Why Trastuzumab Is the Perfect Test Case
When I built my antibody developability predictor, I knew I needed to validate it against a molecule where we actually know the manufacturing story. Trastuzumab (Herceptin) was the obvious choice. It’s one of the most successful therapeutic antibodies ever made, with decades of manufacturing data behind it. More importantly, we know it works well in production. High expression yields, manageable aggregation, stable formulation. If my model couldn’t correctly identify this as a highly developable molecule, something would be fundamentally wrong.
Thank you for reading this post, don't forget to subscribe!But here’s what makes this interesting: Trastuzumab wasn’t always easy to manufacture. The early development had its challenges, and the process optimization took years. So what does the model actually see when it looks at the sequence? Does it recognize the features that made this antibody ultimately successful, or does it flag the issues that required engineering solutions?
Let’s walk through what happened when I fed Trastuzumab’s sequence into the model.
The Sequences
First, the raw inputs. Trastuzumab’s heavy chain variable region is 119 amino acids, and the light chain is 107 amino acids. I’m working with the variable domains because that’s where most developability issues originate. The complementarity-determining regions (CDRs) that give the antibody its specificity are also the regions most prone to aggregation and instability.
The model doesn’t see “Trastuzumab, the blockbuster drug.” It sees patterns in the amino acid sequence. Hydrophobic stretches. Charged residues. Potential glycosylation sites. Deamidation hotspots. It calculates 47 different features from these sequences, each representing some aspect of physicochemical behavior or structural risk.
The Prediction: High Developability
The model’s verdict was clear. Developability score: High. Confidence: 87%.
The probability breakdown was interesting. The model assigned a 12% chance of medium developability and only 1% chance of low developability. This wasn’t a borderline call. The model was confident this antibody would behave well in manufacturing.
And it was right. Trastuzumab is known for good expression in CHO cells, typically yielding 2 to 5 grams per liter in optimized fed-batch processes. It’s stable enough to be formulated as a liquid product. Aggregation is manageable with standard formulation strategies. From a manufacturing perspective, this is about as good as it gets for a therapeutic antibody.
What the Model Saw
Looking at the feature importance breakdown, the model’s confidence came from several key indicators. The overall hydrophobicity (GRAVY score) was moderate, not high. This matters because highly hydrophobic molecules tend to aggregate during production. Trastuzumab sits in a sweet spot where the antibody is stable in solution without excessive self-association.
The charge distribution was also favorable. The model calculated a net positive charge at physiological pH, but not extremely high. Highly charged molecules can have expression issues or require complex formulation buffers. Trastuzumab’s charge profile suggested it would be well-behaved.
The glycosylation site analysis was particularly interesting. The model identified the conserved N-glycosylation site in the heavy chain Fc region, which is present in essentially all IgG antibodies. But critically, it didn’t flag unusual glycosylation patterns in the variable regions. Non-canonical glycosylation sites can cause heterogeneity and stability problems. Trastuzumab was clean in this regard.
The instability index, a calculated metric based on amino acid composition and dipeptide frequencies, came back at 38.2. Anything below 40 is generally considered stable, and Trastuzumab fell comfortably in that range. The model weighted this feature heavily in its decision.
Where Reality Gets Complicated
But here’s where the story gets more nuanced. While the model correctly predicted high overall developability, it didn’t capture everything that happened during Trastuzumab’s actual development. The original clinical material was produced in CHO cells, but the commercial process underwent significant optimization. There were formulation challenges. The early lyophilized formulation had reconstitution issues and eventually gave way to a liquid formulation. There were aggregation concerns during long-term storage that required careful excipient selection.
The model doesn’t know about these process-specific details. It’s predicting intrinsic molecular properties, not the details of how a particular company optimized a particular process. This is actually appropriate. What I’m building is a first-pass filter to identify molecules that are fundamentally difficult versus fundamentally tractable. Trastuzumab was tractable. The engineering effort required was normal for a therapeutic antibody, not exceptional.
False Negatives: What the Model Might Miss
There’s an important lesson here about model limitations. The model correctly identified Trastuzumab as developable, but what if there had been subtle sequence features that caused problems? What if there was a specific dipeptide motif that causes issues in a particular CHO cell line but not others?
This is where the training data matters. I trained on publicly available information about therapeutic antibodies, which means the model learned patterns that are general across many manufacturing contexts. It doesn’t know about proprietary cell lines, specific process conditions, or company-specific formulation strategies. This is both a limitation and a feature. The predictions are broadly applicable, but they won’t catch every edge case.
The other thing the model can’t see is sequence context. Antibodies are not static linear chains. They fold into complex three-dimensional structures. Some sequence features only become problematic in the context of the final folded structure. A hydrophobic patch that looks risky in sequence might be buried in the core of the folded protein and completely fine. Conversely, two residues that look innocuous separately might be adjacent in 3D space and create a problem.
I’ve thought about incorporating structural predictions (AlphaFold) into future versions of the model, but that introduces complexity. For now, the model works with sequence-derived features, and for Trastuzumab, that was enough.
Comparing to Clinical Reality
The real test of any predictive model is whether it aligns with outcomes. Trastuzumab has been in clinical use since 1998. It’s been manufactured at commercial scale by multiple companies. The manufacturing process is mature and reliable. Annual production volumes are measured in kilograms to tons. This is exactly the kind of molecule you want your developability predictor to flag as “high.”
When I compared the model’s predictions to the published literature on Trastuzumab manufacturing, the alignment was strong. Papers describing process optimization consistently reported that Trastuzumab was a well-behaved molecule. Expression was efficient. Purification was straightforward. The main challenges were typical of any therapeutic antibody (optimizing titer, reducing impurities, ensuring batch-to-batch consistency), not molecule-specific disasters.
This is what validation looks like in the real world. Not perfect prediction of every detail, but correct identification of the overall developability profile. The model said “this will work,” and it did.
What This Means for Unknown Candidates
The value of this kind of analysis becomes clear when you apply it to new molecules. When a discovery team designs a novel antibody, they don’t have decades of manufacturing data. They have a sequence and some binding data. The question is: should we invest in developing this candidate, or are there red flags suggesting we’ll hit manufacturability problems downstream?
If the model gives a new candidate a similar profile to Trastuzumab (high developability, 85%+ confidence, favorable feature scores), that’s a strong signal. It won’t guarantee success, but it suggests the molecule is in a good zone. Conversely, if the model flags issues (high aggregation propensity, unfavorable charge distribution, multiple deamidation sites), that’s a reason to either engineer the sequence or prioritize other candidates.
The goal isn’t to replace process development. It’s to avoid investing years of work into molecules that are fundamentally difficult to manufacture. Trastuzumab is an example of a molecule that was worth the investment. The model can help identify which new molecules fall into that category.
Lessons for Model Development
Building this case study taught me several things about how to build useful ML tools for biopharma. First, you need validation examples where ground truth is well-established. Trastuzumab, Rituximab, Adalimumab—these are molecules where we have real manufacturing experience. They’re not just data points; they’re stories.
Second, you need to be honest about what the model can and cannot predict. The model captures intrinsic sequence properties. It doesn’t capture process-specific optimization, cell line engineering, or formulation science. Those are separate problems. A good developability predictor should be one input into a decision, not the only input.
Third, feature engineering matters more than model architecture. I spent more time thinking about which physicochemical properties to calculate than I did tuning hyperparameters. The reason Trastuzumab’s prediction was accurate is because I included features that capture the right biology: hydrophobicity, charge, PTM sites, instability indices. These are features that experienced process development scientists already think about. The model just scales that thinking.
What’s Next
This case study focused on a success story. The next question is: can the model correctly identify problematic molecules? Can it flag sequences that will cause expression issues, aggregation, or instability?
That’s the harder test. In the next post, I’ll look at what the model actually learned about what makes antibodies difficult to manufacture. Not just “this one worked,” but “here’s why some molecules fail, and here are the patterns that predict those failures.”
Because ultimately, that’s where the value is. Not in confirming what we already know about successful drugs, but in preventing investment in molecules that won’t make it.
This is part of a series on applying machine learning to antibody developability. For the technical details, code, and data, see the GitHub repository.
Pingback: Kemal's