B1: Complete variant effect prediction

The asymmetry comes from how the underlying biology constrains and reveals function.

Coding variants change amino acids. Their effects are constrained by what the resulting protein does and how the change affects folding, stability, or activity. The genetic code is universal — the same three-base codon means the same amino acid in any context — so a coding change at position N produces a predictable amino acid change. Protein structure databases (PDB), evolutionary conservation data, and deep mutational scanning experiments provide rich training signal. AlphaMissense achieves ~89% accuracy classifying missense variants as benign vs pathogenic for these reasons.

Non-coding variants affect regulatory elements — enhancers, silencers, splice sites, untranslated regions, chromatin-organising sequences. These elements are:

Heterogeneous in function. Some enhancers activate; others repress. The same element can switch direction depending on which transcription factors are bound. There is no analogue to "the genetic code" for regulatory grammar.

Context-dependent. An enhancer might work in liver hepatocytes but be silent in brain neurons because the relevant transcription factors are tissue-specific. Predicting non-coding variant effects therefore requires predicting effects in the right cell type — multiplying the prediction problem by the number of cell types relevant to a given disease.

Weaker in individual effect. Most non-coding variants modulate expression by 10-30% rather than abolishing function entirely (as a frameshift coding variant might). The signal-to-noise ratio for detection is correspondingly lower, and the combinatorial accumulation of small effects is harder to characterise than single large-effect variants.

ENCODE and the Encyclopedia of Regulatory Elements have catalogued candidate regulatory regions but their detailed grammar remains incompletely understood. Most disease-associated variants from genome-wide association studies are non-coding, so this gap matters substantially for AIHS interpretation — and it is the area where the next wave of foundation models (Enformer, Borzoi, successors) is most actively advancing.

The numbers are sobering. A typical human genome differs from the reference at roughly 3-4 million positions. Most variants have no individual effect of clinical interest. The combinatorial space grows by powers:

Single variants: ~10⁸ common + rare positions in the human genome where variation has been observed. AlphaMissense-class methods can address the coding subset of this; non-coding interpretation is ongoing.

Pairwise interactions: ~10¹³ possible variant pairs from the common-variant catalogue alone. Even with deep mutational scanning of pairs in well-studied proteins (work by Frances Arnold and others, plus systematic efforts in yeast and bacteria), only a vanishingly small fraction of pairs has been experimentally characterised.

Higher-order interactions: real patients carry dozens or hundreds of potentially-interacting variants. The combinatorial space is effectively infinite. No experimental screen can sample it meaningfully, and statistical inference from observational data confounds rare combinations with measurement noise.

For well-characterised single-gene Mendelian disorders, current methods work. For polygenic conditions where effects of many variants combine — most common chronic diseases — prediction is much harder and may not yield to current ML approaches without fundamentally larger interventional datasets.

The realistic ceiling: comprehensive single-variant interpretation for coding variants and well-characterised non-coding regions; probabilistic interpretation with explicit uncertainty for variants in poorly-characterised regions; partial interpretation for common pairwise interactions in well-studied diseases; little capability on higher-order interactions for rare variant combinations.

This is good enough for AIHS in most clinical contexts but is not the universal prediction implied by the original B1 specification. The Phase II decision gate (B3 viability) is partly an assessment of how close this realistic ceiling is to what AIHS actually needs.

These are the two leading experimental techniques that generate the kind of dense functional data prediction models need. Both work by directly measuring variant effects rather than predicting them, at scales that would have been impossible a decade ago.

Massively parallel reporter assays (MPRA) construct synthetic libraries containing many variant sequences — typically 10,000 to 1,000,000 distinct sequences in a single experiment. Each sequence is paired with a unique barcode and a reporter gene whose expression depends on the sequence's regulatory activity. The library is delivered to cells (transfected, transduced, or genome-integrated), the cells are allowed to express the reporters, and high-throughput sequencing reads out which barcodes are associated with which expression levels. This produces direct functional measurements for every variant in the library.

Saturation mutagenesis takes a defined sequence region (a gene, an enhancer, a promoter) and systematically generates every possible single-base change — for a 100-base region, that's 300 variants. The variants are tested either in isolated assays (for small targets) or in pooled screens with phenotypic selection. Variants of Saturation Mutagenesis include "deep mutational scanning" where every possible amino acid substitution at every position of a protein is tested for function.

Throughput for both techniques is improving roughly an order of magnitude every few years through advances in DNA synthesis cost, sequencing depth, and assay miniaturisation. The Atlas of Variant Effects consortium is coordinating systematic application across disease-relevant genes.

By Phase II, comprehensive saturation mutagenesis of coding regions of common disease-relevant genes should be achievable. By late Phase II to Phase III, regulatory-element saturation mutagenesis at scale becomes plausible. These experimental datasets are what will train the next generation of B1 prediction models — the relationship between the experimental and computational sides of B1 is symbiotic: experimental scale-up enables better models, which guide where to focus experimental scale-up.