For the transferrin receptor of T. brucei, we explored the structures of PDBs 6SOY and 6SOZ solved by X-ray diffraction, both bound to human transferrin [30]. For the protein design process, we used PDB 6SOY, as it is better resolved (2.75 Å, compared to 3.42 Å). For the invariant surface glycoprotein 65 (ISG65) of T. brucei, we explored the PDB 7PI6 solved by X-ray diffraction [31] and the AF database model for Q587F5 (Uniprot ID) [32]. As design input, we used PDB 7PI6. For the haptoglobin-hemoglobin receptor (HpHbR) of T. brucei, we explored the structures of PDBs 5HU6 and 4WJG solved by X-ray diffraction [33, 34]. As input for the design, we used 5HU6, as it is better resolved (2.90 Å versus 3.10 Å).

To identify atoms and residues involved in protein interactions between the receptors and their cargo proteins, we used ChimeraX v1.11.1. For this, the chain of the desired cargo protein was selected, and contacts with the receptor chain were computed using a Van der Waals (VDW) overlap ≥ –0.40. For H-bond detection, we used a distance tolerance of 0.4 Å and an angle tolerance of 20°.

Binders were designed using RFdiffusion to generate novel backbones using the motif scaffolding method [15], ProteinMPNN to design their sequences [16], and AF2-multimer to model the resulting sequences in complex with their corresponding targets [35].

For Trypanosoma brucei transferrin receptor (TbTfR), the RFdiffusion design was run using the following configurations:

contigmap.contigs=[A20-342/0 5-30/C349-370/5-30 B18-337/0]

contigmap.inpaint_seq=[C350,C351,C353,C354,C357,C358,C361,C362,C363,C365,C366,C368,C369,C370]

ppi.hotspot_res=[A221,A222,A228,A248,A266,B140,B141,B142,B150,B151]

For ISG65:

contigmap.contigs=[A27-86/A94-154/A196-229/A251-316/0 5-30/B1164-1174/15-25/B1109-1114/5-30]

contigmap.inpaint_seq=[B1111,B1112,B1165,B1166,B1167,B1168,B1169,B1172,B1173]

ppi.hotspot_res=[A70,A73,A74,A77,A282,A283,A286,A289,A290,A293]

For HpHbR1:

contigmap.contigs=[D38-297/0 10-20/B92-92/3-3/B96-96/13-30/B41-45/10-20]

contigmap.inpaint_seq=[B43]

ppi.hotspot_res=[D157,D160,D161,D164,D200,D201,D203,D56,D57,D59,D60,D63,D66,D67,D70,D64,D61,D204,D153,D68,D71]

For HpHbR2:

contigmap.contigs=[D38-297/0 10-35/C300-305/5-30/C275-280/10-35]

contigmap.inpaint_seq=[C276,C277,C279,C301,C302,C303,C304]

ppi.hotspot_res=[D70,D71,D72,D73,D74,D75,D78,D79,D81,D82,D85,D86,D251,D252,D255,D256,D258,D259,D262]

In all cases, we generated 1000 backbones of length 50–80 and we set guiding_potentials = ["type:binder_ROG,weight:5,min_dist:5", "type:binder_ncontacts,weight:10", "type:interface_ncontacts,weight:5"], guide_scale = 0.5 and guide_decay = "linear" as potentials, except for targets HpHbR1 and HpHbR2, for which we tried different potentials without success. The designs shown in the figures change "type:binder_ncontacts,weight:20", "type:interface_ncontacts,weight:5" for the former and "type:binder_ncontacts,weight:5", "type:interface_ncontacts,weight:10" for the latter. We also tested different sizes for both, leaving 70–90 for HpHbR2. These potentials are soft constraints that we optimized iteratively, based on trial and error, to maximize the number of successful designs. They control the final radius of gyration of the binder (binder_ROG: more weight implying more compact designs), the number of intra-molecular contacts of the binder (binder_ncontacts: more weight implying more intramolecular contacts), and the number of intermolecular contacts between the binder and the target (interface_ncontacts: more weight implying more inter-molecular contacts). See the “Availability of data and materials” section for the source code used in this work.

In all cases, we generated a single sequence per backbone with ProteinMPNN by fixing the target chain and allowing the design of all the positions that weren’t a Gly amino acid, omitting Cys residues and using a temperature of 0.0001. Each designed sequence was modelled in complex with the corresponding target with AF2-multimer v3 using LocalColabFold v1.6.1 and the following parameters: --num-models 5, --num-recycle 3, --rank ipTM, --recycle-early-stop-tolerance 0.5, --msa-mode single_sequence, --templates and --custom-template-path ./pdbs [36]. As input, the paired + unpaired single sequences [37] of both target and designed sequences were given, along with the input PDB as template, stored in the ./pdbs directory.

For each of the five AF2 predicted structures, the following metrics were extracted from the per-model PDB and JSON output files. Per-residue predicted local distance difference test (pLDDT) was averaged over all residues assigned to the binder chain to yield a mean binder pLDDT value, with higher values indicating greater confidence in the local structure of the designed chain. Predicted aligned error (PAE) matrices were parsed from the AF2 JSON output. The minimum interaction PAE (miPAE) was defined as the minimum value within the binder-to-target submatrix of the full PAE matrix, representing the single most confident predicted inter-chain distance. Interface predicted template modeling score (ipTM) and whole-complex pTM were extracted directly from the JSON score files.

RMSD to the RFdiffusion backbone was computed by superimposing the Cα atoms of the AF2-predicted binder onto the corresponding chain of the RFdiffusion backbone structure, using a least-squares superposition via BioPython’s Superimposer class. Lower RMSD values indicate that the predicted structure recapitulates the designed backbone geometry.

Within each design, the five AF2 models were re-ranked by their miPAE (ascending), independently of the AF2 internal ranking. This produced a PAE rank (1 to 5 per design), where PAE rank 1 corresponds to the model with the lowest predicted error in the binder-to-target interaction.

All plots were generated using Plotly and exported as interactive HTML files. 3D scatter plot displays each resulting AF2 model as a point in a three-dimensional space with the following axes: min interaction PAE (x), RMSD to RFdiffusion backbone (y), and PAE rank (z). Models belonging to the same design are connected by a line, visualizing how a single design’s predicted properties vary across its five models and PAE ranks. Points were colored by mean binder pLDDT. Hovering over any point highlights all five models from that design, facilitating rapid identification of designs with consistently favorable or inconsistent metrics.

Each design was assigned a scalar score summarizing its structural quality across all five models. For each PAE rank r ∈ {1, 2, 3, 4, 5}, the Euclidean distance from the ideal point (0, 0) was computed in the two-dimensional space of miPAE and RMSD to RFdiffusion backbone:

(1) d r   =   m i P A E r 2   +   R M S D r 2

A decreasing weight (w) was applied to each rank (w₁ = 0.10, w₂ = 0.08, w₃ = 0.06, w₄ = 0.04, w₅ = 0.02). The rationale behind this weighting scheme was to incorporate structural reproducibility across the five AF2 models while assigning greater importance to higher-confidence predictions. Equal weighting would place the same emphasis on all predicted models, whereas progressively decreasing weights prioritize the highest-ranked prediction while still rewarding designs that remain structurally consistent across all five models. A simple linear decay was adopted as a straightforward compromise between these two objectives. The absolute magnitude of the weights was chosen so that successful designs would typically yield scores close to or greater than one, providing an intuitive threshold for candidate prioritization.

The weighted sum of distances was inverted to yield the score S:

(2) S   =   1 ∑ r   =   1 5 w r ∙ d r

Higher raw scores therefore indicate designs whose predicted structures consistently exhibit both low miPAE and close agreement with the RFdiffusion backbone across model ranks.

In the human body, iron is distributed through the bloodstream by transferrin (Tf), a globular protein synthesized by the liver that has two iron-binding sites. Human cells have specialized transmembrane glycoproteins, primarily dimeric in structure, that facilitate cellular iron uptake via receptor-mediated endocytosis of iron-loaded transferrin, known as human transferrin receptors (TfRs: TfR1 and TfR2). For example, when TfR1 binds diferric transferrin at the cell surface, it is internalized into endosomes, releasing the iron due to the low pH, and recycling back to the surface to release apo-transferrin. As a counterpart, kinetoplastid parasites have evolved similar systems to uptake iron from their hosts [38], hence, the involved receptors can be used as targets to design MP modules.

In T. brucei, its transferrin receptor (TbTfR) is a heterodimer located at the FP, composed of the proteins ESAG6, anchored to glycosylphosphatidylinositol (GPI), and ESAG7 (Figure 2A) [30]. It does not have sequence similarity to either TfR1 or TfR2 but binds transferrin and is internalized via clathrin-coated pits. The cargo then traffics through Rab5-positive endosomes, and the receptor recycles via Rab11 to the pocket [39].

In Leishmania species, early biochemical studies detected an integral membrane glycoprotein of ~70 kDa on promastigotes as a single polypeptide that can be purified on transferrin-Sepharose and is recognized by anti-human TfR antibodies. However, its role in receptor-mediated endocytosis and iron acquisition through the incorporation of Tf remains unproven. Instead, Leishmania primarily acquires iron through reduction of ferric to ferrous iron (e.g., ferric reductases such as LFR1) followed by transport via dedicated ferrous iron transporters (e.g., LIT1). Additionally, intracellular amastigotes can access host Tf iron indirectly by recruiting TfR-positive endosomal membranes to the parasitophorous vacuole and import already released iron via iron transporters located at the FP [40].

Similarly, T. cruzi amastigotes show specific Tf binding and internalization, and an approximately 200 kDa surface protein that cross-reacts with anti-human TfR antibodies was reported in early studies [41]. Evidence indicates that Tf uptake occurs into the FP/CC system and is stored in reservosomes. For example, transferrin-gold nanoparticles were shown to localize in reservosomes and to lysosomal-related organelles after internalization [42]. Several functional studies also indicate that Tf endocytosis in T. cruzi is sensitive to disruption of the CC-associated cytoskeleton and membrane cholesterol but is largely independent of classical clathrin-mediated endocytosis, implying an alternative, clathrin-independent receptor-mediated route [43].

Unfortunately, to our knowledge, the coding sequence of putative TfRs from T. cruzi and Leishmania were never cloned or assigned to gene identifiers of any available genome. These proteins remain poorly characterized, with only limited biochemical evidence available, which prevents access to their protein sequences for structural prediction and analysis. Further evidence is needed; hence, we focused only on TbTfR.

Two experimental structures are available in the Protein Data Bank for TbTfR in complex with human Tf (Figure 2A). One of these structures contains N-linked glycans in exposed positions of the receptor, which are thought to be important for immune evasion [30]. Importantly, none of these glycosylations mediate the recognition of Tf. The binding mode involves two distinct interfaces, one for each Tf lobe. The first binding region (Figure 2B, upper panel) comprises residues from a tightly packed α-helix and β-strand of Tf that interact with residues located in loops from both ESAG6 and ESAG7. Molecular interactions include a network of H-bonds mediated by Tf sidechain atoms from residues S359, D356, R352 and H349 and backbone atoms from residues C368 and S370 (Figure S1). The only hydrophobic interaction is established by residue V360. The Tf residue H349, besides being involved in H-bond formation, engages a strong aromatic interaction (π-π stacking) with a tyrosine of ESAG6 (Y248), indicating that it is particularly important for the interaction (Figure S1). Meanwhile, the second binding region involves residues from four different segments of human Tf (Figure 2B, lower panel). H-bond forming residues are Y71 and R324, which are positioned in two different α-helixes from Tf (Figure S2). Other residues involved in electrostatic and hydrophobic interactions are located in two different loops and another α-helix. These include L72, A73, P74 and N76 in the first loop, M256 in the second loop, and K312 in the α-helix (Figure S2).

For all the protein designs in this article, the design strategy we followed was the scaffolding of binding motifs extracted from the receptor-cargo interfaces onto de novo protein backbones [15]. In this case, from the two binding regions described above, we selected the first as the motif input for the design process. This choice was motivated by the structural compactness of the involved Tf segment, given that residues 349–370 form a contiguous stretch of sequence that folds into a well-defined α-helix and β-strand, making it amenable to motif scaffolding as a single, geometrically coherent unit. In contrast, the second binding region involves residues distributed across four separate backbone segments, making it considerably more challenging to extract and scaffold as an isolated motif without losing the spatial relationships between the interacting residues. For the selected motif, the complete backbone of residues 349–370 was preserved while keeping only the sidechains of residues H349, R352, D356, S359, V360 and E367, which engage directly in contacts with TbTfR (Figure 2C, left panel). RFdiffusion was then used to generate 1000 novel protein backbones that scaffold this motif, producing compact MP structures that position the key interacting residues in the same spatial arrangement as in the native Tf-TbTfR complex. Subsequently, ProteinMPNN was used to design the most probable amino acid sequence compatible with each generated backbone, without redesigning motif residues. The resulting sequences were modeled in complex with TbTfR using AF2-multimer, yielding five predicted models per design. Designs were then scored using a scoring function (Equations 1 and 2) that rewards designs for which the greatest number of models have both a confident interaction with the receptor (low miPAE) and a faithful recapitulation of the RFdiffusion backbone (low AF2 versus RFdiffusion backbone RMSD) prioritizing designs where the sequence assigned by ProteinMPNN repeatedly and consistently folds back into the intended structure and engages TbTfR with low predicted error (Figure 2C, right panel; see Materials and methods for details).

The distribution of the resulting design metrics (miPAE, RMSD against RFdiffusion backbone, and pLDDT) is shown in the scatter plots from Figure 2D. The values for the five AF2 models for each of the two top-scoring designs are highlighted in each plot (markers connected with lines), which gave values of 1.55 and 1.46, with miPAE values of 1.65 Å and 1.86 Å and RMSD values of 1.09 Å and 0.97 Å for the rank 1 AF2 model, respectively (Figure 2D). In both cases, AF2 models closely recapitulate the RFdiffusion backbone geometry, and the designs adopt compact folds with excellent binder-receptor surface complementarity. In terms of tertiary structure, the best designs have backbones that continue the natural β-strand of the motif to form a β-sheet and extend the α-helix with a few extra residues. Notably, the design process proved highly successful, with a considerable number of the generated designs (67 candidates) surpassing the miPAE < 5 Å and RMSD < 2 Å thresholds across the ensemble, a level of success rate that, based on experimental validation studies of RFdiffusion-based binder design campaigns, is strongly predictive of retrieving true binding activity in vitro [15].

Many eukaryotic pathogens have evolved to establish and maintain proliferative populations within hosts, and some of them can even maintain a persistent infection that can last for years. These long-term infections only occur if the pathogen population can survive both the innate and adaptive immune responses [31].

Among kinetoplast parasites, African trypanosomes provide a particularly instructive example of how surface architecture supports survival within the host. As extracellular pathogens, they are continuously exposed to the adaptive and innate immune systems. To achieve persistence, they have evolved a remarkable cell surface dominated by many copies of a single variant surface glycoprotein (VSG), which is periodically switched [31]. This VSG coat both protects the plasma membrane from immunoglobulins and underpins a system of antigenic variation, allowing the population to evade the adaptive immune system [44].

However, within the VSG coat are found invariant proteins, including nutrient receptors, and a family of ISGs. Although these invariant proteins are 100- to 200-fold less abundant than the VSG, they are essential for parasite survival [45]. While VSG handles antibody evasion by switching the expressed VSG gene or through VSG recombination, the parasite relies on less abundant, constant proteins hidden within the coat to mitigate innate attacks. Recent work has shown that ISG65 functions as a receptor for the complement protein C3/C3b, an innate immune component, during the bloodstream stage (BS) of infection. Rather than merely blocking complement deposition, ISG65 operates as an active surface-clearing mechanism: it captures deposited C3b and mediates its internalization and trafficking to the lysosome. Together with the observation that ISG65 reduces C3b surface binding and uptake, this suggests that ISG65 participates in complement handling through receptor-mediated endocytosis [31, 44]. This interpretation is consistent with earlier work showing that ISG65 is internalized and sorted through clathrin-dependent, ubiquitylation-dependent endocytic pathways, a general mechanism that trypanosomes use to remove surface proteins from the plasma membrane [46]. Crucially, while the C3b cargo is destroyed in the lysosome, ISG65 itself exhibits a high recycling rate comparable to VSG, allowing it to return rapidly to the surface and sustain its protective role.

The identification of ISG65 as a critical receptor for human complement C3b has redefined the understanding of innate immune evasion in T. brucei [31, 44]. Because this protein is essential for parasite persistence and remains invariant across different strains, it represents a highly attractive target for the development of novel diagnostics or therapeutics [45].

To our knowledge, no putative ISG65 homologs from T. cruzi and Leishmania have been identified. These proteins appear to be a specialized adaptation of T. brucei, specifically restricted to the BS, which prevents structural comparisons across other major trypanosomatids [45]. However, other kinetoplastids deploy functionally analogous invariant surface proteins for immune evasion, such as gp58/68 in T. cruzi, which inhibits alternative pathway activation [47], and GP63 in Leishmania, which cleaves complement C3 to prevent lysis [48]. These examples demonstrate that utilizing constant surface components to manipulate host innate immunity is a widely conserved survival strategy among these diverse human pathogens, and may serve as analogous targets like the ISG65.

Structural analysis has demonstrated that ISG65 specifically binds to the thioester domain (TED) of C3b [31]. Furthermore, experimental evidence indicates that ISG65 is involved in reducing trypanosome susceptibility to C3-mediated clearance, effectively acting as a specialized mechanism for innate immune evasion that complements the adaptive shield of the VSG coat [44]. The structural basis for our target is derived from the crystal structure of ISG65 in complex with the human C3d domain (PDB 7PI6), which reveals the characteristic three-helix bundle architecture of the receptor’s ectodomain (Figure 3A, upper panel). While this experimental model provides high-resolution data on the binding interface, it lacks the C-terminal regions that were truncated for crystallization [31]. A complete picture of the receptor architecture can be seen in the AF2 model of the Uniprot entry Q587F5 (Figure 3A, lower panel), which includes a small, disordered N-terminal tail and a flexible linker that connects to a C-terminal transmembrane domain that anchors the receptor to the parasite’s membrane [44]. The binding mode involves a broad concave interface located within the three-helix bundle, a geometry shaped by a ~20° curvature along the bundle’s longest axis. According to the crystal structure, the natural interaction is stabilized by several conserved residues on the C3d side, including L1109, E1110, K1113, P1114, S1164, T1170, K1171, and D1174 (Figure 3B and Figure S3), which mediate thirteen hydrogen bonds, three salt bridges, and 24 residue-residue contacts [31].

As input for the motif scaffolding design, we used two stretches of backbones from C3d, 1,110–1,114 and 1,164–1,174, where key interface residues are located (Figure 3C). The idea for this design was to connect them with a backbone of variable size and extend the N- and C-terminal extremes with variable segments. We tested several strategies, and the best results were obtained by inverting the order of the motifs (1,164–1,174 first, instead of 1,110–1,114), maintaining the sidechains of residues L1109, E1110, K1113, P1114, S1164, T1170, K1171 and D1174. As before, ProteinMPNN was used to design sequences and AF2-multimer for modelling them in complex with the receptor. The results can be seen in Figure 3D, highlighting the top-two scoring designs, which gave S values of 1.44 and 0.99, with miPAE values of 1.37 Å and 1.46 Å and RMSD values of 1.29 Å and 1.56 Å for the rank 1 AF2 model. In both cases, AF2 models closely recapitulate the RFdiffusion backbone geometry, and the designs adopt compact folds with good binder-receptor surface complementarity. In this case, the total number of successful designs was 17 (miPAE < 5 Å and RMSD < 2 Å), which was lower than for TfR, but enough to suggest true binding activity in vitro among the candidates.

Heme plays a fundamental role in many cellular processes for the vast majority of living organisms. Structurally, it consists of an iron ion coordinated within a tetrapyrrole ring known as protoporphyrin IX. It is an essential cofactor for proteins involved in oxygen transport and storage, mitochondrial electron transport, drug and steroid metabolism, signal transduction, and antioxidant defense [49]. However, when free, it becomes highly toxic due to its capacity to generate free radicals and cause oxidative damage. In the human blood, heme is normally contained within hemoglobin (Hb) in red blood cells. During intravascular hemolysis, released Hb dissociates into reactive αβ dimers that are neutralized by haptoglobin (Hp), an acute-phase glycoprotein that binds Hb dimers in a virtually irreversible manner, forming the HpHb complex, subsequently cleared by macrophages through CD163-mediated endocytosis, enabling heme degradation and iron recycling [50, 51].

Unlike humans and most eukaryotic organisms, which synthesize heme de novo, kinetoplastid parasites present defects in their biosynthetic pathway. Some species have lost the complete pathway, while others retain only the last three biosynthetic steps [52]. As heme auxotrophs, kinetoplastids have evolved to acquire this essential nutrient from their hosts. For example, it has been established that transmembrane proteins belonging to the heme responsive gene (HRG) family are involved in the transport of free heme from the environment in kinetoplastids [53]. In T. brucei, heme acquisition is highly adapted to each stage of its life cycle. During the procyclic stage (PS) in the tsetse fly midgut, heme is obtained from degraded hemoproteins via the TbHRG transporter, localized to the flagellar membrane and the FP [49, 54]. On the other hand, in the BS, the primary route of heme acquisition is bound to Hb, though the parasite’s HpHbR, located in the FP, captures HpHb complexes directly from the bloodstream, similarly to the CD163-mediated endocytosis [55]. Once in endo-lysosomes, HpHb is degraded and heme released, where TbHRG would be responsible for transporting it to the cytosol, making it available for essential cellular processes [56]. While heme acquisition mechanisms vary between different kinetoplastid parasites, here we focused on designing binders for the HpHbR from T. brucei.

Two experimental structures are available in the Protein Data Bank for the T. brucei HpHbR in complex with human HpHb (PDB 5HU6 and PDB 4WJG) (Figure 4A) [33, 34]. This receptor is an elongated protein composed of a three-helix bundle that attaches to the plasma membrane through a GPI anchor covalently bound at its C-terminus. At its membrane-distal end, the receptor widens to form a compact head structure that includes the N-terminus and a 42-residue loop containing two additional helices (Figure 4A, upper panel). Although concentrated within the FP, HpHbR molecules retain lateral mobility within the membrane. This mobility is thought to compensate for the relatively small interaction surface presented by the receptor’s narrow and elongated architecture. Indeed, high-affinity ligand recognition is achieved through a bivalent binding mechanism in which two receptor molecules adopt the appropriate orientation to engage a single dimeric HpHb complex, thereby increasing overall binding avidity (Figure 4A, lower panel) [33, 57].

The interaction between HpHbR and its ligand is mediated by two distinct binding regions (Figure 4B). The first binding interface corresponds to the βHb-binding site (Figure 4B, upper panel) and is formed by residues from helices I and II of HpHbR, together with additional contacts contributed by the loop connecting helices III and IV. Notably, the interaction surface involves direct molecular interactions with the heme of βHb (Figure S4), where its carboxylate moieties participate in receptor recognition through several H-bonds. This observation is consistent with previous biophysical studies demonstrating that the presence of heme is required for efficient binding of the HpHb complex by HpHbR, suggesting that direct heme recognition is necessary to ensure the intake of this compound. Besides the interactions mediated by heme, the amino acid residues R41, F42, E44 and S45 from an α-helix of βHb and L92, K96 and L97 from another are involved in the interaction with the receptor. The second binding interface corresponds to the Hp-binding region (Figure 4B, lower panel), which is mainly mediated by residues located in helices I and VI of HpHbR. In this case, no heme is involved, only amino acid residues from three different segments of Hp. Residues V275, G276, V278, Y280 are located in a loop and β-strand, residues M300 and D305 in another loop and β-strand that forms a β-sheet with the first one, and residues K345 and Y346, located in a turn (Figure S5).

For the design process, we tested several strategies without success. For example, we tried to design over the motif from Figure 4C (left panel), conserving the side chains of residues from βHb involved in the first binding region, filling the gap left by heme with a de novo backbone. We also tried using as input the motif from Hp involved in the interaction with the receptor (Figure 4D, left panel). In both cases, we obtained very low scores for the top-scoring designs (0.32 and 0.39, respectively), with very low reproducibility of the metrics across the AF2 models (Figure 4C and 4D, right panel). We even tried to perform several rounds of completely de novo designs (without motif scaffolding), obtaining similar results (data not shown), indicating that this is a particularly difficult target.

For comparison, Figure 5 shows the distribution of AF2 error (as miPAE) and structural discrepancy between AF2 and RFdiffusion (as RMSD) of the rank 1 AF2-multimer models. We can see how the Tf mimetic designs accumulate the most successful models in the lower-left quadrant (67 candidates), followed by ISG65 binders (17 candidates). Meanwhile, binders designed for HpHbR retrieved only one successful candidate for each of the binding regions, but close to the threshold limits.