Use PALM Online

PALM: Peptide Aggregation Prediction with Protein Language Model Embeddings

PALM (Predicting Aggregation with Language Model embeddings) is a deep-learning model that classifies amyloid-forming peptides and pinpoints aggregation-prone regions at single-residue resolution using transfer learning from pretrained protein language model embeddings. It was developed by Ethan Eschbach, Kristine Deibler, Deepa Korani and Sebastian Swanson of Molecular AI at Novo Nordisk (Lexington, MA, USA) and described in a bioRxiv preprint posted in September 2025.

Amyloid fibrils are filamentous protein aggregates with a cross-beta structure. They are associated with Alzheimer's disease, type-2 diabetes and other conditions, and they complicate the development of biologic drugs by altering physical properties and pharmacodynamics. Models that link amino acid sequence to aggregation are therefore valuable for both disease research and therapeutic design.

Classic sequence-based predictors rely on descriptors such as hydrophobicity, beta-sheet propensity or cross-beta pairing energy, which are proxies for aggregation and cannot easily be improved with new data. Learned models exist, but the main labeled dataset, WaltzDB-2.0, contains only 1,416 hexapeptides, a tiny fraction of therapeutic peptide sequence space. PALM bridges that gap with pretrained ESM2 embeddings.

How PALM Predicts Amyloid Aggregation from Sequence

PALM embeds a peptide with the pretrained ESM2 language model and passes the residue embeddings to an Aggregation Predictor Module (APM) adapted from Light Attention.

1. ESM2 embeddings: the final hidden layer yields a per-residue embedding tensor. The authors compared ESM2 models from 8M to 650M parameters and chose the 8M model based on benchmark performance.

2. Convolutional attention: two independent one-dimensional convolutions (kernel size 5) produce value and attention tensors, combined element-wise after a softmax over sequence length.

3. Residue scoring: a small MLP outputs a residue aggregation score between 0 and 1 for every position.

4. Sequence score: a weighted mean of the residue scores gives the peptide-level score, which makes predictions interpretable down to individual residues.

Residue Padding for Longer Peptides

WaltzDB hexapeptides are much shorter than natural peptides. The authors appended random padding residues at both termini, oversampling each hexapeptide tenfold, so the model sees varied sequence lengths. Non-hydrophobic padding up to 10 residues performed best and significantly improved ROC AUC on Serrano158 and AmyPro22 relative to unpadded training data.

Benchmark Results for PALM Aggregation Prediction

On held-out datasets PALM was compared with TANGO, AggreProt, ANuPP, AggreScan and PASTA using ROC AUC.

PALM scored highest on Serrano158, while PASTA was higher on AmyPro22 residue-level labels. Replacing language model embeddings with one-hot encodings substantially lowered performance on Serrano158, supporting the value of pretrained embeddings.

Single Mutations and Larger Training Data

The authors also tested whether models identify the 13 Abeta42 substitutions known to cause familial Alzheimer's disease by increasing aggregation. PALM trained on WaltzDB-2.0 reached a ROC AUC of 0.51 on this task, effectively failing. Retraining the same architecture on the 100,730-sequence CANYA NNK1-3 dataset raised it to 0.62 (0.70 with one-hot encodings) and reached 0.82 on the NNK4 test set versus 0.72 for the WaltzDB-trained model.

What is Tamarind Bio?

Tamarind Bio is a no-code bioinformatics platform built to give life scientists and researchers access to powerful computational tools. Many cutting-edge machine learning models are hard to deploy and use. Tamarind provides an intuitive, web-based environment that removes the complexity of high-performance computing, software dependencies and command-line interfaces.

The platform is designed for biologists, chemists and other researchers who may not have a background in programming or cloud infrastructure but want to run models on their own data. Key features include:

• A user-friendly graphical interface for setting up and launching experiments

• A robust API for integration into existing research pipelines

• An automated system for managing and scaling computational resources

Tamarind treats information and data security as a top priority, as detailed in its Trust Center and Terms of Service.

Accelerating Discovery with PALM on Tamarind Bio

• Therapeutic peptide screening: flag peptide candidates prone to amyloid formation before costly experimental characterization, since fibrils can alter drug properties.

• Aggregation-prone region mapping: inspect residue-level score profiles, such as those shown for Abeta42 and influenza PB1-F2, to locate the segments that drive aggregation.

• Amyloid research: explore sequence-aggregation relationships for peptides linked to Alzheimer's disease and type-2 diabetes.

• Functional amyloid studies: examine peptides where controlled fibril formation is useful, such as controlled peptide release or unique material properties.

How to Use PALM on Tamarind Bio

1. Log in and open the tool: sign in at tamarind.bio and select PALM.

2. Provide sequences: enter one or more peptide amino acid sequences in standard one-letter code, as the tool accepts them.

3. Choose model options: if exposed, select among the available PALM variants; otherwise use the defaults.

4. Run the job: submit and wait for predictions to finish.

5. Download outputs: retrieve the sequence-level aggregation score and the per-residue scores, if exposed.

6. Interpret scores: scores range from 0 to 1; look for contiguous stretches of high-scoring residues as candidate aggregation-prone regions.

7. Use downstream: prioritize or redesign candidate peptides and confirm results experimentally.

Things to Keep in Mind

• PALM trained on WaltzDB-2.0 failed to identify single mutations that increase Abeta42 aggregation (ROC AUC 0.51).

• Residue scores for Abeta42 mutants looked similar to wild type, so PALM is better suited to classifying peptides than ranking point mutants.

• On the AmyPro22 residue-level benchmark PALM (0.678) did not outperform PASTA (0.707).

• The authors note that hard tasks such as single-mutation effects require more experimental data.

Source: Eschbach E., Deibler K., Korani D., Swanson S. "Predicting peptide aggregation with protein language model embeddings." bioRxiv preprint, September 2025. DOI: 10.1101/2025.09.26.678773 (https://doi.org/10.1101/2025.09.26.678773).