Related Experiment Video
Updated: May 14, 2026
08:48
Development of a Backbone Cyclic Peptide Library as Potential Antiparasitic Therapeutics Using Microwave Irradiation
Published on: January 26, 2016
12.4K
Scaling SMILES-based chemical language models for therapeutic peptide engineering.
Aaron L Feller1,2, Maxim Secor2, Sebastian Swanson2
1Integrative Biology, The University of Texas at Austin, 2500 Speedway, Austin, TX 78712, USA.
Biorxiv : the Preprint Server for Biology
|January 16, 2026
Summary
We developed PeptideMTR, a new suite of language models for therapeutic peptide engineering. These models effectively predict peptide properties, outperforming existing methods and offering a valuable resource for drug discovery.
Area of Science:
- Biochemistry
- Computational Chemistry
- Drug Discovery
Background:
- Therapeutic peptides bridge protein modularity and small molecule versatility but are challenging for current foundation models.
- Protein models are limited to canonical residues, while small molecule models lack peptide sequence context.
Purpose of the Study:
- Introduce PeptideMTR, a suite of nine SMILES-based chemical language models.
- Pretrain models on peptide and small molecule data using masked language modeling and multi-task regression.
- Evaluate model performance on diverse therapeutic peptide engineering tasks.
Main Methods:
- Developed nine SMILES-based chemical language models (32M-337M parameters).
- Pretrained models on a combined dataset of peptides and small molecules.
- Employed masked language modeling and multi-task regression to physicochemical descriptors.
- Evaluated on membrane permeability, aggregation propensity, tumor homing, cell penetration, and antimicrobial activity prediction.
Main Results:
- Identified a scaling transition in model performance based on size and pretraining objectives.
- Descriptor-guided pretraining is crucial at smaller scales for grounding embeddings.
- Large-scale self-supervised models spontaneously learn physicochemical priors.
- PeptideMTR demonstrated superior performance compared to molecular fingerprints and specialized architectures.
Conclusions:
- PeptideMTR provides a powerful and scalable resource for therapeutic peptide engineering.
- The models advance the representation and prediction of peptide behavior.
- This work facilitates the design and development of novel peptide-based therapeutics.
Related Concept Videos
02:00Molecular Models
Physical models representing molecular architectures of chemical compounds play essential roles in understanding chemistry. The use of molecular models makes it easier to visualize the structures and shapes of atoms and molecules.
01:19MALDI-TOF Mass Spectrometry
Mass spectrometry is a powerful characterization technique that can identify and separate a wide variety of compounds ranging from chemical to biological entities, based on their mass-to-charge ratio (m/z). The instruments that allow this detection, known as mass spectrometers, have three components: an ion source, a mass analyzer, and a detector. These spectrometers differ based on the nature of their ion source and analyzers.Matrix-assisted laser desorption ionization (MALDI) is a commonly...