Computational antibody design is a search over a space too large to enumerate, using scores that are proxies for a measurement nobody has made yet. The design of the funnel — what is filtered, in what order — matters more than the choice of generative model.
Where the sequence actually varies
An antibody is mostly conserved. The variable domains of the heavy and light chains each consist of four framework regions, which are close to germline and structurally rigid, interleaved with three complementarity-determining regions, the loops that contact the antigen. Design work almost always holds the frameworks fixed and varies the CDRs, because changing framework residues risks the fold and the developability of the whole molecule for little gain in binding.
Of the six CDRs, the third loop of the heavy chain is the one that matters most. It is the most variable in both length and composition — typically somewhere between about 5 and 25 residues — because it is formed by the junction of three gene segments with imprecise joining and added nucleotides, and it usually contributes the largest share of the antigen contact surface. A design campaign that varies only that loop is already exploring a space of 20 to the power of its length, which at 15 residues is about 3 times 10 to the nineteenth.
Positions are referred to by a numbering scheme rather than by raw index, so that the same position in two antibodies of different CDR lengths gets the same label. IMGT, Kabat and Chothia are the common schemes and they disagree about loop boundaries. Fixing the scheme up front is not pedantry; a script that mixes two of them will mutate the wrong residues.
The generator
Three families of generator are in use and they differ in what information they condition on.
- Sequence models trained on antibody repertoires. Masked or autoregressive models trained specifically on large collections of observed antibody sequences learn the statistics of real repertoires, including which residues appear at which numbered positions and which junctions are plausible. Sampling from such a model conditioned on the framework produces CDRs that look like CDRs, which is a much stronger prior than uniform mutagenesis.
- General protein language models. A model trained on all of UniRef has seen relatively few antibodies and, more importantly, hypervariable loops are exactly where its predictions are least constrained. It is a better scorer of framework plausibility than a generator of CDRs. What a general model is and is not good at is the subject of what a protein language model is trained on.
- Structure-conditioned inverse folding. Given a backbone and an epitope, an inverse-folding model proposes sequences predicted to adopt that backbone. This conditions on the target, which the sequence-only generators do not, but it requires a structure of the complex or a credible model of one, and its output is a sequence that folds — not necessarily one that binds harder.
Scoring, and what each score proxies
Nothing in the loop measures binding. Each score is a stand-in, and being explicit about what for is the difference between a useful funnel and an expensive one.
- Model likelihood, as a proxy for foldability and expressibility. The pseudo-likelihood a repertoire model assigns to a candidate says the sequence is typical of real antibodies. Typical sequences express and fold more reliably. It says nothing about the antigen, which the model never saw.
- Predicted binding energy, as a proxy for affinity. Structure-based scoring of the modelled complex, or a supervised predictor trained on measured affinities. This is the only score pointed at the actual objective and it is also the least reliable: predicted changes in binding free energy correlate with measured ones only moderately, and the correlation degrades exactly where you want it — on large improvements.
- Developability liabilities, as a proxy for whether the molecule survives manufacturing. These are motif rules, not models, and they are cheap and decisive. An asparagine followed by any residue except proline and then a serine or threonine is an N-linked glycosylation sequon. An asparagine-glycine pair is a deamidation hotspot. An aspartate-glycine or aspartate-proline pair is an isomerisation or fragmentation site. An unpaired cysteine invites aggregation. Methionine and tryptophan in an exposed CDR oxidise. Extreme charge patches predict poor solubility and fast clearance. A candidate that binds beautifully and carries a glycosylation sequon in CDR-H3 will cost months.
- Germline distance, as a proxy for immunogenicity risk. Sequences far from any human germline are more likely to provoke an anti-drug antibody response. This is a weak proxy — immunogenicity has many other determinants — and it is used as a tiebreaker rather than a filter.
The funnel, worked
Concrete numbers make the shape clear. The counts below are an illustrative campaign, not a result from one.
generated CDR-H3 variants from repertoire model 5,000
after hard liability filters (sequon, NG, DG,
unpaired Cys, exposed Met) 2,900 (-42%)
after germline-distance cap 2,100
after likelihood percentile cut (top 40%) 840
after structure modelling + predicted dG,
keeping top 5% 42
diversity selection: cluster at 80% CDR-H3
identity, keep one per cluster 24
-> 24 synthesised, expressed, and measured by SPR
or BLI for on-rate, off-rate and KD
Two design points in that funnel. The cheap deterministic filters run first, because there is no reason to spend structure modelling on a candidate that will be rejected for a glycosylation motif. And the last step is diversity, not score: taking the top 24 by predicted energy would return 24 near-identical sequences differing at one position, which tests the model’s ranking rather than the space. Clustering first and taking one per cluster spends the same experimental budget on a broader sample, and it is what makes the resulting measurements useful as training data for the next round.
That next round is the point. The first cycle’s value is not the 24 candidates, it is the 24 labelled measurements, which are the first data you have that ties your proxies to the actual objective. Fitting a model on them is a small-sample transfer-learning problem of exactly the kind worked through in fine-tuning a pretrained embedding on a small labelled set, including the clustered split that keeps the reported accuracy honest.
What the computation cannot decide
The first thing to establish in any campaign is whether your proxy scores correlate with your measurements at all, on your target. That correlation is a property of the target and the model together, and a pipeline that has never checked it is running an unvalidated ranking. Measure the correlation on the first round’s 24 and report it; if it is near zero, the funnel is filtering at random and the fix is a different proxy, not more candidates.
Beyond affinity, none of the properties that decide whether a molecule can become a therapeutic are predictable from sequence with useful accuracy today: thermostability under manufacturing conditions, aggregation at high concentration, viscosity, pharmacokinetics, off-target tissue binding, immunogenicity in humans, efficacy, and safety. These are established by laboratory assay, then by animal studies, then by clinical trials conducted under regulatory oversight.
So the honest description of what this pipeline does is: it reduces the number of molecules that have to be made in order to find good ones. It does not establish that any candidate is safe or effective, no design method does, and the approval pathway for a therapeutic antibody is unchanged by how the sequence was proposed.
Top comments (0)