Hermes: Large DEL Datasets Train Generalizable Protein-Ligand Binding Prediction Models

The paper introduces Hermes, a lightweight transformer model trained exclusively on large-scale, diverse DNA-encoded library (DEL) datasets that successfully generalizes to predict protein-ligand binding across novel targets and chemical scaffolds, demonstrating that unified DEL data can overcome the biases of traditional public affinity datasets for effective virtual screening.

The Gist

Imagine you are trying to teach a robot how to find the perfect key for a specific lock. In the world of medicine, the "lock" is a protein inside your body (often a culprit in diseases), and the "key" is a drug molecule. If the key fits the lock perfectly, it can stop the disease. This is called Protein-Ligand Binding.

For a long time, teaching robots to do this has been like trying to learn a language by reading thousands of different dictionaries written by different people, in different languages, with different spellings. It's messy, inconsistent, and full of errors.

Enter Hermes, a new AI model created by a team at Leash Biosciences. Here is the story of how Hermes learned to be a master key-finder, explained simply.

1. The Problem: The "Noisy Library"

Traditionally, scientists taught AI models using data from public databases. These databases are like a giant library where millions of people have dropped off notes about which keys fit which locks.

• The Issue: One scientist wrote a note in 1990 using a specific test; another wrote one in 2020 using a totally different test. The notes are inconsistent, biased, and often wrong.
• The Result: AI models trained on this "noisy library" struggle to generalize. They memorize the specific notes but fail when they see a new lock or a new type of key.

2. The Solution: The "Massive, Unified Factory"

The authors decided to stop reading the messy library and instead built their own massive, perfectly organized factory. This factory uses a technology called DEL (DNA-Encoded Libraries).

• How it works: Imagine you have a library of 6.5 million tiny keys (chemical compounds). Each key has a unique DNA barcode attached to it, like a serial number.
• The Test: You dump all 6.5 million keys into a pool with a specific protein lock. You wash away everything that doesn't stick. You then scan the DNA barcodes of the keys that did stick.
• The Advantage: Because this is done in one giant, automated experiment with the same rules every time, the data is clean, consistent, and huge. It's like having a factory that tests billions of keys against hundreds of different locks in a single day, all following the exact same instruction manual.

3. The Star: Hermes

The team built Hermes, a lightweight AI brain trained only on the data from this DEL factory.

• What makes it special? Hermes has never seen a "traditional" lab report. It has never been told "this key fits with a strength of 5 stars." It only knows "this key stuck, and that one didn't."
• The Magic: Despite never seeing the complex, messy data of the real world, Hermes learned the fundamental rules of how keys and locks interact. It learned the "shape" of a good fit so well that it can apply those rules to locks it has never seen before.

4. The Test: Can it handle the real world?

To see if Hermes was truly smart or just memorizing the factory, they put it through three tough exams:

1. The "New Lock" Test: They gave Hermes locks (proteins) it had never seen in the factory.
   • Result: Hermes did great. It figured out the new locks because it understood the underlying logic, not just the specific examples.
2. The "New Key" Test: They gave Hermes keys made of materials (chemical structures) it had never seen.
   • Result: It handled these well too, showing it learned general principles, not just specific shapes.
3. The "Real World" Test: They tested it against data from other labs (the messy library).
   • Result: Hermes performed surprisingly well, proving that the clean data from the factory taught it skills that transferred to the messy real world.

5. The Superpower: Speed

There is another reason Hermes is a game-changer.

• The Competitor (Boltz-2): Imagine a super-smart detective who solves a case by building a 3D model of the crime scene, analyzing every angle, and simulating physics. It's incredibly accurate, but it takes hours to solve one case.
• Hermes: Imagine a detective who has seen millions of cases and instantly recognizes the pattern. It doesn't build a 3D model; it just looks at the description of the key and the lock.
• The Result: Hermes is 500 to 700 times faster than the super-detective. It can screen billions of potential drug keys in the time it takes the other model to check a few thousand.

The Big Picture

This paper tells us that we don't need to rely on messy, inconsistent historical data to train AI for drug discovery. By using clean, massive, and consistent data (like the DEL factory), we can train smaller, faster AI models that are actually better at generalizing to new problems.

In a nutshell: Hermes is a fast, efficient AI that learned to find the right drug keys by watching a massive, perfectly organized factory, and it can now find those keys for diseases it has never even met before. This could speed up the discovery of life-saving medicines by years.

Technical Summary

1. Problem Statement

The field of Protein-Ligand Interaction (PLI) prediction faces a critical bottleneck: the quality and consistency of training data.

• Data Heterogeneity: Public affinity datasets (e.g., BindingDB, ChEMBL) aggregate millions of measurements from thousands of different labs, assays, and protocols. This introduces significant noise, biases, and standardization issues that hinder model generalization.
• Evaluation Challenges: Existing benchmarks often suffer from data leakage (e.g., temporal or structural overlaps between training and test sets) or are too small to provide statistically significant evaluations.
• Limitations of Current Models: While structure-based models (like AlphaFold3 adaptations) are powerful, they are computationally expensive for large-scale virtual screening. Conversely, models trained on noisy public data often fail to generalize to novel targets or chemical scaffolds.

2. Methodology

A. The Hermes Architecture

Hermes is a lightweight, sequence-only transformer model designed for binary binding prediction (bind vs. no-bind).

• Input Representation:
  • Proteins: Uses pre-trained ESM2 (150M or 650M parameters) embeddings.
  • Ligands: Uses pre-trained ChemBERTa embeddings for SMILES strings.
• Core Mechanism:
  • Joint Cross-Attention: Interleaves self-attention blocks (processing protein and ligand sequences independently) with cross-attention blocks (facilitating information flow between protein and ligand tokens).
  • Attention Pooling: Aggregates sequence representations into fixed-length vectors using learned scoring functions.
  • Output: A Multi-Layer Perceptron (MLP) predicts a scalar binding probability.
• Efficiency: By avoiding 3D structure prediction, Hermes enables protein-embedding caching and achieves inference speeds orders of magnitude faster than structure-based models.

B. Training Data: DNA-Encoded Libraries (DELs)

Hermes is trained exclusively on DEL screening data, bypassing traditional affinity measurements (IC50, Kd).

• Dataset: Derived from the "Kin0" library (6.5 million compounds) screened against 239 unique protein targets (approx. 2/3 kinases).
• Labeling Strategy:
  • Data is binarized based on enrichment over control screens (DEL-only, bead-only, no-target).
  • Positives: Compounds with sequencing counts significantly exceeding background.
  • Negatives: Compounds with counts confidently below expected enrichment.
  • Sampling: To address class imbalance (vast majority of compounds do not bind), the authors employ stratified sampling, capping positives per protein and sampling hard negatives to prevent memorization.

C. Evaluation Framework

The authors evaluated Hermes on four distinct benchmarks to test generalization at multiple levels:

1. DEL Protein Split: "Cold target" generalization (same library/assay, unseen proteins).
2. DEL Chemical Library Split (STRELKA): "Cold ligand" generalization (unseen library "AMA020", seen proteins).
3. Public Binders/Decoys: External validation using curated public data (Papyrus++) and synthetic decoys (GuacaMol).
4. MF-PCBA: High-throughput screening (HTS) data from PubChem BioAssay.

3. Key Contributions

• DEL-Only Training: Demonstrates that massive, unified DEL datasets can train models that generalize to traditional affinity benchmarks and novel protein targets without ever seeing explicit affinity values (IC50/Kd) during training.
• Generalizable Representations: Proves that PLI representations learned from DEL enrichment data are transferable across different experimental systems (DEL vs. HTS vs. crystallography).
• Speed vs. Accuracy Trade-off: Introduces a model architecture that is ~500–700x faster than state-of-the-art structure-based models (Boltz-2), making it viable for screening billions of compounds.
• Rigorous Benchmarking: Provides a leak-proof evaluation framework, explicitly addressing data leakage issues common in PLI research by using strict protein-family and temporal splits.

4. Key Results

Performance Metrics

• DEL Protein Split: Hermes achieved a mean AUROC of 0.68 (0.73 for kinases), outperforming the structure-based Boltz-2 (0.64) and the XGBoost baseline (0.74) on this specific split. The XGBoost success here suggests that DEL data contains highly reproducible signals for similar targets.
• Public Binders/Decoys: Hermes achieved a mean AUROC of 0.60, significantly outperforming the XGBoost baseline (0.48) and demonstrating that DEL-trained representations generalize to non-DEL data.
• Kinase Specificity: Performance was consistently higher for kinases (AUROC ~0.66–0.73) compared to non-kinases, reflecting the kinase-enriched nature of the training data.

Inference Speed

• Hermes: ~28.2 samples/second/GPU (on H200).
• Boltz-2: ~0.04 samples/second/GPU (on H100).
• Speedup: Hermes offers a ~500–700x speedup, enabling large-scale virtual screening that is computationally infeasible for structure-based models.

Limitations Observed

• Cold Ligand Performance: Performance dropped significantly on the DEL Chemical Library Split (AUROC ~0.56), suggesting challenges in generalizing to entirely new chemical scaffolds not present in the training library.
• Memorization: The XGBoost baseline performed well on the DEL Protein Split, indicating that some of Hermes's success may stem from memorizing specific hit profiles of similar targets rather than purely learning generalizable physics.

5. Significance and Conclusion

This paper establishes DEL data as a superior, unified source for training generalizable PLI models.

• Paradigm Shift: It challenges the reliance on noisy, heterogeneous public affinity datasets, showing that massive, consistent DEL screens can capture transferable interaction representations.
• Practical Impact: The Hermes model bridges the gap between accuracy and speed. While structure-based models (like Boltz-2) offer higher theoretical accuracy for specific targets, Hermes provides a practical solution for ultra-large-scale virtual screening where speed is the primary constraint.
• Future Directions: The authors suggest that future improvements could come from incorporating structural data (e.g., co-folding models) and moving beyond binary labels to model relative binding strengths using normalized sequencing counts.

In summary, Hermes proves that "big data" from unified experimental protocols (DELs) can overcome the data quality issues of traditional drug discovery, enabling fast, generalizable AI models for protein-ligand binding prediction.