🧬 biology

Protein language models predict artemisinin partial resistance and prioritize genetic modifiers in Plasmodium falciparum

This study demonstrates that protein language models, specifically ESM-3, can effectively predict artemisinin partial resistance in *Plasmodium falciparum* beyond established K13 variants and successfully prioritize context-dependent genetic modifiers like EXO E415G that enhance resistance only in specific genetic backgrounds.

Original authors: Chang Li, Anongruk Chim-Ong, Bi Zhao, Camilla Pires, Min Zhang, Jenna Oberstaller, Justin Gibbons, Jinyong Pang, Samantha Barnes, Yicheng Tu, Geoffrey Siwo, Kami Kim, Shannon Takala-Harrison, Thomas O

Published 2026-09-23

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Chang Li, Anongruk Chim-Ong, Bi Zhao, Camilla Pires, Min Zhang, Jenna Oberstaller, Justin Gibbons, Jinyong Pang, Samantha Barnes, Yicheng Tu, Geoffrey Siwo, Kami Kim, Shannon Takala-Harrison, Thomas Otto, John Adams, Jun Miao, Liwang Cui, Xiaoming Liu, Chengqi Wang

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ⚕️ This is an AI-generated explanation of a preprint that has not been peer-reviewed. It is not medical advice. Do not make health decisions based on this content. Read full disclaimer

Malaria remains one of the most persistent threats to human health, causing hundreds of millions of infections and hundreds of thousands of deaths annually, with the heaviest burden falling on the African continent. The primary weapon against the disease is a class of treatments known as artemisinin-based combination therapies. These drugs work by rapidly killing the malaria parasite in the blood, but a dangerous phenomenon called partial resistance has emerged. In resistant cases, the drug does not kill the parasites as quickly as it should, leaving them lingering in the body for longer periods. This delay forces the treatment to rely more heavily on the partner drug in the combination, which can eventually lead to total treatment failure. For years, scientists have monitored this resistance by looking for specific changes in a single protein called PfK13, which acts like a molecular switch for the parasite's survival. However, relying on just one switch is becoming insufficient. The parasite is evolving in complex ways, and new mutations are appearing in other parts of its genetic code that might be helping it survive, but these are difficult to spot using traditional methods that only look for known, pre-defined changes.

A team of researchers has now developed a new way to see these hidden patterns by treating the parasite's proteins not as a list of individual parts, but as a complete, evolving language. Instead of checking for a single known mutation, they used a type of artificial intelligence trained on millions of natural protein sequences to understand the entire structure of the parasite's proteins. This approach allows the computer to recognize subtle shifts in the protein's shape and composition that signal resistance, even if those specific changes have never been seen before. The researchers tested this method on over 1,200 malaria parasites collected from patients across Southeast Asia over several years. They divided the data into groups based on when the samples were collected, training the computer on older samples and asking it to predict the resistance status of newer, unseen samples. The system proved remarkably accurate, correctly identifying resistant parasites in future collections with high reliability. More importantly, it successfully predicted resistance in parasites from Africa, a region where the disease is evolving independently, even though the computer had only been trained on data from Southeast Asia. This suggests that the AI learned the fundamental rules of how the protein changes to survive, rather than just memorizing specific mutations.

The power of this method extends beyond the main switch, PfK13. The researchers applied the same language-based analysis to 167 other genes that had been suspected of playing a role in resistance. By looking at the full protein sequence of each gene, the model could prioritize which ones were most likely to be involved in the parasite's survival strategy. The results confirmed many genes that scientists already knew about, but they also highlighted new candidates that had been overlooked. To prove that these computer predictions were real, the team selected a specific gene called EXO and a specific change within it, known as E415G, to test in the laboratory. They used a precise gene-editing tool to introduce this change into two different types of malaria parasites: one that already carried the main resistance switch and one that did not. The experiment revealed a crucial detail: the change in the EXO gene did not make the parasite more resistant on its own. It only increased the parasite's survival when it was already carrying the main resistance switch. This finding confirms that resistance is not just about one bad mutation, but about how different genetic changes work together. In contrast, another gene they tested, ATP4, showed no effect on survival in either scenario, demonstrating that the computer model could distinguish between meaningful signals and noise.

This work establishes a new framework for watching the malaria parasite evolve. By using protein language models, scientists can now look at the entire genetic code of a parasite and predict how it might respond to treatment, even before a specific mutation becomes common enough to be detected by older methods. The study shows that resistance is a complex trait shaped by the background of the parasite's entire genome, not just a single gene. While the computer model cannot yet replace the need for laboratory testing, it provides a powerful way to scan for emerging threats and prioritize which genetic changes deserve the most urgent attention. As the parasite continues to adapt, this ability to read the full language of its proteins offers a vital tool for staying ahead of the disease, ensuring that the treatments we rely on remain effective for as long as possible.

Technical Summary: Protein Language Models Predict Artemisinin Partial Resistance and Prioritize Genetic Modifiers in Plasmodium falciparum

Problem Statement
Artemisinin partial resistance (ART-R) in Plasmodium falciparum, characterized by delayed parasite clearance, poses a significant threat to malaria control. While molecular surveillance has been transformed by the identification of nonsynonymous mutations in the PfK13 gene, current marker-based approaches face limitations. These include the inability to interpret emerging sequence variations that lack sufficient epidemiological evidence for immediate classification and the failure to account for genetic backgrounds or sequence contexts that modulate resistance phenotypes. Furthermore, traditional Genome-Wide Association Studies (GWAS) often lack power for rare variants and cannot directly estimate the phenotypic relevance of substitutions absent from discovery populations.

Methodology
The authors developed a protein language model (PLM)-based framework to predict ART-R directly from protein sequences using clinical parasite-clearance phenotypes.

Data Source: The study utilized 1,293 P. falciparum clinical isolates from the Tracking Resistance to Artemisinin Collaboration (TRAC I and II) studies, collected between 2011 and 2018. ART-R was defined by an in vivo parasite clearance half-life $\ge$ 5 hours.
Temporal Splitting: To evaluate generalizability, data were stratified temporally: training (pre-2015, n=769), validation (2015–2016, n=315), and independent testing (2017–2018, n=209).
Feature Extraction: The framework employed ESM-3, a pretrained protein language model, as a frozen feature extractor. Coding sequences for individual genes were translated into protein sequences and encoded into 1,536-dimensional vectors via mean-pooling of residue embeddings.
Classification: These sequence representations served as input for Random Forest classifiers trained to predict ART-R status.
Scope: The approach was first benchmarked on PfK13 and subsequently applied to a curated set of 167 candidate genes previously implicated in ART-R.
Experimental Validation: Model-derived prioritization identified specific amino acid substitutions. EXO E415G and ATP4 G1128R were selected for functional validation. Using CRISPR–Cas9, these mutations were introduced into isogenic P. falciparum lines: the ART-R Cam2 line (carrying K13 C580Y) and a revertant line (Cam2_rev, wild-type K13). Ring-stage survival assays (RSAs) were performed to measure artemisinin susceptibility.

Key Contributions and Results

Generalizable PfK13 Prediction: The PfK13 sequence model achieved strong predictive performance across temporally distinct cohorts (auROC > 0.88; auPR > 0.83). Crucially, the model generalized to geographically distinct and previously unseen K13 substitutions. It correctly classified 13 of 17 variants absent from the training data, including 10 of 13 unseen African variants, without retraining or fine-tuning.
Candidate Gene Prioritization: Extending the framework to 167 candidate genes successfully recovered established resistance loci (e.g., K13, VPS13, FD, ARPS10, MDR2) and prioritized genes with emerging evidence (e.g., ATG18, WD11). Model performance correlated significantly with independent genomic evidence scores but showed weak correlation with transcriptomic or functional evidence scores, consistent with the model operating on protein sequence variation.
Superiority Over Single-Variant Markers: For many genes with informative sequence-level signals, the full protein-sequence framework provided complementary predictive information beyond the strongest single-variant marker. Specifically, 60.3% of genes with an auROC > 0.60 showed an improvement of at least 0.05 in auROC over the training-selected single-variant benchmark.
Context-Dependent Genetic Modifiers: In silico permutation analysis identified EXO E415G and ATP4 G1128R as significant drivers of model predictions. Experimental validation revealed that EXO E415G increased ring-stage survival only in parasites carrying the K13 C580Y mutation, with no measurable effect in the K13 revertant background. Conversely, ATP4 G1128R had no significant effect in either background. This confirms EXO E415G as a background-dependent modifier rather than an independent determinant of ART-R.

Significance and Claims
The paper establishes protein-sequence modeling as a complementary approach for malaria surveillance and biological discovery. The authors claim that PLM-derived representations capture evolutionary and biochemical constraints that allow for the prediction of resistance phenotypes from combinations of amino acid substitutions, even for rare or previously uncharacterized variants.

The study demonstrates that sequence-based modeling can:

Prioritize emerging resistance-associated variation that lacks immediate epidemiological classification.
Identify genetic modifiers whose phenotypic effects are dependent on the specific genetic background (e.g., the presence of K13 C580Y).
Provide predictive information that complements, rather than replaces, traditional GWAS and single-variant surveillance.

The authors emphasize that while the framework successfully recovers signals supported by genomic evidence and nominates candidates for experimental testing, it does not establish causality on its own. They note that population structure and geographic stratification may contribute to predictive signals, and future work requires lineage-aware validation and larger, geographically diverse datasets to distinguish conserved functional effects from population structure. The approach is presented as a tool to help identify resistance-associated sequence changes before they become established surveillance markers.

More like this