The Old Maps Were Missing Something

Here's the lay of the sea. We have two approved malaria vaccines now - RTS,S (Mosquirix) and R21/Matrix-M - and they're saving lives across Africa. But they're like boats that can only handle calm waters. RTS,S clocks in around 39% efficacy over four years (Laurens, 2020). R21 does better, hitting roughly 75% in seasonal settings, but both share the same blind spot: they target the NANP repeat region of the Plasmodium falciparum circumsporozoite protein (CSP) while ignoring the junctional epitopes where the most fearsome antibody in the fleet - an antibody called L9 - actually docks.

L9 is something special. It binds NPNV junctional motifs on CSP, and it does this trick where three copies of L9 latch onto adjacent epitopes simultaneously, locking arms through Fab-Fab contacts like sailors linking up in rough weather. This homotypic grip makes L9 arguably the most potent anti-CSP antibody ever described. The catch? Current vaccines don't present those junctional epitopes. It's like building a harbor and forgetting the deepest berth.

Letting the AI Take the Helm

Wu and colleagues decided to build protein scaffolds - essentially custom-designed molecular docks - that present multiple NPNV epitopes with the exact spatial geometry that L9 expects to find (Wu et al., 2026). And they didn't draft these blueprints by hand. They handed the wheel to RFdiffusion's inpainting module and ProteinMPNN, the same deep learning tools from the Baker lab that have been reshaping protein design since 2022 (Wang et al., Science 2022; Watson et al., Nature 2023).

The generative model's job: design compact protein scaffolds that hold up to three junctional repeat epitopes in exactly the right positions and orientations to mimic what L9 sees when it wraps around native CSP. Think of it as asking the AI to build a model ship inside a bottle, except the bottle is a protein and the ship has to be aerodynamic enough to trigger an immune response.

Two Out of Three Ain't Bad (Yet)

The results? Two epitopes landed right where they should - confirmed by crystal structures and affinity measurements. The third epitope drifted from its intended coordinates, like a waypoint knocked off course by an unexpected current. Not ideal, but navigating three simultaneous structural constraints on a single protein scaffold is genuinely hard, and getting two right is a substantial feat.

When mounted on nanoparticles and tested in mice, the scaffold immunogens blocked malaria liver invasion as effectively as nanoparticles displaying short junctional peptides, though longer peptide versions still outperformed them. The ship made it to port, but didn't break any speed records.

A companion effort from a completely independent crew - Nelson et al. at UPenn and St. Jude - tackled the same problem with their own machine learning pipeline called MESODID and published months earlier (Nelson et al., PNAS 2025). Two independent teams reaching for the same horizon is usually a sign that the wind is blowing in the right direction.

Why This Heading Matters More Than the Headwind

The real treasure here isn't the mouse data - it's the proof that generative AI can scaffold multiple epitopes with predetermined spatial relationships. Castro et al. recently showed this same approach works for RSV, where multiepitope scaffolds actually outperformed single-epitope versions in eliciting cross-reactive antibodies (Castro et al., Nature Chemical Biology 2026). If you can control where epitopes sit on a designed protein, you can start engineering the geometry of immune responses themselves.

For malaria, that means potentially coaxing the immune system into producing L9-like antibodies with those critical Fab-Fab contacts - antibodies that bind CSP like a crew lashing themselves to the mast before the storm hits, rather than just grabbing the nearest rope.

With tools like mapb2.io for visually mapping out these complex protein interaction networks, or with RFdiffusion now in its third generation and getting faster and more capable by the season, the barriers to designing next-generation immunogens are dropping fast.

Charting What's Ahead

I've seen many a computational design look brilliant on paper and founder on the rocks of in vivo testing. This crew is honest about their limitations - the third epitope displacement, the modest advantage over simple peptides. But they've drawn a new chart. The coordinates are set. The next voyage will know where the shallow waters are.

Malaria still kills a child roughly every minute in sub-Saharan Africa. The current vaccines are a lifeline, not a cure. If deep learning can help us design immunogens that trigger the right antibodies in the right configuration, this weathered old field might finally find favorable winds.

References

1. Wu NR, Castro KM, Beutler N, et al. Deep learning-enabled scaffolding of spatial arrays of PfCSP epitopes. PNAS. 2026;123. DOI: 10.1073/pnas.2521914123
2. Nelson JAD, Garfinkle SE, Lin ZJ, et al. Machine learning enables de novo multiepitope design of Plasmodium falciparum circumsporozoite protein to target trimeric L9 antibody. PNAS. 2025;122(49):e2512358122. DOI: 10.1073/pnas.2512358122
3. Castro KM, Watson JL, Wang J, et al. Accurate single-domain scaffolding of three nonoverlapping protein epitopes using deep learning. Nature Chemical Biology. 2026;22(4):604-611. DOI: 10.1038/s41589-025-02083-z
4. Watson JL, Juergens D, Bennett NR, et al. De novo design of protein structure and function with RFdiffusion. Nature. 2023;620:1089-1100. DOI: 10.1038/s41586-023-06415-8
5. Wang J, Lisanza S, Juergens D, et al. Scaffolding protein functional sites using deep learning. Science. 2022;377(6604):387-394. DOI: 10.1126/science.abn2100
6. World Health Organization. World Malaria Report 2025. WHO Global Malaria Programme.

Disclaimer: This blog post is a simplified summary of published research for educational purposes. The accompanying illustration is artistic and does not depict actual model architectures, data, or experimental results. Always refer to the original paper for technical details.