Gene therapy often uses a gutted virus as a delivery truck: the viral shell (the capsid) carries a therapeutic gene into a patient's cells. The main obstacle is that most people have already been exposed to viruses like these, so their immune systems hold antibodies that recognize the shell and destroy the vector before it delivers anything — and a first dose immunizes even naive patients against any second dose. The obvious fix is to mutate the shell so antibodies no longer bind it, but this creates a brutal engineering dilemma: the shell must still self-assemble around its cargo, and mutations that dodge antibodies frequently break assembly or wreck the vector's ability to reach the target tissue. Machine learning models can now predict capsid assembly fitness and antibody escape from sequence quite accurately, but they are oracles — they say whether a sequence works, never which part of the sequence serves which purpose. As a result, the field's design loops are empirical: screen, validate, repeat.
This proposal argues that the missing piece is exactly what the causal-interpretability literature has spent the last decade formalizing: explaining a model means answering "what-if-things-had-been-different" questions by intervention, not by correlation. We propose to learn a causal abstraction of black-box capsid fitness predictors — a small set of interpretable latent mechanisms (assembly, per-epitope antibody escape, tissue tropism) whose interventions provably commute with matching interventions on sequence positions — and then to run the abduction–intervention–prediction counterfactual pipeline over that abstraction to answer design questions directly: what is the minimal mutation set that raises escape above a threshold while leaving predicted packaging fitness and muscle tropism intact?
Core Claim
The claim is that the notorious "fitness desert" — the collapse of viable capsids beyond a handful of simultaneous mutations, where more than half of random variants fail to assemble in an adeno-associated virus (AAV) surface-loop library Optimal trade-off control in machine learning-based library design, with application to adeno-associated virus (AAV) for gene therapy — is partly an artifact of treating capsid fitness as a single monolithic output. If assembly, escape, and tropism are governed by partially separable mechanisms whose interactions are the source of the observed non-additive epistasis — which Machine-learning a virus assembly fitness landscape demonstrates directly, with a neural network reaching $R^2 = 0.83$ on held-out assembly configurations where a multi-linear model manages only $0.57$ — then a predictor that decomposes fitness into commuting causal mechanisms should predict high-mutation-depth behavior better than a monolithic model of the same capacity, and, more importantly, should support counterfactual design queries that no oracle can answer.
Formally, let $\phi$ map low-level sequence variables to a small set of high-level mechanism variables. The abstraction is valid when intervening on a high-level variable produces the same distribution over outcomes as intervening on the aligned low-level sequence positions:
$$P\big(Y_{\text{hi}} \mid do(a)\big) \;=\; P\big(Y_{\text{lo}} \mid do(\phi^{-1}(a))\big) \quad \text{for every high-level intervention } a,$$
where $do(\cdot)$ is Pearl's intervention operator — forcibly setting a variable and cutting its natural causes, which distinguishes a causal driver from a correlated bystander. Given a valid abstraction, the design question becomes a search over minimal low-level interventions:
$$\Delta^\ast \;=\; \arg\min_{\Delta} |\Delta| \;\; \text{s.t.} \;\; \hat{F}_{\text{escape}}(s \oplus \Delta) \ge \tau_{\text{esc}}, \;\; \hat{F}_{\text{pack}}(s \oplus \Delta) \ge \tau_{\text{pack}}, \;\; \hat{F}_{\text{tropism}}(s \oplus \Delta) \ge \tau_{\text{trop}},$$
i.e., the smallest mutation set $\Delta$ applied to a reference sequence $s$ that clears the escape bar without dropping below the assembly and tissue-targeting bars.
Research Program
Three components, each reusing a technique from one field in the other field's setting.
First, we extend unsupervised causal abstraction to non-linear, discrete-sequence systems. Unsupervised Causal Abstraction learns high-level causal variables without annotations, but only for linear structural causal models with orthogonal linear aggregations — a limitation its authors state plainly, and one that viral fitness landscapes violate everywhere. We replace its linear consistency and projection losses with neural structural equations and a contrastive interventional-commutativity loss estimated by active sequence edits, keeping its anti-collapse projection term. The validity condition above is our training objective, not a post-hoc check.
Second, we translate the abduction–intervention–prediction counterfactual pipeline of Towards Robustness and Explainability of Automatic Algorithm Selection from discrete solver selection to capsid engineering. There, the system inverts a neural structural model to infer latent instance variables, perturbs a hypothesized driver, and propagates forward to see if the recommendation flips — inducing a selection change in 79 instances with feature interventions under 10%. Here, the "instance" is an immune environment (a polyclonal serum profile), the "tool" is a capsid sequence, and the flip is a design counterfactual: hold the inferred serum fixed, intervene on an abstracted escape mechanism, and read off the minimal sequence edit realizing it.
Third, we define an evaluation protocol with two novel metrics: an interventional commututation score (measured agreement between high-level and sequence-level interventions on held-out data) and a counterfactual design success rate (the fraction of issued design queries whose realized edit, scored by a frozen, held-out fitness predictor, satisfies all three thresholds). This follows the evaluation logic of NEMO: Benchmarking Natural-Language Explanations of Vision Model Errors, which scores a hypothesized error factor by the measured prediction shift after neutralizing it — a hypothesis is credited only if the counterfactual edit moves the model, the same standard we apply to an abstracted mechanism.
Proposed Experiments
The dodecahedral capsid assembly landscape of Machine-learning a virus assembly fitness landscape is our ground-truth testbed: all $3^{12} = 531{,}441$ configurations of twelve packaging-signal sites are enumerable, so every intervention can be checked against truth rather than a surrogate. We train on the same 30,000 configurations and compare against the paper's 3-layer perceptron, which reports $R^2 = 0.83$ on the remaining ~501,000 configurations (multi-linear regression: $R^2 = 0.57$). Decisive checks: does the abstraction-augmented model match or exceed $R^2 = 0.83$ at the same training budget, and do its mechanism-level interventions commute with sequence-level interventions measurably better than a matched autoencoder bottleneck?
The simulated receptor-binding-domain deep mutational scanning dataset from A biophysical model of viral escape from polyclonal antibodies tests the escape side: 30,000 variants spanning three known neutralizing-antibody epitope classes, with ground-truth escape weights and neutralization titers. Decisive check: can our unsupervised abstraction recover the epitope structure without being told the epitope classes (the biophysical model assumes them), at accuracy approaching its reported escape-weight correlation of $R^2 = 0.91$ and its $R^2 = 0.98$ on IC90 of unseen multi-mutants averaging three mutations?
The Dyno Therapeutics AAV2 dataset (~140,000 hypervariable-region-VIII sequences) from AAVDiff tests high-mutation-depth behavior, where that discrete diffusion model maintains over 90% packaging viability at 7–20 mutations and roughly 50% zero-shot viability on AAV9 at 9–10 mutations. We hold out multi-mutant viability prediction, then issue design queries and compare our counterfactual edits against diffusion-generated sequences of matched mutation depth, scored by a frozen predictor and, in a stretch goal, by wet-lab packaging assays. The AAV5 seven-amino-acid insertion library of Optimal trade-off control in machine learning-based library design, whose entropy-guided D2 library achieved a five-fold higher packaging titer, provides the coupled design benchmark; the N271D liver-detargeting result of A single amino acid variant in the variable region I of AAV capsid confers liver detargeting — a single substitution cutting liver vector DNA by over 100-fold while sparing muscle transduction — provides the tropism-truth probe for validating that our tropism abstraction identifies such single-position causal levers from screening data rather than merely fitting it.
The generalization motivation is documented, not speculative: Therapeutics Data Commons shows drug-target-interaction Pearson correlation falling from roughly 0.7 to 0.4 under temporal distribution shift across 21 architectures, and explicitly calls for causal-inference methods for therapeutics machine learning — a call it does not itself answer, and one we answer in a concrete subdomain.
Baselines
The do-nothing reference point is random mutagenesis at the target mutation depth. In the NNK degenerate library of Optimal trade-off control in machine learning-based library design, over 50% of random variants fail to assemble or package at all, and per AAVDiff, random mutagenesis yields near-zero viability at 9–10 mutations on AAV9. Any designed set of variants that is majority-viable at that mutation depth beats this reference.
The strongest existing method depends on the sub-task, and we will run each head-to-head with the same training data, the same splits, and the same tuning budget (a fixed grid search, identical model-parameter counts where architecture allows, and evaluation by predictors frozen before any design work begins). For escape prediction, it is the polyclonal biophysical model of A biophysical model of viral escape from polyclonal antibodies, which reports $R^2 = 0.98$ on IC90 of unseen multi-mutants (averaging three mutations) and $R^2 = 0.91$ on escape weights for dominant epitopes; beating it means matching that accuracy on its home data without predefined epitope labels, and exceeding it on multi-mutants deeper than three mutations, where its Hill-equation assumptions have not been demonstrated. For assembly prediction, it is the 3-layer perceptron of Machine-learning a virus assembly fitness landscape, reporting $R^2 = 0.83$ on the dodecahedral landscape; beating it means exceeding $R^2 = 0.83$ at the same 30,000-sample training budget, or matching it while additionally passing the commutativity checks that a plain predictor cannot even be asked. For generation, it is AAVDiff, which maintains over 90% packaging viability at 7–20 mutations; beating it is not about raw viability but about targeted design — achieving comparable viability under explicit escape and tropism constraints that the diffusion model is not conditioned on.
The key ablation removes the interventional-commutativity loss while keeping the projection term and all architecture and data fixed, reducing the method to a learned autoencoder bottleneck over the same predictor. If counterfactual design success rates survive this ablation, the causal ingredient is doing no work and the core claim is wrong; if they collapse, the commutativity requirement — not representation learning — is what enables design queries.
Why not just…?
Why not just fit the polyclonal biophysical model? This is the most serious alternative, and on pure escape prediction it may win: the model already reports $R^2 = 0.98$ on unseen multi-mutants. But it assumes the epitope partition is known in advance, models a single antigen rather than a 60-subunit capsid, and has nothing to say about assembly fitness or tropism — the coupled trade-offs that make anti-vector design hard. The head-to-head on the receptor-binding-domain dataset will settle the accuracy question honestly; the decisive evidence for us must come from the coupled design benchmarks, where the polyclonal model has no mechanism to answer "escape more, assemble still, target muscle still" at all. Its own requirement of at least 20,000 functional variants per serum also raises the cost question our transfer experiment targets.
Why not just run SHAP or feature attributions on a bigger predictor and mutate the top-ranked positions? Attribution methods rank positions by correlation with the output, and the causal-interpretability literature is built on why that fails: Meaningful Explanations of Black Box AI Decision Systems shows correlational attributions cannot support counterfactual reasoning, and The explanation game: a formal framework for interpretable machine learning proves such methods are restricted subcases vulnerable to "counterfactual eclipse" — an explanation that is accurate yet conceals the actionable cause. Three Desiderata for Faithfulness in Machine Learning Explanations: The Case for Causal Abstraction adds that internal-representation interventions can create interpretability illusions unless commutativity is enforced. Rather than argue, we will test it: a fixed mutation budget spent on attribution-ranked positions versus commutativity-guided interventions, scored by escape achieved subject to the viability threshold. If the attribution heuristic matches us, our mechanism is unnecessary — that is a real risk we accept, and the experiment is designed to detect it.
Why not just scale up the diffusion generator and screen harder? AAVDiff already generates deep mutants with over 90% viability, and directed evolution needs no theory. But generation without explanation cannot accept a design query — it has no variables corresponding to "escape mechanism" or "tropism mechanism" to hold fixed, which is exactly the gap its own authors leave open. The pragmatic resolution is complementary, not adversarial: our abstraction can condition or filter the diffusion model, and if raw diffusion sampling at matched compute matches our counterfactual designs on the constrained benchmarks, the causal layer is redundant and we will say so.
Decisive Checks
The proposal succeeds if three things hold: the abstraction recovers the known epitope classes unsupervised on the receptor-binding-domain data at accuracy approaching the polyclonal model's $R^2 = 0.91$; the commutativity ablation shows a clear gap in counterfactual design success relative to a matched autoencoder and to attribution-guided mutation; and the coupled design benchmarks (AAV2 depth, AAV5 insertions) yield majority-viable variant sets under explicit escape-plus-assembly-plus-tropism constraints, at depths where random mutagenesis is near-dead. It fails cleanly if the ablation gap disappears, and that failure mode is worth knowing too — it would say the fitness desert is irreducibly entangled, and that no interpretable decomposition can steer through it.
References
- Optimal trade-off control in machine learning-based library design, with application to adeno-associated virus (AAV) for gene therapy
- Machine-learning a virus assembly fitness landscape
- Unsupervised Causal Abstraction
- Towards Robustness and Explainability of Automatic Algorithm Selection
- NEMO: Benchmarking Natural-Language Explanations of Vision Model Errors
- A biophysical model of viral escape from polyclonal antibodies
- AAVDiff: Experimental Validation of Enhanced Viability and Diversity in Recombinant Adeno-Associated Virus (AAV) Capsids through Diffusion Generation
- A single amino acid variant in the variable region I of AAV capsid confers liver detargeting
- Therapeutics Data Commons: Machine Learning Datasets and Tasks for Therapeutics
- Meaningful Explanations of Black Box AI Decision Systems
- The explanation game: a formal framework for interpretable machine learning
- Three Desiderata for Faithfulness in Machine Learning Explanations: The Case for Causal Abstraction