Chemically informed representations of amino acids enable learning beyond the canonical protein alphabet

This paper introduces a chemically informed peptide representation based on 2D molecular structures and convolutional autoencoders that enables machine learning models to generalize beyond the canonical amino acid alphabet to unseen post-translational modifications while providing chemically interpretable insights.

The Gist

Imagine you are trying to teach a computer how to understand proteins. For decades, scientists have taught computers to read proteins like they read a book: using a fixed alphabet of 20 letters (A, C, D, E, etc.), where each letter stands for a specific amino acid.

The Problem:
This "letter-based" system works great for standard proteins. But it has a major blind spot. It's like trying to describe a complex painting using only a list of paint names. If an artist adds a special, glowing gold leaf to the canvas (a chemical modification), the letter-based system can't really describe what that gold leaf looks like or how it feels. It just sees a new, strange symbol it doesn't understand. In biology, these "gold leaves" are called Post-Translational Modifications (PTMs), like phosphorylation (adding a phosphate group). They change how proteins work, but old computer models often ignore them or get confused by them.

The New Idea:
The authors of this paper asked: What if we stopped teaching computers to read the "letters" and started teaching them to look at the "pictures"?

Instead of giving the computer a string of letters like A-T-G-C, they gave it a mosaic of images.

• They took the chemical structure of every amino acid and turned it into a 2D drawing (like a blueprint).
• They stitched these drawings together side-by-side to create a long, strip-like image of the whole protein chain.
• Now, the computer isn't reading words; it's looking at a picture of the molecule's actual shape, size, and chemical "furniture."

The Magic Tool (The Autoencoder):
To make sense of these pictures, the team used a special type of AI called a Convolutional Autoencoder. Think of this AI as a highly skilled art student who is given a complex mosaic and asked to:

1. Compress it: Squish the whole image down into a tiny, 256-number "summary" (a latent vector) that captures the essence of the shape.
2. Rebuild it: Try to draw the original mosaic back from that tiny summary.

By practicing this "squish and rebuild" game, the AI learns to understand the physics of the molecule. It learns that a phosphate group looks like a specific cluster of atoms, regardless of which letter it's attached to.

The Test: The Immune System's Bouncer
To see if this worked, they tested it on a classic biology problem: MHC Class I binding.

• The Analogy: Imagine the immune system has "bouncers" (MHC molecules) at a club. They only let specific peptides (guests) in. If a peptide fits the bouncer's shape, it gets in. If not, it's rejected.
• The Challenge: Predicting which peptides get in is usually done by looking at the sequence of letters.
• The Result: The new "image-based" AI performed almost as well as the best "letter-based" AI. But here's the kicker: It could understand modified guests.

The "Magic" Moment: Generalization
The most exciting part happened when they tested the AI on a guest it had never seen before: a peptide with a phosphorylated amino acid (a modified guest).

• The AI had never been trained on this specific modified guest.
• However, because the AI learned the chemical picture, it realized: "Hey, this phosphorylated serine looks a lot like a negatively charged aspartic acid, which is a known VIP guest for this bouncer."
• The AI correctly predicted that the modified guest would get in, simply because it understood the chemistry, not just the name of the letter.

Why This Matters:

1. No More "Ad-Hoc" Fixes: You don't need to invent new letters for every new chemical modification. You just show the computer the picture of the new molecule, and it figures it out.
2. Explainable AI: With letter-based models, it's hard to know why the AI made a decision. With this image-based model, you can use "heat maps" to see exactly which part of the chemical drawing the AI was looking at. It's like pointing to the specific atom in the drawing that convinced the bouncer to let the guest in.
3. Future-Proof: This opens the door to studying synthetic proteins and rare diseases caused by weird chemical changes that don't fit in the standard 20-letter alphabet.

In a Nutshell:
The authors swapped the dictionary for a camera. By teaching computers to see the chemical structure of proteins as images rather than just reading them as text, they created a system that can understand the "flavor" and "shape" of molecules, allowing it to predict how modified proteins behave—even ones it has never seen before.

Technical Summary

1. Problem Statement

Current computational models for protein analysis predominantly rely on a symbolic 20-letter amino acid alphabet (one-hot or tokenized sequences). While effective for evolutionary pattern recognition, this approach has critical limitations:

• Abstraction of Chemistry: It discards the underlying physicochemical structure of residues (charge, steric size, functional groups).
• Inability to Handle PTMs: It struggles to represent Post-Translational Modifications (PTMs) like phosphorylation, glycosylation, or citrullination. Standard models either ignore these modifications, replace them with ad-hoc symbols (e.g., 'B' for phospho-serine), or fail to generalize to modified residues not explicitly seen during training.
• Biological Relevance: Many immune responses (e.g., in autoimmune diseases like rheumatoid arthritis) are driven by chemically modified self-antigens. Current models cannot naturally capture the structural changes these modifications induce.

2. Methodology

The authors propose a paradigm shift from symbolic sequence encoding to chemically informed image-based representation.

A. Data Construction

• Dataset: Utilized publicly available immunopeptidomics datasets containing experimentally identified MHC class I ligands, including phosphorylated peptides.
• Preprocessing: Filtered for human 9-mer peptides. Used the MHCMotifDecon tool to assign peptides to specific HLA alleles. Created three dataset variations:
  • D_Human9: Peptide sequence + binding label.
  • D_HLA9: Peptide + assigned HLA allele.
  • D_HLA9P: Peptide + MHC pseudo-sequence (key polymorphic residues) + binding label.

B. Chemically Informed Representation (The Core Innovation)

• Molecular Depiction: Instead of letters, amino acids are represented as standardized 2D molecular images.
  • SMILES strings for canonical amino acids and phosphorylated variants (Ser, Thr, Tyr) were converted to RDKit molecule objects.
  • 2D coordinates were computed and aligned to a common peptide backbone template to ensure consistent orientation of side chains.
• Peptide Mosaics: Individual residue images are concatenated horizontally to form a single 2D "mosaic" image representing the entire peptide sequence. This preserves both sequential order and structural features.

C. Model Architecture

• Stage 1: Feature Learning (Unsupervised):
  • A Convolutional Autoencoder was trained to reconstruct the peptide mosaic images.
  • Encoder: Four convolutional blocks (3x3 conv, batch norm, LeakyReLU, 2x2 max pooling) compress the image into a 256-dimensional latent vector.
  • Decoder: Mirrored transpose convolution and upsampling layers reconstruct the input image.
  • Goal: Learn a compact latent space that captures physicochemical features directly from molecular structure.
• Stage 2: Prediction (Supervised):
  • The frozen encoder extracts fixed-length feature vectors from new peptides.
  • These vectors are fed into a feedforward neural network classifier to predict MHC class I binding.
• Evaluation: Nested cross-validation (5 outer folds) across multiple HLA alleles. Performance compared against One-Hot Encoding (23 tokens: 20 canonical + 3 phosphorylated) and sequence similarity baselines (BLAST).

D. Interpretability

• Attribution Analysis: Gradient-based saliency methods were applied to the image-based classifier to generate heatmaps. These maps highlight specific regions of the molecular depiction (e.g., phosphate groups) that drive the binding prediction.

3. Key Results

• Competitive Performance:
  • The image-derived embeddings achieved meaningful predictive performance across multiple HLA alleles.
  • While One-Hot encodings generally achieved higher AUC (due to the strong signal of position-specific residue identity in MHC binding), the structural representations remained competitive, particularly in the low false-positive regime.
  • The structural approach outperformed simple sequence similarity baselines (BLAST).
• Generalization to Unseen PTMs:
  • The model successfully predicted the binding of phosphorylated peptides (specifically pSer at anchor position P2 for HLA-B*40) that were not present in the training set for the classifier.
  • The model recognized that the phosphorylated serine structurally mimics the negatively charged side chains of Glutamic/Aspartic acid, allowing it to generalize to the modified residue based on physicochemical similarity rather than symbolic identity.
• Interpretability:
  • Attribution maps successfully localized predictive signals to specific chemical features.
  • For HLA-B*40, the model assigned high importance to the phosphate group of the pSer residue at the anchor position, confirming it learned the correct chemico-structural logic (negative charge compatibility) rather than memorizing sequence tokens.
• Data Dependency:
  • Reconstruction quality and model stability were highly dependent on dataset size. Rare PTMs or sparsely represented alleles led to diffuse reconstructions, highlighting the need for larger, balanced datasets for robust PTM modeling.

4. Key Contributions

1. Novel Representation Paradigm: Introduced a framework that replaces the symbolic amino acid alphabet with explicit 2D structural depictions, enabling models to "see" chemistry rather than just read sequences.
2. Handling PTMs without Ad-hoc Tokens: Demonstrated that chemically modified residues can be represented naturally within the same space as canonical residues without expanding the vocabulary or requiring special tokens.
3. Physicochemical Generalization: Proved that models trained on structural images can infer the behavior of unseen modified residues by learning physicochemical relationships (e.g., charge mimicry) rather than symbolic co-occurrence.
4. Enhanced Interpretability: Enabled direct visualization of which molecular substructures (e.g., specific functional groups) drive model predictions, a capability difficult to achieve with standard sequence embeddings.

5. Significance and Future Outlook

• Bridging the Gap: This work bridges the gap between deep learning and chemical intuition, offering a pathway to model the "broader chemical space" of proteins, including non-canonical amino acids and synthetic peptides.
• Immunology Applications: The ability to model modified self-antigens is crucial for understanding autoimmune diseases and designing vaccines that target neo-epitopes generated by PTMs.
• Limitations & Future Work:
  • Current performance lags slightly behind one-hot encodings, likely due to the data-hungry nature of image-based learning and the dominance of motif-based signals in MHC binding.
  • Future directions include using Graph Neural Networks (GNNs) for atom-level relationships, pre-training on larger peptide libraries, and integrating sequence-based and structural representations to combine the strengths of both approaches.

Conclusion: The paper successfully demonstrates that representing peptides as structural images allows machine learning models to learn transferable physicochemical principles, enabling generalization to chemically modified residues that traditional sequence-based models cannot handle.