Human-to-mouse variant mapping for better preclinical bets

Most preclinical programs don’t fail because the biology was unknowable. They fail because the model was “close enough” until it wasn’t.

That’s why the new H2M (human-to-mouse) pipeline, published in December 2025, matters for anyone working at the intersection of genetics, disease models, and drug discovery. H2M computationally predicts how clinically observed human genetic variants map onto the mouse genome—and, crucially, whether you can engineer a mouse variant that matches the human change at the DNA level, the protein level, or both.

For teams building genetically engineered mouse models (GEMMs), prioritising CRISPR edits, or trying to translate variant biology into therapeutic hypotheses, this shifts a lot of work from “manual detective work” into a more standardised, scalable workflow. And in an AI in pharmaceuticals and life sciences context, it’s a good example of where computational systems create immediate operational leverage: fewer dead-end models, faster iteration cycles, and more defensible preclinical decisions.

Why variant-to-model mapping is still a mess in 2025

The problem is simple: human genetics is variant-rich; mouse models are variant-poor.

Human sequencing has produced millions of germline and somatic variants. But when a program wants to test a specific mutation (say a ClinVar pathogenic variant or a recurrent oncology lesion from a clinical cohort), the mouse equivalent isn’t always obvious.

Here’s what typically goes wrong:

• Ortholog complexity: “Same gene” isn’t always a clean one-to-one relationship. Paralog and homolog mappings complicate what should be a straightforward edit.
• Sequence context changes the effect: A nucleotide change in a corresponding locus can produce a different amino acid substitution because codons differ across species.
• Noncoding variants are harder than they look: Regulatory elements and deep intronic regions are less conserved, so “equivalent” often becomes debatable.
• Manual workflows don’t scale: Teams bounce between orthology resources, transcript tables, alignment tools, variant notations, and guide design platforms—often with bespoke scripts and a lot of human glue.

This matters because preclinical timelines are tightening. By late December, most pharma and biotech groups are planning Q1–Q2 execution, and model-building backlogs are a familiar bottleneck. If variant selection is a strategic decision, model selection shouldn’t be a spreadsheet scavenger hunt.

What H2M actually does (and why it’s practical)

H2M is a Python pipeline designed to take human variation data and output predicted mouse equivalents, with standardised variant representations suitable for downstream genome editing.

At a high level, H2M runs four steps:

1. Query orthologous genes (using integrated homology catalogs)
2. Align transcripts or peptides (transcript alignment for noncoding; peptide alignment for coding)
3. Simulate mutations (apply the change)
4. Check and model functional effects at both nucleotide and protein levels

The part I like is that H2M doesn’t pretend there’s only one definition of “equivalent.” It explicitly separates:

• NCE (nucleotide change effect): the DNA-level modification
• PCE (peptide change effect): the amino acid-level modification

That separation is exactly what teams need when deciding what they’re trying to conserve: a regulatory disruption, a splice defect, or a protein change.

Three modelling strategies that match real-world use cases

H2M uses three modelling strategies depending on what you’re trying to replicate:

1. NCE-only modelling

   • Used for noncoding variants and frame-shifting mutations
   • Goal: introduce the same DNA-level change in the mouse at the corresponding site

2. NCE-for-PCE modelling

   • Used when the same DNA-level change yields the same amino acid change in both species
   • Goal: conserve both DNA and protein effects

3. Extended NCE-for-PCE modelling

   • Used when the DNA change doesn’t produce the same protein effect in mouse
   • Goal: find an alternative DNA edit (via codon redundancy) that yields the same amino acid substitution

If you’ve ever had a scientist say, “We want the human mutation, not some mouse-ish version,” this is the computational framework that lets you answer, quickly, what’s feasible.

The scale: 3.17 million human-to-mouse mappings

H2M was used to build a database (version 1) covering 3,171,709 human-to-mouse mutation mappings (as of May 2024 input aggregation) drawn from clinically observed sources including major oncology and clinical variant resources.

A few numbers worth pulling forward because they’re operationally meaningful:

• 96% of input human genes mapped to mouse orthologs.
• H2M predicts more than 80% of human variants can be modelled in mouse.
• Modelling rates are higher for coding variants than for noncoding variants, consistent with higher conservation.

These aren’t just feel-good coverage stats. They tell you something actionable: for most therapeutic areas with strong human genetics, you can probably create a mouse model that is closer to the clinical allele than you think—if you stop doing this manually.

“Flank size”: a useful way to think about conservation

H2M introduces a parameter called flank size: the length of conserved sequence on both sides of the variant site (nucleotides for noncoding; amino acids for coding).

This is more than a nice-to-have metric.

• Small flank sizes mean you can likely map the site, but local context differs, which can complicate interpretation.
• Larger flank sizes indicate higher local conservation, which correlates with a greater chance that the functional effect is conserved.

In the database:

• 50% of coding mutations have flank size ≤ 18 amino acids
• 50% of noncoding mutations have flank size ≤ 14 nucleotides

If you’re building a prioritisation rubric for which variants to model first, flank size is an elegant feature to include—especially when your goal is translational fidelity.

Where AI and computational biology show up in day-to-day drug discovery

H2M isn’t “AI” in the marketing sense. It’s a computational pipeline with alignment, simulation, and rule-based modelling strategies.

But in the AI in pharmaceuticals and life sciences series, it belongs here because it demonstrates a core pattern: computational standardisation turns an artisanal workflow into an industrial one.

Here are three practical ways this connects to AI-driven pharma operations.

1) Better variant triage before you spend months building models

Most organisations can’t model every variant of interest. What you can do is triage variants into:

• High-fidelity model candidates (NCE-for-PCE, strong flank conservation)
• Protein-faithful candidates (extended NCE-for-PCE)
• Risky candidates (low conservation, noncoding deep intronic, ambiguous mapping)

Then you layer AI-enabled scoring on top:

• Pathogenicity predictions (e.g., missense effect models)
• Literature and cohort recurrence signals
• Pathway and target tractability features

The result is a more defensible “why this model” decision, which is exactly what translational and portfolio governance groups want.

2) Faster, more standardised genome editing design

H2M was paired with guide design tooling to generate a sizable set of base-editing and prime-editing guides for a subset of cancer-associated variants.

In practical terms, this is what teams want: not just “here’s the mouse coordinate,” but here’s the edit plan.

For preclinical platforms using CRISPR, base editing, or prime editing, this reduces cycle time in:

• feasibility assessment
• guide design iterations
• library design for pooled screening

And it also supports better automation: once variant mapping and representation are standardised, you can push downstream processes (guide ranking, off-target prediction, QC planning) into more consistent pipelines.

3) More credible translational immunology experiments

One of H2M’s showcased applications is mapping variants in the context of neoantigen prediction—whether a human mutation-derived peptide has a plausible mouse counterpart that could be presented by mouse MHC.

Why this matters for drug discovery:

• Immuno-oncology relies heavily on mouse experiments for early validation.
• But many programs quietly assume that mouse neoantigen relevance will “mostly translate.” That assumption often breaks.

Computational mapping gives teams a way to be honest upfront:

• Some human neoantigens will have workable mouse equivalents.
• Others won’t, and you should switch models or switch questions.

That kind of clarity saves quarters, not weeks.

A concrete example: KIT variants and conserved functional impact

H2M’s analysis of the proto-oncogene KIT (and mouse Kit) is a good illustration of how computational mapping can do more than create a lookup table.

The key observation: variants in highly conserved functional domains are more likely to be modellable and to have conserved functional impact.

In KIT, recurrent cancer-associated mutations cluster in functionally critical regions like the transmembrane/juxtamembrane and kinase domains. H2M found a higher proportion of missense mutations in these domains that can be accurately modelled in mouse, and conservation metrics correlated with predicted pathogenicity signals.

If you’re building or validating a KIT inhibitor program (or working in adjacent RTK biology), this supports a practical stance:

Don’t treat all “KIT mutations” as equal candidates for a GEMM. Model the ones in conserved domains first, because they’re more likely to behave like the human lesion.

That’s a simple sentence, but it can reshape a model roadmap.

How to operationalise H2M in a pharma or biotech workflow

If you want this to turn into leads and impact (not just a nice paper), the workflow needs to be straightforward. Here’s a practical implementation pattern I’ve seen work.

Step 1: Start with your real variant list

Use the variants you already care about:

• clinical sequencing hits from trials
• ClinVar pathogenic or likely pathogenic variants
• recurrent somatic mutations for oncology indications

Standardise them into a consistent format (MAF-style representations make downstream work easier).

Step 2: Generate a “modellability report,” not a database dump

The most useful output isn’t every field—it’s a prioritised report with:

• gene and variant identifier
• mapping strategy (NCE-only, NCE-for-PCE, extended)
• flank size / conservation indicator
• engineering feasibility flags (base/prime editing amenability where relevant)

This becomes your cross-functional artefact for biology, in vivo, and genome engineering teams.

Step 3: Tie modellability to decision gates

Make modellability a formal checkpoint:

• Before greenlighting GEMM creation
• Before committing to a pooled editing screen
• Before selecting an in vivo efficacy model for a mechanism claim

If you’re already running AI-assisted target discovery or trial stratification, this is a natural companion: it converts “human genetics insight” into “preclinical test plan” with fewer unspoken assumptions.

Where this is heading in 2026

The direction is clear: preclinical accuracy will increasingly depend on computational matchmaking between human evidence and model design.

H2M is a strong example because it focuses on a hard but solvable part of the chain: mapping, representing, and engineering equivalence at scale. Combine tools like this with phenotype prediction, multimodal disease modelling, and automated editing design, and you get a future where model selection is less tradition and more evidence.

If your 2026 roadmap includes heavier use of human genetics for target validation, don’t stop at the association signal. Ask a more practical question: Can we test the actual allele, in a model we trust, on a timeline the program can afford?