Loading…
Peptide-focused instruction-tuned LLM that describes function, designs sequences, predicts eight bioactivity properties and edits physicochemistry.
No providers recorded yet. Browse all providers
A twelve-residue antimicrobial peptide carries almost no evolutionary signal to read. There is no deep alignment behind it and no conserved fold to recognise; what decides whether it works is net charge, a hydrophobic face, and a handful of local residue contacts. Protein language models such as ESM-2 are built on precisely the signal that short peptides lack — co-evolutionary statistics gathered across full-length proteins — which is why their representations transfer unevenly to peptide bioactivity.
Pep-Instruction answers that gap from the other direction. Rather than pretraining a new sequence encoder, it treats amino acid strings as ordinary text and adapts an instruction-following chat model to read and write them, using natural-language supervision in place of structural or evolutionary priors. The peptide knowledge arrives entirely through supervised instructions — predict this peptide's function, design a sequence that does X, raise its net charge — so a new capability is a new instruction template, not a new architecture or output head.
Kaijun Yang, Tianxiang Wu, Wenbo Zhang and Pengyong Li at Xidian University built the model for a 2026 Bioinformatics paper. Its training corpus, Pep-Instructions, is the artifact the paper foregrounds and the name the released weights inherited; the model itself is that corpus applied to Llama-2-7b-chat through low-rank adaptation. That construction — a general chat LLM given a biological vocabulary by LoRA rather than by pretraining — is the same one behind Protein-Llama-3-8B and ProLLaMA, though here the adaptation is a single instruction-tuning pass with no continual pretraining stage.
The adapter is rank 64 with alpha 128 and 0.1 dropout, applied to all seven projection matrices — query, key, value, output, gate, up and down — in each of Llama-2-7b-chat's 32 layers. That is 159,907,840 trainable parameters against a frozen 6.74B backbone, roughly 2.4% of the served model's 6.9B total. Training runs once over the pooled corpus for seven epochs at a 2e-4 cosine-decayed learning rate with 5% warmup, an effective batch of 128 and a 256-token sequence limit. Pep-Instructions itself holds 482,639 instruction–response pairs — 386,037 for training, 48,291 for validation, 48,311 held out — with physicochemical optimization the largest task family, followed by sequence design, function description and property prediction. Evaluation is metric-per-task: ROUGE-L, BLEU and BERTScore for functional descriptions; global alignment identity plus a BLOSUM80-weighted similarity to the reference peptide for design; F1 for the property calls; and validity and directional-improvement rates computed from the peptidy descriptor library for optimization. Baselines are the unmodified base model, Galactica-6.7B and GPT-3.5-turbo, and backbone agreement for designed sequences is spot-checked with ESMFold.
The practical target is early peptide triage, where the candidate list is long and synthesis is the bottleneck. A group with a lead sequence can ask for a functional annotation, screen it against the eight bioactivity properties, request variants biased toward higher charge or lower instability, and read the answers back as text, without assembling a separate predictor per property. The instruction format also makes the corpus reusable as a benchmark: the same held-out splits score any other language model on identical peptide tasks.
Pep-Instruction is a compact demonstration that peptide-specific competence can be bought with curated instructions rather than large-scale biological pretraining, and its corpus is the piece most likely to outlive it — a shared benchmark that lets peptide tasks be compared across general-purpose models. The claims come with real limits. Every result is in silico, with design quality judged by sequence similarity to a reference rather than by assay; the zero-shot pipeline additionally depends on a second, unreleased discriminator adapter, so those numbers cannot be reproduced from the released artifacts alone. Neither the code repository nor the weights carry a license, and because the adapter is a derivative of Llama-2-7b-chat, the base model's community license governs any use regardless.
Much of this page is generated or calculated automatically. Flag anything that looks off and we will re-run it.