PepCABO: Latent-space Bayesian optimization for peptide-MHC binding using contrastive alignment

PepCABO is a novel latent-space Bayesian optimization framework that leverages a dual variational autoencoder with contrastive alignment to efficiently discover high-affinity peptide-MHC binders by enabling structured knowledge transfer across alleles and improving sample efficiency in both low- and high-budget experimental settings.

The Gist

The Big Picture: Finding the Perfect Key for a Specific Lock

Imagine your body is a fortress, and the immune system is the security guard. To spot a virus or a cancer cell, the guard needs to see a specific "ID card" (a peptide) displayed on the surface of your cells. This ID card must fit perfectly into a specific slot on the guard's uniform (the MHC molecule).

The Problem:
There are millions of possible ID cards (peptides) and thousands of different types of slots (MHC alleles).

• The Challenge: If you want to design a vaccine, you need to find the perfect ID card for a specific slot.
• The Difficulty: The number of possible combinations is astronomical (like trying to find a specific grain of sand on all the beaches on Earth). Testing them one by one in a lab is incredibly slow and expensive.
• The Current Issue: Existing computer methods try to guess the best ID card, but they often start with a "random guess" (throwing darts in the dark) and struggle to use knowledge from similar slots to help them find the answer faster.

The Solution: PepCABO (The Smart Guide)

The authors created a new tool called PepCABO. Think of it as a super-smart GPS for navigating the vast ocean of possible peptides.

Here is how it works, broken down into three simple steps:

1. The Dual-VAE: A "Translator" and a "Map Maker"

Imagine you have a library of books (peptides) and a library of locks (MHC alleles).

• The Translator (VAE): The system first learns to translate these complex biological sequences into a simple, continuous "language" (a latent space). Instead of dealing with messy strings of letters (A, C, G, T), it turns them into coordinates on a map.
• The Map Maker (Contrastive Alignment): This is the magic part. Usually, these maps are just random. But PepCABO uses a technique called Contrastive Alignment.
  • Analogy: Imagine you are teaching a student. You don't just show them a map of a city; you show them that "High-Value Peptides" (the best keys) for "Lock A" are physically located right next to "Lock A" on the map.
  • By training the system to pull the best keys close to their matching locks on this map, the system learns the shape of the problem before it even starts searching.

2. The Gaussian Process: The "Weather Forecaster"

Once the map is drawn, the system needs to guess where the "treasure" (the strongest binding) is hidden without checking every single spot.

• It uses a Gaussian Process, which acts like a weather forecaster. Based on the few data points it has, it predicts the "temperature" (binding strength) of the surrounding areas.
• The Trick: Because PepCABO learned from other locks (alleles) during its training, it doesn't start with a blank slate. It has a "prior knowledge" of what good keys look like. It knows that if a key works well for Lock A, it might look similar to keys that work for Lock B. This is Knowledge Transfer.

3. Guided Initialization: Starting in the Right Neighborhood

Most methods start their search by picking a random spot on the map.

• PepCABO's Advantage: Because it aligned the locks and keys during training, it knows exactly where to look first.
• Analogy: If you are looking for a specific type of coffee in a giant city, a random search might start in the middle of a desert. PepCABO starts its search right in the "Coffee District" because it knows the geography. This saves a massive amount of time and money.

The Results: Why It Matters

The researchers tested PepCABO against other methods using a computer simulation of a lab experiment.

• Speed: It found the best peptides much faster (fewer "trials" needed).
• Quality: The peptides it found were stronger binders (better keys).
• Efficiency: Even when the "budget" for experiments was very tight (low budget), PepCABO still outperformed everyone else.

The Takeaway

PepCABO is like upgrading from a blindfolded person throwing darts at a board to a person who has studied the board's layout, knows where the bullseyes usually are, and uses a smart guide to aim their first few throws directly at the target.

This means scientists can design better vaccines and immunotherapies faster, using fewer expensive lab tests, by letting the computer do the heavy lifting of "smart searching" based on patterns it learned from related biological data.

Technical Summary

1. Problem Statement

The optimization of peptide sequences to bind specifically to Major Histocompatibility Complex (MHC) class I alleles is critical for immunotherapy and vaccine design. However, this task faces three major challenges:

• Combinatorial Explosion: The search space of peptide sequences grows exponentially with length, making exhaustive search impossible.
• Non-linearity and Context-Dependency: The relationship between peptide sequence and binding affinity is highly non-linear and difficult to predict accurately.
• Data Scarcity and Cost: Experimental measurement of binding affinity is expensive and low-throughput. While there are over 20,000 known human MHC class I alleles, experimental binding data exists for fewer than 300. Consequently, for most target alleles, existing Latent-space Bayesian Optimization (LSBO) methods must rely on random initialization, leading to inefficient early-stage exploration and poor sample efficiency.

Current LSBO methods often fail to leverage binding data from related alleles to inform the optimization of a new target allele, and they frequently suffer from a mismatch between the latent space geometry and the objective function (binding affinity).

2. Methodology: PepCABO

The authors propose PepCABO (Peptide Contrastive-Aligned Bayesian Optimization), a framework that integrates a dual Variational Autoencoder (VAE) with Bayesian Optimization (BO) using contrastive alignment to enable structured knowledge transfer across alleles.

A. Dual Variational Autoencoder Architecture

PepCABO employs a dual-VAE framework that jointly learns continuous latent representations for both peptides ($z_p$) and MHC alleles ($z_m$):

• Peptide VAE: Uses a Transformer encoder and a CNN decoder to map discrete peptide sequences to a continuous latent space.
• Allele VAE: Uses ProtBERT embeddings (focusing on the 34 binding-groove residues) to map MHC allele sequences to a latent space of the same dimensionality as the peptide space.

B. Multimodal Ranked Contrastive Alignment (Multi-RNC)

To align the latent spaces and facilitate knowledge transfer, the authors introduce a Multi-RNC loss function. Unlike standard contrastive losses that use binary labels, Multi-RNC ranks samples based on their relative binding affinities.

• Mechanism: For a given allele $x_m$, the loss encourages the latent representation of high-affinity peptides to be geometrically closer to the allele's latent representation than low-affinity peptides.
• Effect: This induces a latent geometry where the distance between a peptide and an allele correlates with their binding strength, effectively organizing the space to reflect the binding landscape.

C. Surrogate Pre-training and Knowledge Transfer

Before the actual optimization begins, the model undergoes a pre-training phase:

1. Joint Training: The peptide VAE, allele VAE, and a Gaussian Process (GP) surrogate model are trained simultaneously on a large dataset of triplets $(x_p, x_m, y_{p,m})$ using a combined loss function ($L_{pre}$).
2. Surrogate Modeling: A single GP is defined over the joint latent space $(z_p, z_m)$. This allows the model to share information across related alleles via the kernel structure, rather than treating each allele as an independent task.
3. Weighted Loss: A data-dependent weighting function prioritizes high-affinity triplets during training to ensure the surrogate accurately models the regions of interest (high binders).

D. Guided Initialization and Optimization

• Guided Initialization: Instead of random sampling, PepCABO initializes the optimization by sampling from the neighborhood of the target allele's latent representation ($z_m$) in the peptide latent space. This biases the search toward regions structurally similar to known high-affinity binders for that specific allele.
• Bayesian Optimization Loop: During optimization, the system performs end-to-end updates. The surrogate model is updated while preserving the informative prior learned during pre-training (by freezing kernel hyperparameters and updating only variational parameters). The CoBO (Correlated Latent Space BO) loss is combined with the Multi-RNC term to refine the latent geometry dynamically.

3. Key Contributions

1. Contrastive Alignment for LSBO: Introduction of a Multi-RNC loss that aligns peptide and allele latent representations based on relative binding rankings, creating a geometry conducive to optimization.
2. Cross-Allele Knowledge Transfer: Development of a dual-VAE and joint GP surrogate that leverages data from thousands of alleles to optimize for alleles with little to no prior binding data.
3. Guided Initialization Strategy: A novel initialization method that utilizes the learned latent geometry to start optimization in high-probability regions, significantly reducing the number of required experimental evaluations.
4. Unified Framework: A complete pipeline that handles both quantitative (IC50) and qualitative (mass spectrometry) data, using MHCflurry 2.0 as a unified proxy for systematic evaluation.

4. Experimental Results

The method was evaluated on 12 held-out MHC alleles (simulating scenarios with no prior binding data) under both low-budget (200 oracle calls) and high-budget (1,000 oracle calls) settings.

• Performance Metrics: PepCABO was compared against vanilla LSBO, InvBO (Inversion-based LSBO), and PepPPO (a Reinforcement Learning baseline).
• Key Findings:
  • Superior Convergence: PepCABO consistently outperformed all baselines in terms of the Area Under the Optimization Curve (AUOC) and the best-found binding affinity.
  • Impact of Initialization: The "Guided Initialization" variant of PepCABO significantly outperformed the "Random Initialization" variant. In the low-budget setting, guided initialization found better solutions in the second batch compared to baselines.
  • Sample Efficiency: PepCABO achieved faster convergence and higher final binding affinities, demonstrating improved sample efficiency.
  • Experimental Validation: When tested against experimental IC50 data (using percentile ranks), the guided initialization produced an initial batch with an average rank of the 73rd percentile, compared to the 49.9th percentile for random initialization. This confirms the method's effectiveness translates to real-world experimental constraints.

5. Significance

PepCABO addresses a critical bottleneck in vaccine and immunotherapy design: the inability to efficiently optimize peptides for rare or data-scarce MHC alleles. By effectively transferring knowledge from related alleles and structuring the latent space through contrastive alignment, the method drastically reduces the number of expensive wet-lab experiments required to identify high-affinity binders.

The work demonstrates that pre-training surrogates on related tasks and aligning latent geometries are powerful strategies for biological optimization. This approach moves beyond "black-box" optimization by incorporating biological structure (allele-peptide relationships) directly into the optimization framework, offering a scalable solution for personalized medicine where patient-specific MHC alleles often lack sufficient training data.