCLM. The results the authors report are competitive rather than dominant; in the earlier conference version, scored against ProtBERT, Ankh-large and ESM-C on four of these benchmarks, the model led on anti-inflammatory peptide prediction and ranked second to Ankh-large on cross-task stability by Matthews correlation coefficient. It does not match ProtT5-3B on raw accuracy; the claim the authors make is stability across tasks at four orders of magnitude less capacity, plus a latent space that can be interrogated directly.
The intended user is a domain scientist building a predictor for a specific peptide property on a few thousand labeled examples — antimicrobial screening, hemolysis triage, solubility filtering — where a billion-parameter encoder is impractical and an uninterpretable feature vector is hard to defend to experimentalists. Because the embedding is fixed and cheap to compute, it also suits large virtual library screens and structure-property analyses where the point is to identify which physicochemical factors drive an observed activity, not only to rank candidates.
dcBiLSTM-AE is a concrete counterexample to the assumption that peptide representation learning has to scale with parameter count, and it does so without the usual interpretability tax. Its practical reach is bounded: pretrained weights are not distributed, so users must retrain from the released code and data, neither of the project's two GitHub repositories carries a license at all, and the padded 96-residue input caps the peptide length the encoder accepts. The authors are explicit that scaling the approach to longer chains, noncanonical residues and branched or cyclic topologies remains open work.
Much of this page is generated or calculated automatically. Flag anything that looks off and we will re-run it.