Why most molecules fail CMC — and 8 things you can predict at design time

A molecule looks perfect in Discovery. Target hit. Strong affinity. The team celebrates.

Six months later in CMC, it aggregates. Or its viscosity at clinical dose makes it impossible to administer. Or its PTM hot spots create batch-to-batch variability your QA team cannot accept. The team restarts.

We have looked at this pattern across the Inscinstech CMC v2.2 knowledge base — 82+ curated entries covering real process development histories. Most of these failures were predictable from the sequence. Not "obvious in retrospect" predictable — actually computable at design time, with current open-source tools.

Here is the list.

1. Aggregation propensity

The single biggest reason mAbs fail downstream. Computational prediction has converged in the past three years:

• TAP / SAP scores (Oxford OPIG) — surface aggregation patches and hydrophobicity
• TANGO / AGGRESCAN — sequence-based aggregation hot spots
• Boltz-2 structure prediction — for context-dependent aggregation calls

The honest accuracy bar: low-aggregation predictions are very reliable. High-aggregation predictions need wet-lab confirmation but flag the candidate for early triage. Either outcome saves a quarter.

2. Viscosity at high concentration

mAbs delivered at >100 mg/mL — the new normal for subcutaneous dosing — frequently hit viscosity walls. Sequence-level predictors (charge patches, hydrophobic patches, electrostatic complementarity) get you 80% of the way. The remaining 20% needs DSC + DLS data, but you should not be running those if the sequence is already disqualified.

3. Solubility

Solubility at clinical dose is rarely a problem for IgG1 mAbs in well-behaved frameworks. It is frequently a problem for engineered formats (bispecifics, IgG4, IgA fusions). Sequence-level solubility scoring catches most of the disasters.

4. PTM hot spots

The four PTMs that ruin batch records:

• Asparagine deamidation — sequence-determined, very predictable
• Aspartate isomerization — same
• Methionine oxidation — context-dependent but the sequence narrows it
• N-glycosylation site placement — fully sequence-determined

A scientific scorer should give you all four in under 30 seconds. If you have a candidate with three or more PTM hot spots in the CDRs, you know what you are looking at before you commit a tank.

5. Immunogenicity (humanness)

For mAbs that started as murine antibodies, humanization quality predicts immunogenicity risk. BioPhi's OASis humanness score is the current standard — calibrated against real clinical immunogenicity data.

The honest bar: humanness scoring is excellent at flagging clearly humanized vs clearly not-humanized molecules. Mid-range scores need wet-lab data. But the worst offenders show up immediately.

6. Purification friendliness

This is the "will Protein A actually work on this?" question. Answers are sequence-deterministic:

• Protein A binding behavior — Fc region structure
• IEX behavior — pI calculation + surface charge distribution
• HIC behavior — surface hydrophobicity scoring

A 30-second prediction tells you whether your standard mAb platform will work, or whether this molecule will need a custom downstream train. That changes the cost estimate by ~$300k–$1M.

7. Synthesis difficulty (oligos)

For siRNA / ASO designs, synthesis difficulty is sequence- and modification-pattern-determined. Repetitive sequences, certain modification combinations, and length all predict yield and impurity profile. Our team's experience: a synthesizability scorer trained on real OligoMS data catches >90% of the candidates that will not synthesize at acceptable yield.

8. Characterizability

The least-discussed one. Some sequences are easy to characterize by HPLC and mass spec; some sequences are not. Highly disulfide-bonded mAbs, mAbs with many close-mass PTMs, ADCs with heterogeneous DAR distributions — these create analytical headaches that show up in late-stage development as "we cannot quality-control this batch."

Predict this at design time. A molecule that scores poorly on characterizability is a molecule that will eat your QC budget.

What this looks like in practice

A developability scorecard for a mAb should fit on one page:

Aggregation         0.3% HMW         LOW
Viscosity           12 cP @ 150 g/L  LOW
Solubility          > 200 g/L        LOW
PTM hotspots        0 critical       LOW
Humanness           OASis 92         LOW
Protein A           binds            LOW
IEX (pI)            7.8              MEDIUM (mid-range pI)
HIC                 standard         LOW
─────────────────────────────────────────
Overall             9.1 / 10  ★
Recommended route:  Protein A → CEX → UVC inact → TFF

This used to take 2–4 weeks across multiple labs. With current tools, 30 seconds. The cost of generating it is rounding error compared to the cost of starting a CMC program with a molecule that should have been ranked second.

Why teams still do not do this

Three reasons:

1. Tool fragmentation. The eight dimensions live in eight different tools, each from a different academic group with different input formats. Stitching them together is an engineering project, not a science one.
2. Calibration trust. Predictions need to be calibrated against real wet-lab outcomes. Academic tools are often calibrated on benchmark sets, not your project's actual candidates.
3. Workflow integration. A scorecard that lives in a tool nobody opens during Discovery → CMC handoff is a scorecard that does not get used.

That is the gap we have been working on. InAnvil integrates the eight predictions plus Inscinstech's CMC v2.2 calibration data into a single 30-second pass.

But you can do most of this with open source today. The bar should be: every molecule entering CMC has a scorecard. The 8 dimensions above are the right starting list.