production_ss9_model

Production SS9

A nine-class secondary structure predictor, used in the DSM study to compare the secondary structure of generated sequences with natural ones at finer resolution than the four-class model. It is Synthyra/ESMplusplus_large, the ESMC-600 architecture, with a token classification head.

Classes follow DSSP conventions plus a label for disordered residues: 0 = B (beta bridge), 1 = C (coil or loop), 2 = D (disordered), 3 = E (beta strand), 4 = G (3-10 helix), 5 = H (alpha helix), 6 = I (pi helix), 7 = S (bend), 8 = T (turn). The config carries them as LABEL_0 to LABEL_8.

Predicting secondary structure

import torch

from transformers import AutoModelForTokenClassification, AutoTokenizer

model = AutoModelForTokenClassification.from_pretrained(
    "GleghornLab/production_ss9_model", trust_remote_code=True
).eval()
tokenizer = AutoTokenizer.from_pretrained("GleghornLab/production_ss9_model", trust_remote_code=True)

tokens = tokenizer("MKTAYIAKQRQISFVKSHFSRQ", return_tensors="pt")
with torch.no_grad():
    classes = model(**tokens).logits.argmax(-1)[0]      # (l,)
print("".join("BCDEGHIST"[index] for index in classes[1:-1].tolist()))

Training

Trained on the Proteinea secondary structure training set, validated on CB513 and TS115, and tested on CASP12, CASP13, and CASP14. Base models were first screened with a single transformer block probe over full-residue embeddings, hidden size 512 with rotary embeddings; ESMC-600 was chosen for the production model, balancing performance against throughput, and fine-tuned with LoRA on its attention layers instead of an external probe.

Evaluation

The preprint reports this model's test-set performance as ROC curves with 95% confidence intervals from DeLong's test, in supplementary figure S4b. The screening that chose its base model is in supplementary figure S3.

Links

• Code, demos, and the evaluation pipeline: Gleghorn-Lab/DSM
• Preprint: arXiv:2506.08293
• The Gleghorn Lab and Synthyra
• Terms of use: the LICENSE file in this repository

Cite

@misc{hallee2025diffusionsequencemodelsenhanced,
      title={Diffusion Sequence Models for Enhanced Protein Representation and Generation},
      author={Logan Hallee and Nikolaos Rafailidis and David B. Bichara and Jason P. Gleghorn},
      year={2025},
      eprint={2506.08293},
      archivePrefix={arXiv},
      primaryClass={q-bio.BM},
      url={https://arxiv.org/abs/2506.08293},
}

---

The DSM project

What follows is the project README, shared with github.com/Gleghorn-Lab/DSM.

Table of Contents

• Introduction
• Models
• Usage
• Demos
• Local installation
• Training
• Evaluation
• Results
• Experimental validation
• Cite

Introduction

DSM (Diffusion Sequence Model) is a novel Protein Language Model (pLM) developed in collaboration between the Gleghorn Lab and Synthyra. It was trained with masked diffusion to enable both high-quality representation learning and generative protein design. This repository contains the code for training, evaluating, and applying DSM and its variants.

DSM is capable of generating diverse, biomimetic sequences that align with expected amino acid compositions, secondary structures, and predicted functions. Furthermore, DSM's learned representations match or exceed those of comparably sized pLMs on various downstream tasks. DSM is detailed extensively in our preprint (which is currently in review). Beyond the base and PPI variants, we are currently training versions to jointly diffuse over sequence and foldseek tokens, as well as Annotation Vocabulary tokens. Since the preprint release, Synthyra has trained Synthyra/DSM_ppi_full which neglects the LoRA PPI training in favor for full finetuning. Additionally, the sequences SeqA and SeqB are jointly masked instead of just SeqB in the original version. We plan on adding the many new results to the second version of the preprint and eventual journal article.

Models

Relevant Huggingface hosted models and datasets

• Base DSM Models:

  • GleghornLab/DSM_150 - 150M parameter DSM model
  • GleghornLab/DSM_650 - 650M parameter DSM model

• DSM-ppi Models: (LoRA versions - results reported in paper but not recommended for real use)

  • GleghornLab/DSM_150_ppi_lora - 150M parameter LoRA DSM-ppi model
  • GleghornLab/DSM_650_ppi_lora - 650M parameter LoRA DSM-ppi model
  • GleghornLab/DSM_150_ppi_control - Control version of LoRA DSM-ppi

  (Fully finetuned - recommended for real use)

  • Synthyra/DSM_ppi_full - 650M parameter DSM-ppi model

• Datasets:

  • Synthyra/omg_prot50 - Open MetaGenomic dataset clustered at 50% identity (207M sequences)
  • GleghornLab/stringv12_modelorgs_9090 - STRING database model organisms (653k sequences)

• Utility Models:

  • GleghornLab/production_ss4_model - Secondary structure prediction (4-class)
  • GleghornLab/production_ss9_model - Secondary structure prediction (9-class)

Usage

This section outlines how to use a trained DSM model for common generation tasks. The core generation logic is provided by the GenerateMixin class, used by DSM models.

First, ensure you have a trained model (either one you trained or a pre-trained one from Hugging Face Hub) and the necessary environment set up.

import torch
from models.modeling_dsm import DSM # Or DSM_ppi for binder generation

# Load a pre-trained model
model_name_or_path = "GleghornLab/DSM_650" # Replace with your model of choice
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
model = DSM.from_pretrained(model_name_or_path).to(device).eval()
tokenizer = model.tokenizer

You are using a model of type esm_diff to instantiate a model of type dsm. This is not supported for all configurations of models and can yield errors.

This warning is normal - all good!

1. Unconditional Sequence Generation

To generate a novel sequence of a specific length. DSM uses a progressive denoising approach.

### Unconditional generation
length = 100
mask_token = tokenizer.mask_token
# optionally, enforce starting with methionine
input_tokens = tokenizer.encode('M' + ''.join([mask_token] * (length - 1)), add_special_tokens=True, return_tensors='pt').to(device)
output = model.mask_diffusion_generate(
    tokenizer=tokenizer,
    input_tokens=input_tokens,
    step_divisor=100,          # lower is slower but better
    temperature=1.0,           # sampling temperature
    remasking="random",        # strategy for remasking tokens not kept
    preview=False,             # set this to True to watch the mask tokens get rilled in real time
    slow=False,                # adds a small delay to the real time filling (because it is usually very fast and watching carefully is hard!)
    return_trajectory=False    # set this to True to return the trajectory of the generation (what you watch in the preview)
) # Note: output will be a tuple if return_trajectory is True

generated_sequences = model.decode_output(output)
print(f"Generated sequence: {generated_sequences[0]}")

Generated sequence: MFRVDALQVAQQETLAIGRSTAYDKQESPSMAQRQVLTQLAAYGGENDLRQICIPAERRNFLSIANGASYQFVEEDNEANGGYWSPHKAGLPESACKRFI

2. Mask Filling (Inpainting)

To fill in masked regions of a template sequence:

# Mask Filling / Inpainting
template_sequence = "MA<mask><mask><mask>KEG<mask><mask>STL"
input_tokens = tokenizer.encode(template_sequence, add_special_tokens=True, return_tensors='pt').to(device)

output = model.mask_diffusion_generate(
    tokenizer=tokenizer,
    input_tokens=input_tokens,
    step_divisor=100,          # lower is slower but better
    temperature=1.0,           # sampling temperature
    remasking="random",        # strategy for remasking tokens not kept
    preview=False,             # set this to True to watch the mask tokens get rilled in real time
    slow=False,                # adds a small delay to the real time filling (because it is usually very fast and watching carefully is hard!)
    return_trajectory=False    # set this to True to return the trajectory of the generation (what you watch in the preview)
) # Note: output will be a tuple if return_trajectory is True

generated_sequences = model.decode_output(output)
print(f"Generated sequence: {generated_sequences[0]}")

Generated sequence: MAVKFKEGGISTL

3. Conditional Generation (e.g., Binders - using DSM-ppi)

# from models.modeling_dsm import DSM_ppi
# model_binder = DSM_ppi.from_pretrained("GleghornLab/DSM_650_ppi_lora").to(device).eval()
# The lora version from the paper leads to unreliable outputs
# Synthyra has generously trained a version through full fine tuning

model = DSM.from_pretrained("Synthyra/DSM_ppi_full").to(device).eval()

# BBF-14
target_seq = "MGTPLWALLGGPWRGTATYEDGTKVTLDYRYTRVSPDRLRADVTYTTPDGTTLEATVDLWKDANGVIRYHATYPDGTSADGTLTQLDADTLLATGTYDDGTKYTVTLTRVAPGSGWHHHHHH"
# For binder generation, the 'interactor' (SeqB) part is what gets generated/filled.
# Start with a fully masked interactor of desired length.
interactor_template_len = 256
interactor_template = ''.join([mask_token] * interactor_template_len)

combined_input_str = target_seq + '<eos>' + interactor_template

input_tokens = tokenizer.encode(combined_input_str, add_special_tokens=True, return_tensors='pt').to(device)

output = model.mask_diffusion_generate(
    tokenizer=tokenizer,
    input_tokens=input_tokens,
    step_divisor=100,          # lower is slower but better
    temperature=1.0,           # sampling temperature
    remasking="random",        # strategy for remasking tokens not kept
    preview=False,             # set this to True to watch the mask tokens get rilled in real time
    slow=False,                # adds a small delay to the real time filling (because it is usually very fast and watching carefully is hard!)
    return_trajectory=False    # set this to True to return the trajectory of the generation (what you watch in the preview)
) # Note: output will be a tuple if return_trajectory is True

target, binder = model.decode_dual_input(output, seperator='<eos>')
# Parse out the generated interactor part based on EOS tokens.
# Example: generated_full_seq_str.split(model_binder.tokenizer.eos_token)[1]
print(f"Generated binder {binder[0]}")

Generated binder HRHHHRRPTHARETEWLARMRLGIAEHQRIAVPRSDLEPDQMRERAADNQRLVKEYDQVIDHQTEGSTERLFEVLRVWEQVNTEQAHHEASAALEFGRVGYPDDEGGRAFYTQANAHKKDLVEYIGGIDEDAKWDPRIAWLMPEGGQPVKATVIGVSEERINGLKVLDDHWGRERRLWLINLFTALQAYDDPTRPTQVTLTPATDQLTNDVQYLLLSTRYTPPGVTTAVKIRKLDGRTLKVLTTEAPYVVRGATLS

Folded with Chai1:

Synthyra/DSM_ppi_full was actually trained to fill masks from any part of SeqA and SeqB. That means you can fully hallucinate plausibly interacting protein pairs.

seq_a_length = 128
seq_b_length = 128

seq_a_template = ''.join([mask_token] * seq_a_length)
seq_b_template = ''.join([mask_token] * seq_b_length)

combined_input_str = seq_a_template + '<eos>' + seq_b_template

input_tokens = tokenizer.encode(combined_input_str, add_special_tokens=True, return_tensors='pt').to(device)

output = model.mask_diffusion_generate(
    tokenizer=tokenizer,
    input_tokens=input_tokens,
    step_divisor=10,          # lower is slower but better
    temperature=1.0,           # sampling temperature
    remasking="random",        # strategy for remasking tokens not kept
    preview=False,             # set this to True to watch the mask tokens get rilled in real time
    slow=False,                # adds a small delay to the real time filling (because it is usually very fast and watching carefully is hard!)
    return_trajectory=False    # set this to True to return the trajectory of the generation (what you watch in the preview)
) # Note: output will be a tuple if return_trajectory is True

seqa, seqb = model.decode_dual_input(output, seperator='<eos>')
# Parse out the generated interactor part based on EOS tokens.
# Example: generated_full_seq_str.split(model_binder.tokenizer.eos_token)[1]
print(f"SeqA: {seqa[0][5:]}") # remove cls token
print(f"SeqB: {seqb[0]}")

SeqA: MVNLAKMRQRTEQNLREVSSFVKILFHTVLKFPMKINIGIHVHINMQAAQNAAADQNMQATNVIDLHNFKMGKDIGVDNKASATAHIYDEAHHTFLQLGAIKLLHAIPMIAGPVRCRLPIGFGHRFRG
SeqB: HYKNPMHSLLDSNVLHKDVVEVRLPIKIGMELDVMASAMREFLMPGTQQGDLRVIAEKRPVNKLHTYRRDLVKLLLAGAKLGTEAKSVELDLYRTELGGLVVYIININIATWDIIFAKVKICRGNDKP

Folded with Chai1:

Demos

There are various demos with many more to come. For example, in demo_dsm_ppi_full.py (run by python -m demos.demo_dsm_ppi_full) we perform a test on DSM-ppi. We take 1000 protein pairs from BIOGRID (real protein-protein interactions) and 1000 from Negatome (non interacting protein pairs) and mask the second sequence (SeqB) by 50%. This acts as a sanity check, as we expect the accuracy on reconstructing real positive PPIs to be higher than the accuracy on non-interacting proteins. Indeed, this is the case:

==================================================
RESULTS COMPARISON
==================================================
Positive examples:
  Mean accuracy: 0.495 ± 0.322
  Processed:     1000 examples

Negative examples:
  Mean accuracy: 0.227 ± 0.231
  Processed:     1000 examples

Difference (Positive - Negative): 0.267
T-test: t=21.331, p=0.000
Difference is statistically significant (p < 0.05)

Installation

1. Clone the repository:

   git clone https://github.com/Gleghorn-Lab/DSM.git
   cd DSM
   

2. Initialize the submodules:

   git submodule update --init --remote --recursive
   

3. Set up the Python virtual environment: The setup_bioenv.sh script creates a virtual environment named bioenv in your home directory (~/bioenv), installs PyTorch with CUDA 12.6 support, and then installs all other dependencies from requirements.txt.

   Make the script executable:

   chmod +x setup_bioenv.sh
   

   Run the script:

   ./setup_bioenv.sh
   

   If you are not on a linux machine, you can install the requirements directly

   python -m pip install -r requirements.txt
   

4. Activate the environment: Each time you want to work on this project, activate the virtual environment:

   source ~/bioenv/bin/activate
   

5. To deactivate the environment:

   deactivate
   

All together

git clone https://github.com/Gleghorn-Lab/DSM.git
cd DSM
git submodule update --init --remote --recursive
chmod +x setup_bioenv.sh
./setup_bioenv.sh
source ~/bioenv/bin/activate

Logging in

Log in once, and the scripts find your credentials themselves:

hf auth login   # Hugging Face; or set HF_TOKEN
wandb login     # Weights & Biases; or set WANDB_API_KEY

Training pushes the trained model to the Hugging Face Hub and logs to Weights & Biases, so it needs both. Reading a private model or dataset needs the Hugging Face login.

Training

The primary script for training models is training/train_dsm.py. This script further pretrains an ESM2 checkpoint using the DSM objective (masked diffusion based on LLaDA) on a large protein sequence dataset like OMG-prot50.

Main Training Script: train_dsm.py

• Base Model: DSM models are extended from pre-trained ESM2 checkpoints (e.g., ESM2-150M, ESM2-650M).
• Training Objective: Masked diffusion loss, where the model predicts masked tokens. The loss is scaled by 1/(t + epsilon) where t is the corruption level, penalizing errors more at low mask rates.
• Language Modeling Head: Uses a modified head with a soft-logit cap (tau=30) and tied output projection weights to the token embeddings.
• Data Handling:
  • Training data can be streamed from datasets like Synthyra/omg_prot50 (a version of Open MetaGenomic dataset clustered at 50% identity).
  • Uses data.dataset_classes.SequenceDatasetFromList for validation/test sets and data.dataset_classes.IterableDatasetFromHF for streaming training.
  • data.data_collators.SequenceCollator is used for batching.
• Training Process:
  • Utilizes Hugging Face TrainingArguments.
  • A custom IterableTrainer (from training.iterable_trainer.py) handles iterable datasets.
  • Uses AdamW optimizer and a cosine learning rate scheduler with linear warmup.
  • Supports logging to Weights & Biases (wandb).
  • The trained model can be pushed to Hugging Face Hub.
  • Example checkpoints mentioned in the paper: DSM-150 (from ESM2-150M, 100k steps, batch 32, seqlen 512, LR 1e-4) and DSM-650 (from ESM2-650M, 100k steps, global batch 128, seqlen 2048, LR 1e-4).

Usage Example:

python -m training.train_dsm \
    --model_path facebook/esm2_t33_650M_UR50D \
    --save_path GleghornLab/DSM_650 \
    --lr 1e-4 \
    --batch_size 8 \
    --grad_accum 16 \
    --max_steps 100000 \
    --save_every 1000 \
    --fp16 \
    --wandb_project "DSM_Training"

Key Command-Line Arguments for train_dsm.py:

• --model_path: Path to the base ESM2 model to start from.
• --save_path: Path to save the trained DSM model on Hugging Face Hub.
• --lr: Learning rate.
• --batch_size: Batch size per device.
• --grad_accum: Gradient accumulation steps.
• --max_steps: Maximum training steps.
• --wandb_project: Wandb project name (default: DSM).
• --max_length: Maximum sequence length.
• --save_every: Save model and evaluate every N steps.
• --fp16: Enable mixed-precision training.
• --bugfix: Use small batch size and max length for debugging.

Other Training Scripts (e.g., for DSM-ppi)

The training/ directory may also contain scripts like train_dsm_bind.py.

• DSM-ppi (e.g., DSM-150-ppi, DSM-650-ppi) is fine-tuned on PPI datasets.
• Training involves conditioning on a target sequence (SeqA) to generate an interactor (SeqB) using the format [CLS]--SeqA--[EOS]--[MASKED~SeqB]--[EOS].
• LoRA (Low-Rank Adaptation) can be applied to attention layers for efficient fine-tuning.

And training/iterable_trainer.py provides the get_iterable_trainer function used by train_dsm.py to enable training with iterable datasets.

Evaluation

The repository includes a comprehensive suite for evaluating model performance, focusing on:

1. Sequence Reconstruction (Mask Filling):

   • Evaluated by masking validation/test sets at various corruption rates (5% to 90%) and measuring cross-entropy loss, weighted F1 score, and Alignment Score (ASc) for the masked positions.
   • The script evaluation/mask_filling.py is central to this.

2. Unconditional Generation Quality:

   • Generate a corpus of sequences based on lengths from a reference set (e.g., validation data).
   • Compare distributions (1-mers, 2-mers, 3-mers) of amino acids and predicted secondary structures between generated and natural sequences using χ² test and Jensen-Shannon (JS) divergence.
   • Compare distributions of predicted functional annotations (e.g., using Annotation Vocabulary - AV terms).
   • Scripts involved: evaluation/unconditional_generation_tuning.py (to find optimal generation parameters like temperature and step divisor s), evaluation/unconditional_generation.py, evaluation/ss_pred.py (using production_ss4_model or production_ss9_model), evaluation/annotate_comparisons.py, evaluation/compare_distributions.py, evaluation/plot_distribution_comparisons.py.
   • The run_eval_pipeline.py script automates this workflow.

3. Representation Quality (Model Probing):

   • Evaluate learned embeddings by training linear probes (or simple transformer blocks) on various downstream tasks (e.g., secondary structure prediction, localization prediction, etc.).
   • Performance is compared against random vectors, randomized transformers, and other established pLMs.
   • The assessment was done with Protify, an open-source framework that can be used for pLM training and evaluation.

4. Conditional Generation (Binder Design for DSM-ppi):

   • Evaluate DSM-ppi on benchmarks like BenchBB.
   • Generate binders for target proteins using template-based masking strategies.
   • Assess generated binders using in-silico tools like Synteract2 for predicted binding affinity (ppKd).

The evaluation/ directory also contains a readme.md which provides further details on some evaluation workflows. Key metrics used include:

• Alignment Score (ASc): A normalized Needleman-Wunsch global alignment score (using BLOSUM62) to measure sequence similarity, robust to length variations. ASc(a, b) = l/(f(a, a) - f(a, b) + l).
• Jensen-Shannon (JS) Divergence: To compare distributions of k-mers and functional terms.

Functional annotations (evaluation/annotate_comparisons.py) and binding-affinity predictions (evaluation/conditional_binder.py, evaluation/unconditional_binder.py) come from the Synthyra API, and the scripts read its key from the SYNTHYRA_API_KEY environment variable. The pipeline below needs it for its annotation step. The binder scripts' --test flag substitutes random scores and needs no key.

Running the Full Unconditional Evaluation Pipeline:

python run_eval_pipeline.py --data_dir ./evaluation_results

Refer to run_eval_pipeline.py --help for more options, such as --skip_tuning.

Mask Filling Evaluation

The script evaluation/mask_filling.py is used to evaluate models on their ability to predict masked tokens in a sequence across various masking rates.

• Functionality:

  • Evaluates different models (DSM, DPLM, standard ESM models).
  • Tests across multiple datasets (Synthyra/omg_prot50, GleghornLab/stringv12_modelorgs_9090).
  • Calculates metrics: loss, perplexity, precision, recall, F1, accuracy, MCC, and alignment score.
  • Saves detailed results to CSV files.
  • Can generate a summary plot comparing model performance across different mask rates using evaluation/plot_mask_fill_results.py.

• Usage Example:

  python -m evaluation.mask_filling \
      --batch_size 4 \
      --mask_rates 0.15 0.30 0.50 \
      --data_splits valid test \
      --results_dir ./results/mask_fill_custom
  

  To generate a comparison plot from existing results:

  python -m evaluation.mask_filling --generate_comparison_plot --results_dir ./results/mask_fill_custom --plot_output ./results/mask_fill_custom/comparison.png
  

Other Evaluation Scripts

The evaluation/ directory contains additional scripts for more specific analyses. These are typically run independently:

• evaluation/all_targets_uncond.py and evaluation/all_targets_cond.py: Likely for evaluating generation towards specific targets, unconditionally and conditionally.
• evaluation/conditional_binder.py and evaluation/unconditional_binder.py: Suggest evaluation focused on generating protein binders.
• evaluation/unconditional_by_length.py: May evaluate unconditional generation focusing on sequence length distributions.
• evaluation/utils.py: Utility functions for evaluation scripts.

Users should refer to individual scripts (e.g., using python -m evaluation.<script_name> --help) for their specific usage and arguments. The evaluation/ directory also contains a readme.md which provides further details on the unconditional generation evaluation workflow.

Results

DSM demonstrates strong performance in both protein sequence generation and representation learning, establishing masked diffusion as a powerful paradigm.

• Biomimetic Sequence Generation: Unconditionally generated DSM sequences closely mimic natural protein distributions in terms of amino acid k-mers, predicted secondary structures (JS divergence < 0.01 for AA k-mers), and predicted functional annotations (AV terms, JS divergence ~0.1). This suggests DSM captures underlying biological principles.

• Superior Sequence Reconstruction: DSM models significantly outperform MLM-based ESM2 models in reconstructing sequences from highly corrupted inputs (up to 90% masking).

  • At 90% masking, DSM achieves an Alignment Score (ASc) of ~0.27, considerably higher than random.
  • DSM models show higher F1 scores in reconstruction tasks compared to DPLM models, especially at high mask rates.

• High-Quality Embeddings: DSM embeddings match or exceed the quality of those from comparably sized pLMs (ESM2, DPLM) and even larger autoregressive models (ProtCLM 1B) on various downstream tasks evaluated by linear probing. DSM-650 generally provides the best representations among tested models of similar size.

• Effective Binder Design (DSM-ppi):

  • DSM-ppi fine-tuned on protein-protein interaction data, demonstrates the ability to generate protein binders conditioned on target sequences.
  • On the BenchBB benchmark, DSM-generated binders (both unconditional DSM and conditional DSM-ppi) show promising predicted binding affinities, in some cases superior to known binders. For example, designs for EGFR showed high predicted pKd and good structural metrics (ipTM, pTM with AlphaFold3).

• Efficiency: DSM can generate realistic protein sequences from a single forward pass during reconstruction tasks at high mask rates, offering potential efficiency advantages over iterative AR or some discrete diffusion models.

These results highlight DSM's capability to unify high-quality protein representation learning and biologically coherent generative modeling within a single framework.

Experimental validation

We validated various DSM generated binders using biolayer interferometry through Adaptyv Bio - sending in 20 designs for EGFR and PD-L1. Sequences generated via unconditional generation were sorted hierarchically by predicted binding affinity (Synteract2), then ESMfold pLDDT, ESM2 PLL, and finally Chai1 iPTM.

EGFR

Of the 13 expressed, 12 designs bound to EGFR with 11 of them binding strongly. Noteably, the top design, dsm_egfr_10, presented a mean KD in the picomolar range (861 pM), which is a ~30% increase in binding affinity vs. the winner (and our starting template) of the Adaptyv EGFR competition at 1.21 nM, and ~90% over the original starting scFV Cetuximab at 664 nM.

<img width="1593" height="527" alt="image" src="https://github.com/user-attachments/assets/f6f6a614-d5f4-4e9e-b7c9-fc0fdf9dc1e2" />

dsm_egfr_10

QVQLQQSGPGLVQPSQSLSITCTVSGFSLTNYGVHWVRQSPGKGLEWLGVIWSGGNTDYNTPFTSRLSISRDTSKSQVFFKMNSLQTDDTAVYYCARALTYYDYEFAYWGQGTLVTVSAGGGGSGGGGSGGGGSDILLTQSPVILSVSPGERVSFSCRASQSIGSNIHWYQQRTNGSPKLLIRYASESISGIPSRFSGSGSGTDFTLSINSVDPEDIADYYCQQNNNWPTTFGAGTKLEIK

• KD - 861 pM
• pKD - 9.06
• PPI probability (Synteract2) - 0.9991
• predicted pKD (Synteract2) - 8.94
• mask rate for DSM - 2%
• mutations from template - 3
• mutations - I92V, T165S, L240I
• average esm2 pll - 1.23
• esmfold plddt - 0.75
• pTM (AlphaFold3) - 0.81
• ipTM (AlphaFold3) - 0.91

<table> <tr> <td> <img width="596" height="355" alt="image" src="https://github.com/user-attachments/assets/0c6f690d-3134-44d2-9540-12c277e187b3" /> </td> <td> <img src="https://github.com/Gleghorn-Lab/DSM/blob/main/wetlab_result_analysis/egfr/kinetics/dsm_egfr_10_2.png" width="400"> </td> </tr> </table>

This sequence would have won the EGFR competition by a wide margin! Of course, we piggybacked on the winning entry as our template. We have been using DSM-PPI-full to attempt to replicate competition winning binders from the Cetuximab starting point instead. Stay tuned!

<img width="1865" height="548" alt="image" src="https://github.com/user-attachments/assets/ce486326-f4ba-4604-af47-259f7bbe496f" />

PD-L1

All 20 PD-L1 designs had high expression rates, with 15/20 binding. 1 weak, 10 medium, and 3 strong. The strongest presented with an average KD of 8.06 nM (pKD) , which is markedly less than the original template at 0.8 pM. We attribute the consistent binding but worse performance overall to the higher error between Synteract2 ppKD and true pKD of the template, implying it is not modeled well by our affinity system.

<img width="1592" height="516" alt="image" src="https://github.com/user-attachments/assets/6d2cde0e-75a4-4f29-999f-a8c601286845" /> <img src="https://github.com/Gleghorn-Lab/DSM/blob/main/wetlab_result_analysis/pdl1/kinetics/dsm_pdl1_7_1.png" width="400">

Cite

@misc{hallee2025diffusionsequencemodelsenhanced,
      title={Diffusion Sequence Models for Enhanced Protein Representation and Generation}, 
      author={Logan Hallee and Nikolaos Rafailidis and David B. Bichara and Jason P. Gleghorn},
      year={2025},
      eprint={2506.08293},
      archivePrefix={arXiv},
      primaryClass={q-bio.BM},
      url={https://arxiv.org/abs/2506.08293}, 
}