Expression yield is the manufacturing metric that gets the least attention at the antibody design stage, and the most attention six months later when it becomes a problem. The standard early-stage selection funnel — affinity, selectivity, developability flags — doesn't include expression yield as a hard filter because yield isn't easily measurable from sequence alone, and measuring it experimentally requires transient CHO expression, which is resource-intensive to do on 50+ candidates.
The result is a predictable pattern: programs advance leads that look good on all the standard metrics, reach the point of stable cell-line development, and discover that the candidate of interest expresses at 0.3–0.8 g/L in CHO when the program needs 3–5 g/L to make preclinical supply feasible. Back-engineering a yield fix at that stage — without disrupting the binding or safety profile — is expensive and not always successful.
We've been working on this problem specifically since early 2024, developing sequence-level predictors of CHO expression yield that can be applied early enough to actually influence design decisions. This post describes what we've found: which sequence features have predictive value, what the limits of the approach are, and how to integrate yield prediction into an early-stage design pipeline without creating false confidence.
Why CHO Expression Yield Is Hard to Predict
CHO expression yield is determined by a chain of cellular processes: transcription efficiency, mRNA stability, co-translational folding, glycosylation, assembly of heavy and light chains, secretion pathway throughput, and protein stability in the cell culture medium. Each step is influenced by sequence features, and those features are not the same ones that determine binding affinity or biophysical stability in buffer. A sequence that folds perfectly in the test tube can still express poorly if it saturates the ER folding capacity, if the HC/LC pairing is mismatched, or if a charged patch interacts unfavorably with the secretion machinery.
Published correlations between sequence features and CHO yield generally explain 30–50% of yield variance in training datasets. That's real but limited — the rest of the variance comes from process conditions, cell line background, and stochastic variation in clonal selection that sequence features simply can't predict. Any group claiming high R² for sequence-to-yield prediction across diverse antibody sequences is overfitting their training data.
The honest version of the prediction problem is: which sequences are at high risk of low yield (say, < 1 g/L in standard CHO batch), and can we identify them at the design stage well enough to avoid them? That's a classification task with noise, not a precise quantitative prediction, and that's how we frame our models internally.
Sequence Features That Predict Expression Problems
From our analysis of a curated set of 280 therapeutic antibody sequences with matched transient CHO expression yield data (drawn from public disclosures in patents and published CMC datasets), six sequence features emerged as statistically significant predictors of low-yield outcomes (yield < 1.5 g/L in transient CHO):
1. VH/VL Interface Hydrophobic Patch Area
The interface between the VH and VL domains is stabilized by a set of conserved hydrophobic contacts. Mutations in the CDRs that extend hydrophobic patches into this interface region — typically via large aromatic side chains at CDR H3 base positions 94, 101, or light chain positions 87–90 — correlate with lower expression yield in our dataset. The proposed mechanism is misfolding or aggregation in the ER during early assembly. Effect size: sequences in the top quartile for VH/VL interface hydrophobic patch area expressed at a median yield 40% lower than the bottom quartile in our dataset.
2. Framework Region Net Charge at Secretion pH
The secretion pathway operates at pH 5.5–6.5 (ER and Golgi lumen). Net charge of the antibody at this pH range matters for interactions with the lectin-based chaperone system (specifically GRP78/BiP binding). Highly negative net charge (< −5) at pH 6.0, driven by framework region acidic residue loading, is associated with increased BiP retention time and lower secretion efficiency. This is a framework-region feature, not a CDR-region feature — it's influenced by germline selection and framework mutations introduced during humanization.
3. CDR H3 Proline Content
Prolines in CDR H3, particularly at positions 95–100, introduce kinetic traps in cis/trans isomerization that slow folding. Antibodies with two or more Pro residues in CDR H3 show a bimodal yield distribution in our dataset — some express normally, suggesting the proline positions allow the loop to adopt its native conformation efficiently, but roughly 35% of multi-Pro H3 antibodies in our set expressed below 1.5 g/L. The prediction value is lower than for VH/VL interface features, but Pro content is trivially computed from sequence and worth flagging for review.
4. HC/LC Charge Complementarity
Correct heavy chain / light chain pairing in the ER is driven partly by electrostatic complementarity at the HC/LC assembly interface. Sequences where both HC and LC carry the same net charge sign (both positive or both negative at physiological pH) showed 28% lower median yield in our dataset, consistent with electrostatic interference in assembly. This is a design-level consideration that's easy to check before synthesis and can often be resolved by germline selection or framework modification without affecting CDR regions.
5. CDR Asparagine Deamidation Motifs in Flexible Loops
NxS/T sequons (asparagine glycosylation sites) in CDR loops — particularly NGT and NGS — are not only chemical liabilities post-expression but also can interfere with ER processing if they're in positions that are inappropriately glycosylated during folding. We observed yield suppression in sequences carrying CDR glycosylation sequons that are buried in a predicted binding-competent conformation but accessible in the unfolded state, suggesting ER processing interference.
6. VH Germline Family Usage
This is less actionable at late stages but important for library design: VH3 family germlines as a group show consistently higher median expression yields in CHO compared to VH1 and VH4 family sequences in our dataset. The difference is approximately 1.2-fold at median (statistically significant, p < 0.01). This doesn't mean VH1 or VH4 sequences are unworkable — dozens of approved therapeutics use these germlines — but for a library design decision where you're choosing between equal-performance scaffolds, germline family yield bias is a real consideration.
A Specific Program Example
In a program we engaged with in mid-2024 targeting a metabolic disease antigen, the lead antibody from phage display had KD = 0.7 nM and excellent selectivity, but our early manufacturability screen flagged a VH/VL interface hydrophobic patch score in the top 5% of our dataset and a CDR H3 with two Pro residues (positions 98 and 100b). Transient CHO expression confirmed 0.6 g/L, well below the 3 g/L target for preclinical supply.
We ran a targeted CDR scan focusing on framework-adjacent CDR H3 positions, looking for substitutions that reduced the VH/VL interface hydrophobic patch score while maintaining ΔΔG within ±0.5 REU of the parent. Three variants from a 30-variant synthesis set showed expression yields of 2.1–2.8 g/L in transient CHO with ≤ 1.5-fold affinity loss confirmed by SPR. The best balance was a variant with one H3 position changed from Phe to Ser (reducing both the hydrophobic patch score and, incidentally, eliminating one of the Pro-adjacent structural constraints), yielding 2.6 g/L and KD = 1.1 nM. That's a candidate you can actually make preclinical material from.
Integrating Yield Prediction Without Creating False Confidence
We're not saying sequence-based yield prediction is reliable enough to replace transient CHO expression screening. It isn't. The correlation between our sequence features and yield is real but explains a minority of variance; there will be false positives (flagged sequences that express fine) and false negatives (sequences that look clean by sequence but express poorly). The value of sequence-level screening is in prioritization, not elimination.
A practical integration point: when ranking a panel of 30–50 lead candidates by combined affinity + developability score, adding the yield risk tier as a tiebreaker changes the synthesis queue in a useful way. Candidates with similar affinity and developability profiles but different yield risk tiers — where the lower-risk candidate gets prioritized — represent the marginal cases where sequence-level prediction earns its keep. The high-risk candidates don't get dropped from the list entirely; they get deprioritized unless there's a strong affinity argument that justifies the risk of a yield remediation cycle later.
What we're actively avoiding is using yield prediction as a hard filter at the sequence stage. The false negative rate is high enough that a hard filter would eliminate candidates that would have expressed normally. The appropriate use is advisory: flag sequences in the high-risk tier, note the specific features driving the flag, and let the program team make the call knowing the risk is there. That's a better outcome than either ignoring yield entirely or using an unreliable predictor to hard-eliminate candidates that might have been your best leads.
The Upstream Design Implication
The more useful application of this knowledge is upstream of lead selection, in library design. If you know that VH3 family germlines express better on average, and that VH/VL interface hydrophobic patches increase expression risk, those facts should inform CDR library construction and humanization scaffold selection before a single sequence is ever synthesized. The phage display or yeast display campaign produces sequences from whatever library you put in; if you've built the library around low-expression-risk scaffolds from the start, the lead candidates that emerge are more likely to express well without remediation.
That's a genuine shift in how manufacturability is addressed — from reactive (catch the yield problem in transient CHO and fix it) to proactive (design the variant space so low-yield sequences are less likely to be selected in the first place). It's not a complete solution — some yield problems emerge from the specific CDR sequence context that no scaffold pre-selection could anticipate — but it shifts the base rate favorably, and that compounds across a program with multiple optimization cycles.