---
title: "The malaria vector selection atlas"
shorttitle: "The malaria vector selection atlas"
slug: selection-atlas
date: 07/21/2025
thumbnail: '/thumbnails/atlas.png'
tag: malariagen_data, genetics
canonicalUrl: 'https://sanjaycnagi.com/blog/2025-07-21-selection-atlas/'
---

[![selection atlas signals](/blog/signals.png)](doi.org/10.1101/2025.07.16.664900)

Dual-active ingredient ITNs are now being deployed in sub-Saharan Africa. We know that these nets are not only highly effective but effective against pyrethroid-resistant malaria vectors, which should help in the fight against malaria. But given the evolutionary adaptability of malaria mosquitoes, we must rapidly find ways to detect resistance before it becomes widespread, in order employ insecticide resistance management practices and protect the lifespan of these critical tools. 

When a beneficial mutation arises — like one that helps a mosquito survive insecticide exposure — it begins to spread through a population, leaving a signature of reduced genetic diversity around it. We can exploit this signature to pinpoint regions of the genome that are under selection and threatening malaria vector control efforts. 

In a new pre-print on [bioRxiv](https://www.biorxiv.org/content/10.1101/2025.07.16.664900), we present a web resource, the [**Malaria Vector Selection Atlas**](https://anopheles-genomic-surveillance.github.io/selection-atlas/), a database of selection signals in wild Anopheles mosquitoes within the [Vector Observatory](https://www.malariagen.net/vobs/). We analysed whole-genome sequence data from over 4,300 mosquitoes collected across 21 African countries, using population genomic methods to scan the genomes of our cohorts for telltale signs of recent positive selection. The atlas reveals both familiar loci and new threats: we confirm selection at established resistance genes like the *Vgsc* (the target of DDT and pyrethroids) and metabolic enzyme clusters like the *Cyp6p* locus that break down insecticides. But we also discovered novel signals including a diacylglycerol kinase on the X chromosome that may represent a previously unknown resistance mechanism. The selection atlas is powered by an automated computational workflow that enables continuous updates as new data become available. As we integrate more recent data, we envisage that the resource will provide a crucial early warning system for tracking mosquito evolution in near real-time, helping to inform evidence-based decisions about where and how best to deploy our limited vector control tools.
 
---

*Explore the pre-print on [bioRxiv](https://www.biorxiv.org/content/10.1101/2025.07.16.664900v1) and the [selection-atlas web resource](https://anopheles-genomic-surveillance.github.io/selection-atlas/)*


The malaria vector selection atlas

---
title: "AnoKin: Using genomics to estimate dispersal in malaria mosquitoes"
shorttitle: "AnoKin: a close-kin mark-recapture study"
slug: anokin-ckmr-dispersal
date: 03/05/2025
thumbnail: '/thumbnails/dispersal.png'
tag: ckmr, dispersal
canonicalUrl: 'https://sanjaycnagi.com/blog/2025-03-05-ckmr-dispersal/'
---

![anokin logo](/blog/anokin-logo.png)

---

Despite hundreds of years of research, we do not understand how far the malaria mosquito can fly.

This is challenging for many reasons. How can we design a cluster randomised controlled trial (cRCT), without knowing if mosquitoes will fly from a control cluster into an intervention cluster? How can we model the spread of gene drives without knowing how far a single mosquito could transport the drive? 

---

Last year, I was awarded some funding from the [Liverpool School of Tropical Medicine](http://lstmed.ac.uk/) to work on this problem, as part of their [Directors Catalyst Fund](https://www.lstmed.ac.uk/study/research-degrees/director%E2%80%99s-catalyst-fund-0). The idea was to use a genomic approach known as close-kin mark-recapture (CKMR). The method is analogous to traditional mark-release recapture, except rather than marking individuals and releasing them, we use kin to mark each other. We sequence mosquitoes, infer their kinship, and use the distances between kin to estimate how far they disperse. It was first developed in fisheries science, and has many applications; it can be used to investigate dispersal, but also population size as well as survivorship. 

For the last couple of weeks, I have been based in Kisumu, western Kenya, working with Dr Eric Ochomo's team and Brian Polo of the [Kenya Medical Research Institute (KEMRI)](https://www.kemri.go.ke/) to set up the sampling. We are working at Lake Kanyaboli, a site of high malaria transmission that provides a stable habitat for the malaria mosquito *Anopheles funestus* throughout the year. Everything is going super well so far - I am really fortunate to be working with an awesome team!

#### The AnoKin site

In the spirit of transparency, part of my proposal was to set up a [public-facing website for AnoKin](https://sanjaynagi.github.io/anokin/). The site contains some background to the project, some protocols, and important links for the field team.

In ultimate nerdiness 😎, I've set up a [daily GitHub actions](https://github.com/sanjaynagi/anokin/blob/main/.github/workflows/docs-auto.yml) which downloads the sampling and morphological ID data from the ODK server where we store data. It then runs a jupyter notebook to perform some analysis, and re-builds and re-publishes the [webpage](https://sanjaynagi.github.io/anokin/). 

So far, the sampling is going really well. Numbers are relatively low for Lake Kanyaboli, but this may actually be a positive - sampled mosquitos are probably more likely to be related. Follow along on here and the AnoKin website, where I'll post regular updates as our project attempts to yield new insights into mosquito dispersal.

AnoKin: a close-kin mark-recapture study

AnoKin: Using genomics to estimate dispersal in malaria mosquitoes

---
title: Neural networks for taxon prediction
shorttitle: Neural networks for taxon prediction
slug: neural-nets-taxon
date: 02/19/2025
thumbnail: '/thumbnails/phylo.png'
tag: deep-learning
canonicalUrl: 'https://sanjaycnagi.com/blog/2025-02-19-neural-networks-for-taxon-prediction/'
---

<center>
  <figure>
    <img
      src="/blog/tree-river.png"
      alt="phylo tree and"
      width="600"
      height="600"
    />
  </figure>
    <figcaption>
      Figure 1. Some trees.
    </figcaption>
</center>


These two trees are topologically identical. One of them is a river system creating some of the most fertile land on earth, supporting a population of over 300 million people. And the other is a phylogenetic tree of the An. gambiae complex. One of them is also tattooed on my arm. extra points if you can guess the rivers. 

#### Anopheles gambiae; a *complex* world

The major malaria mosquito *Anopheles gambiae*, is a species complex of at least seven recognised species, and at least four additional cryptic taxa. These taxa have varying contributions to malaria transmission, with *An gambiae s.s* and *An. coluzzii* considered the most important vectors, due to their anthropophilic nature, or in other words, their preference for human blood. It is estimated that the ancestor of *An. gambiae* and *coluzzii* diverged from the rest of the complex [approximately 2 million years ago](https://doi.org/10.1126/science.1258524).  

<center>
  <figure>
    <img
      src="/blog/range.png"
      alt="Range of species in the gambiae complex"
      width="350"
      height="350"
    />
    <figcaption>
      Figure 2. Range of species in the gambiae complex <a href="https://doi.org/10.1126/science.1258524">Fontaine et al., 2016</a>.
    </figcaption>
  </figure>
</center>


*An. arabiensis* on the other hand, shows more flexible host preferences but remains an important vector, particularly in drier areas, where it tends to dominate. Meanwhile, *An. merus* and *An. melas* are salt-tolerant coastal species and minor vectors. *An. quadriannulatus* is zoophilic (it prefers to feed on other animals) and is thought to play virtually no role in malaria transmission despite being competent to carry the parasite. 

Despite these important phenotypic differences, these species are morphologically identical - you can't tell them apart even under a microscope. Traditionally, researchers have relied on molecular methods to identify members of the *An. gambiae* complex. The gold standard has been PCR-based assays targeting the ribosomal DNA intergenic spacer regions (IGS), developed by [Scott et al. in 1993](https://doi.org/10.4269/ajtmh.1993.49.520) and later with the SINE200 method [(Santolamazza et al.,  2008)](https://doi.org/10.1186/1475-2875-7-163). Importantly, both methods rely on single genetic loci, which can miss hybridization between species, and these methods cannot detect cryptic taxa, and in those cases provide incorrect species assignments.

#### Neural networks for taxon prediction

I've long been intrigued by the potential of machine learning to solve biological problems. For our recent development of a [platform for targeted genomic surveillance of vectors](https://doi.org/10.1101/2025.02.14.637727), we needed to be able to identify species within the gambiae complex. The amplicon panel itself, originally designed long before my entry into the field by collaborators at the Sanger institute, was built for discriminating between *An. gambiae* and *An. coluzzii*, but not necessarily other species. I had been trying out some traditional machine learning methods, but with Large language models (LLMs) demonstrating remarkable capabilities in so many fields, I thought it was finally time to code up a neural network of my own.

Neural networks are computing systems inspired by the human brain's architecture. They consist of interconnected nodes ("neurons") organized in layers - an input layer that receives data, one or more hidden layers that process the information, and an output layer that produces the final result. Each connection between neurons has an associated weight, determining how much influence one neuron has on another. The network "learns" by adjusting these weights through a process called backpropagation, guided by a loss function that measures how far predictions deviate from known values.

Using free GPUs provided in Google Colab, alongside the TensorFlow and Keras libraries, I trained the network on ancestry informative markers within whole-genome sequence data from the Anopheles gambiae 1000 Genomes Project (Ag1000G). The neural network performed exceptionally well, achieving 100% accuracy in classifying all five species in the complex with sufficient numbers of samples: *An. gambiae*, *An. coluzzii*, *An. arabiensis*, *An. melas*, and the recently described Bissau molecular form. You can find the notebook for deep learning taxon prediction [here](https://colab.research.google.com/drive/1rcanKIJyD5Pnzg_17HCBUuVeIWe5QA5j?usp=sharing).

#### Four SNPs to Rule Them All, and in the Darkness, Predict Taxa

So, pretty amazing that a neural network can classify 5 species with 100% accuracy. Or is it? well, it turns out that neural nets are kind of overkill for this problem. In fact, what I found in the targeted surveillance paper, was that a simple decision tree using only four ancestry informative markers (AIMs) could predict the same taxa to a very high level. 

The decision tree (Figure 3) showed remarkable accuracy for most species: *An. gambiae* (F1 = 0.995), *An. coluzzii* (F1 = 0.995), *An. arabiensis* (F1 = 1.000), and *An. melas* (F1 = 1.000). Although these SNPs were selected to distinguish between gambiae and coluzzi, it seems that they get fixed in different ways in the various species, allowing us to use these combinations to predict taxa. It performed slightly less well for the Bissau form (F1 = 0.849), likely because this cryptic taxon harbors less consistent AIM genotypes compared to the other species. 

![decision tree](/blog/tree.svg)
Figure 3. A simplified representation of a decision tree trained on four AIMs to predict species in the Anopheles gambiae complex

--- 

And back to my opening question - the right image depicts the Five Rivers of Punjab; the Jhelum, Chenab, Ravi, Beas, and Sutlej, tributaries of the Indus (admittedly, not in their correct orientation). The word Punjab literally means 'land of the five (panj) waters (ab)' The similar topology between these two systems - one representing my scientific work, and the other an ancestral home - was an intriguing reminder of the unexpected connections we can find in nature.

Neural networks for taxon prediction

---
title: Claude AI as a language learning companion
shorttitle: Claude AI as a language learning companion
slug: claude-ai-hindi
date: 01/06/2025
thumbnail: '/thumbnails/hindi.png'
tag: llm
canonicalUrl: 'https://sanjaycnagi.com/blog/2025-01-06-llms-language-companion/'
---

![punjab](/blog/punjab.jpeg)

As I’m writing this, I'm sat in the back of my cousins car, driving through the chaotic, cacophonous streets of Ludhiana, a large, sprawling city in the heart of the Indian state of Punjab. I am midway through a trip to visit family here, a trip that I tend to make every year or two. My father moved to England from Punjab in the early 1980s. It was only ever intended to be a temporary visit, but he has remained there ever since, building a life, career and family, far away from home. Whilst we visited India when we were young, me and my siblings never had the chance to learn our father's two native tongues, Punjabi and Hindi. It was something I always felt compelled to address. 

Since submitting my [PhD thesis](https://archive.lstmed.ac.uk/23310/) two years ago, I've been learning Hindi, primarily through online tutoring on [Preply](https://preply.com/). Progress has been steady, despite the fact I'm terrible at doing homework set by my tutor (if you are reading, I’m sorry, Shivaani!). I've recently found a tool, however, which has revolutionised my learning - using a large language model, Claude, as a personal language tutor and learning companion. 

Claude is a family of large language models (LLM) developed by Anthropic, many of which are currently world-leading. These AI models are neural networks trained on vast amounts of textual data, allowing them to understand the context and relationships between words. LLMs are creating enormous interest for their potential to transform many aspects of daily life, such as assisting software developers in writing and debugging code more efficiently, automating routine tasks like email drafting and scheduling, or revolutionizing customer service through intelligent chatbots. In this post, I'll share how I've been using Claude to help me learn Hindi.

---

One of its biggest strengths lies in its ability to generate engaging learning materials. I first started using generative AI for Hindi with OpenAI’s ChatGPT, where I used it to generate stories which I could then translate for practice. It's fun - you can personalise them, and then read the story as you translate (my personal favourite was the story of Sanjay, a cricketer for India who moonlighted as a Bollywood actor, until one day, his secret was discovered 😅). This was with an earlier model (GPT-3.5), and the stories often contained many small mistakes which my tutor would need to correct before I started. Now, when I do this with Claude Sonnet 3.5, the stories are essentially mistake-free. Beyond these stories, it helps to create personalised homework assignments and other translation exercises tailored to my learning needs.

I've found it incredibly useful for breaking down complex grammar structures. It can offer multiple perspectives until the concept resonates, and provides detailed explanations accompanied by practical examples. Each sentence can be analyzed in detail, showing both its literal meaning (like how 'mujhe chai peeni chahiye' directly translates to 'to-me tea drink should) and its natural English equivalent ('I should drink tea'). Typically, I’ve found it really challenging to rapidly access this information online. 

Day to day, my family speak Punjabi to each other rather than Hindi - although they can understand me speaking Hindi, this can make it quite difficult for me to understand them. I’ve found Claude helpful for this, because I can ask it to give me the Punjabi version of a phrase in Hindi, with helpful explanations that go beyond Google Translate. It is also able to easily switch between scripts (for example, Devanagari and Latin script). And most importantly, unlike any normal tutor or learning buddy, Claude is available 24 hours a day, 7 days a week, and responds immediately. It responds thoughtfully, and never judges when responding, giving the user freedom to continue asking questions until understanding arrives. 

My intuition is that whilst LLMs are particularly useful at a beginner level (as I am), as you get more and more proficient in a language, their utility may decrease. And of course, there is only so much a text-based tutor can help with learning a language - nothing will replace real time conversation with native speakers. The nuances of pronunciation, the rhythm of natural conversation, and the subtle cultural clues that emerge through face-to-face interaction remain essential components of achieving fluency. 

While I'm still early in my Hindi learning journey, I'm excited to see how AI tools like Claude continue to evolve as learning companions. As an example of this evolution, Duolingo now have a new subscription tier, [Duolingo Max](https://blog.duolingo.com/duolingo-max/), powered by GPT-4. If you're interested in learning more about Claude and the future of LLMs, I'd recommend having a [play around](https://claude.ai/new), or listening to this episode of the Lex Fridman [podcast](https://open.spotify.com/episode/69V7CtdbB8blcxNPXvpnmk?si=AEsAvzaKQ3iZZp6qA0d8YA), with Dario Amodei (the CEO of Anthropic), Amanda Askell (a philosopher, responsible for shaping Claude's personality), and Chris Olah (co-founder at Anthropic, reverse engineering neural networks).


Claude AI as a language learning companion

---
title: Karyotyping chromosomal inversions in malariagen-data
shorttitle: Karyotyping chromosomal inversions in malariagen-data
slug: karyotyping
date: 11/28/2024
thumbnail: '/thumbnails/chrom.png'
tag: malariagen_data
canonicalUrl: 'https://sanjaycnagi.com/blog/2024-11-28-karyotyping/'
---

Chromosomal inversions, where a region of a chromosome is flipped, play an important role in the evolution of malaria vectors, influencing their ability to exploit ecological niches and potentially affecting their ability to transmit parasites and tolerate insecticide exposure.

The Vector Observatory's [malariagen_data package](https://github.com/malariagen/malariagen-data-python) itself continues to evolve, providing researchers with powerful tools for accessing and analyzing genomic datasets from major malaria vectors. A recent addition to this toolkit is a karyotyping method that allows users to determine the karyotype of specimens in the Vector Observatory by genotyping inversion tagging SNPs. 

Here I plot karyotype frequencies of the 2La inversion in samples from phase 3 of the ag1000g resource, including additional cohorts in VObs we've been working on at the Liverpool School of Tropical Medicine. 

<HTMLPlot
  title="2La frequencies in Ag3"
  pathname="/blog/2La.html"
  height_before_scale={550}
  width_before_scale={600}
/>

You can see that the 2La karyotype (in green) is at higher frequencies as you move towards the more arid Sahel. The Colab notebook to generate this plot is [here](https://colab.research.google.com/drive/1qlhygvn1N9XFn5oQgWIoQOMIzMQkSGoE?usp=sharing).

#### Usage of karyotyping in malariagen_data

The functionality uses tagging SNPs from [compkaryo](https://doi.org/10.1534/g3.119.400445). Previously, the tool only reported the positions of the tagging SNPs, which would cause issues at sites which are multi-allelic - we would not know which alternate allele is the inversion tagging SNP. I've now found the exact alternate alleles, making the estimates more robust.

Here's an example of how to use the new karyotyping function in malariagen_data:

```python
df_karyo = ag3.karyotype(
    inversion="2La",  # Inversions to analyze
    sample_sets="1244-VO-GH-YAWSON-VMF00149",  # The sample set
    sample_query=None
)

df_karyo.head()
```

I hope this is a useful addition for users of malariagen_data! Any questions or feedback, please get in [touch](mailto:sanjay.c.nagi@gmail.com).

*Thank you to Anastasia Hernandez-Koutoucheva and Alistair Miles for the code to plot pie charts with Bokeh!* 


Hey, I'm Sanjay

I'm a research scientist at the Ellison Institute of Technology, developing methods and resources to improve the diagnosis and treatment of infectious pathogens. I enjoy developing open-source computational tools that can aid the community and empower researchers to analyse their own data.

Latest Posts

Contact Me!