Add helper to compute simple SNP feature matrix for ML workflows

[ishaniray1]

Downstream users often want to feed MalariaGEN SNP/genotype data into machine‑learning models (e.g. taxon classification, resistance prediction), but there is no small, documented helper that turns SNP data into a simple, ML‑ready feature matrix. Each user currently has to re‑implement feature extraction on top of snp_allele_frequencies() and snp_genotype_allele_counts().
Proposed
Add a lightweight helper snp_feature_matrix(...) on the Anopheles SNP frequency analysis API that:
Accepts the standard parameters (transcript or region, sample_sets, sample_query, optional cohorts).
Returns a pandas.DataFrame where rows are samples or cohorts and columns are:
total_snp_count
nonsynonymous_snp_count
mean_allele_frequency
Internally uses existing public methods such as snp_allele_frequencies() and snp_genotype_allele_counts() (no ML inside the package).

Implementation details
Implemented as AnophelesSnpFrequencyAnalysis.snp_feature_matrix in malariagen_data/anoph/snp_frq.py, decorated with _check_types and @doc(...) like other public methods.

Supports:
Cohort mode (cohorts provided): one row per cohort; counts based on unique (contig, position) sites; mean_allele_frequency from each frq_ column.
Sample mode (cohorts is None): one row per sample; per‑sample counts from snp_genotype_allele_counts; mean AF from snp_allele_frequencies(..., cohorts={"all": "True"}).

Includes two tests in tests/anoph/test_snp_frq.py that check expected columns and non‑empty results using ag3_sim.

I’ve opened a PR implementing this and am very happy to adjust the feature set or signature if you’d prefer.

[ishaniray1]

Hi @jonbrenas, I have raised a PR(#1186) against this issue, please look at it once whenever you are available

[jonbrenas]

Hello @ishaniray1,

Downstream users often want to feed MalariaGEN SNP/genotype data into machine‑learning models

Do they? I have not met a user yet who wanted to do this. Can you justify why the feature matrix would be useful? In both examples that you cite, insecticide resistance detection and species classification, the frequency of specific SNPs is a much better metric than the number of SNPs (synonymous or not) that are observed in a region.

[ishaniray1]

Hi @jonbrenas, thank you for the feedback, you are right in thstaggregate SNP counts are far too coarse for biological ML tasks. For both IR prediction and species classification, specific SNP frequencies are more informative . My intent with this initial PR was to familiarize myself with the API internals and build a scaffold for an ML-feature exporter, as I am preparing a proposal for the GSoC 26 Building a machine-learning taxon classifier project. I wanted my first contribution to look at the data structures that an ML pipeline would eventually ingest.
Given your point, would it be valuable to pivot this PR into an exporter that returns a true (n_samples, n_snps) numeric genotype matrix (like alternate allele dosage) specifically formatted for direct injection into scikit-learn/PyTorch?
If that kind isn't something needed right now, I am fine closing this issue and PR. Either way, this was a learning exercise for me on how the API handles cohorts and allele frequencies under the hood. I plan to email support@malariagen.net with my proposal skeleton for the FASTQ-based taxon classifier, would you be open to discussing the feature engineering strategy for that project there?

[jonbrenas]

Hi @ishaniray1,

Because the deadline is fast approaching, the email you will get back from support@malariagen.net will probably tell you that it is unlikely your assigned mentor (who will likely be me) will have time to review your proposal, but I am interested in your ideas so I will do my best to send you my comments.

I need to point out, however, that the classifier is supposed to work with fastq files which are in a very different format from the genotypes available through the API. The classifier will be run upstream of the API to decide to which reference genome to align each sample.

[ishaniray1]

Hi @jonbrenas, thank you so much! I really appreciate you taking the time to look out for my proposal despite the tight deadline.
And thank you for that clarification re the fastq requirement, I finalized the proposal skeleton now with this, and just sent it to the support email .

[Farangiz1505]

Hi! I'm Farangiz, a Junior Data Analyst with a background in Applied Linguistics. I’ve been exploring the MalariaGEN GSoC 2026 project, specifically the "Natural-Language Interface" idea, and I’d love to take a crack at this issue.
In my current work as a BI Developer Intern, I frequently build data pipelines that transform raw, complex datasets into structured formats for analysis. I believe creating a helper function for SNP feature matrices is a perfect bridge between low-level data and high-level ML workflows—which is exactly what the natural-language interface aims to simplify.
My plan is to:

1. Review the existing malariagen_data API to ensure the helper integrates seamlessly.
2. Develop a Python function that efficiently computes the matrix while handling missing values or specific genomic variants.
3. Ensure the output is directly compatible with standard ML libraries like Scikit-learn or PyTorch.
   Could you please assign this to me? I’m excited to contribute!

[jonbrenas]

Hi @Farangiz1505 ,

As I told @ishaniray1, I am not sure that such a function is needed but you are welcome to look at the PR they created and propose improvements or make other comments.