Computational De Novo Design of Miniprotein Candidate Neutralizers Targeting the IL-23 Cytokine Subunit p19

Raghav Paladugu; Fardina Fathmiul Alam

q-bio.BMOct 23, 2026

Computational De Novo Design of Miniprotein Candidate Neutralizers Targeting the IL-23 Cytokine Subunit p19

Authors: Raghav Paladugu, Fardina Fathmiul Alam

Organizations: Department of Computer Science, University of Maryland, College Park, MD, USA

Abstract

Interleukin-23 (IL-23) drives T-helper-17 inflammation underlying inflammatory bowel disease and psoriasis; anti-p19 antibodies are clinically effective but large, costly, and injectable. De novo miniproteins are an attractive alternative, yet published designs target the IL-23 receptor; the IL-23 cytokine itself has not been addressed. A further obstacle is that the standard confidence score (ipTM) tracks construct length rather than interface quality. We computationally design miniprotein binders against the IL-23R-binding face of p19 (PDB 3DUH) and score every design with ipSAE, a length-robust interface metric derived from AlphaFold-initial-guess predicted aligned error. A conventional RFdiffusion/ProteinMPNN/ColabFold pipeline yields no reliable predicted interface across 972 designs (best ipSAE 0.30), and a scoring artifact, in which single-sequence ColabFold under-rates binder pLDDT by approximately 56 points, explains an earlier misleading "binders do not fold" signal. An interface-optimized hallucination method (BindCraft), scored identically, yields 42 of 50 designs above ipSAE 0.60 (best 0.804). An AlphaFold-independent geometric check reproduces the conventional failure: no top-ranked conventional design reaches the buried area or interface hydrogen-bond count of the weakest lead. Five all-alpha-helical leads (72-104 residues) collectively engage all six targeted hotspots; two dock within the intact p40/p19 heterodimer and all show favorable Rosetta energetics. A p40-alone counter-screen returns ipSAE 0.00 for all five, indicating no spurious p40 interface. To our knowledge, these are the first computationally proposed de novo miniprotein candidates designed to occlude the IL-23R-binding face of IL-23 p19. All results require experimental validation.

Figures & tables

Panel of Figure 1 — Fig. 1 : Pipeline overview and the central result. Both tracks begin from the same IL-23 p19 target and the same six IL-23R-face hotspots, and every design from both tracks is judged by one identical interface scorer (AlphaFold-initial-guess PAE processed into ipSAE, asymmetric maximum), so the comparison reflects the design method rather than differences in evaluation. Reading the outcome boxes against the design counts above them shows the conventional pipeline did not fail for want of designs.

Fig. 2 : The single-sequence ColabFold scoring artifact. For 50 designs the same backbone is folded two ways: single-sequence ColabFold rates the binders as essentially disordered (mean pLDDT 30.1) while AlphaFold-initial-guess, seeded with the designed pose, rates them well-folded (mean 86.3). The gap is a property of the scorer. Source data: Supplementary S6.

	Conventional	BindCraft
Metric	( $n{=}20$ )	leads ( $n{=}5$ )
dSASA (Å 2 )	1149 (144–1755)	2204–3018
Shape compl.	0.44 (0.07–0.70)	0.62–0.69
Interface H-bonds	0 (0–3)	3.5–10.5
Meets all three	0/20	5/5

TABLE I : AlphaFold-independent interface geometry for the 20 top-ranked conventional designs with retained structures; median (range). No conventional design reaches the weakest lead on buried area or hydrogen bonds, so the pipeline contrast does not depend on AlphaFold-derived scoring. Conventional values are scored against the merged p40/p19 target, which offers more surface to bury, making the comparison conservative. Metrics are geometric only;

\Delta

G is excluded (Methods G).

Fig. 3 : ipSAE distribution for every design from both pipelines, shared axis and logarithmic count scale. The conventional distribution is collapsed against zero, reaching neither the 0.60 threshold (black dashed) nor the 0.70 band (green dashed), whereas BindCraft sits largely above both. Because the conventional designs nonetheless have high predicted monomer confidence (Fig. 2 ), the failure is localized to the interface rather than backbone quality. Source data: Supplementary S5.

ipSAE cutoff	Conventional ( $n{=}972$ )	BindCraft ( $n{=}50$ )
$\geq$ 0.50	0	45
$\geq$ 0.60	0	42
$\geq$ 0.70	0	38
$\geq$ 0.80	0	2

TABLE II : Threshold-sensitivity of the pipeline contrast: designs at or above each ipSAE cutoff. The conventional pipeline yields zero at every cutoff, so the contrast does not depend on the 0.60 choice.

Panel of Figure 6 — Fig. 4 : BindCraft outcome under the shared scorer. (a) Distribution of binder–p19 ipSAE over the 50 designs; dashed lines mark the 0.60 and 0.70 thresholds, the conventional best (dotted) sits far left. (b) ipSAE versus binder length by regime; rings mark the five leads.

#	Design	Regime	Len	ipSAE	minPAE (Å)	RMSD (Å)	Hotspots (n/6)	%Hel	$\Delta$ G (REU)	dSASA	SC	HB	Cont.
1	l94_s482991_mpnn8	p19-only	94	0.804	0.94	1.69	51,80,81,89 (4)	91	$-57.5$	2204	0.68	3.5	35
2	l104_s309115_mpnn6	p19-only	104	0.801	0.99	1.50	51,137,140 (3)	91	$-84.9$	2962	0.69	10.5	34
3	l72_s467308_mpnn8	p19-only	72	0.785	0.92	1.37	80,81,89 (3)	88	-87.3	3018	0.68	8.5	39
4	l92_s675453_mpnn18	full	92	0.789	1.05	3.17	51,80,81,89,137 (5)	78	$-77.0$	2960	0.62	9.5	28
5	l83_s957191_mpnn2	full	83	0.763	1.12	1.86	51,80,81,89 (4)	81	$-75.0$	2943	0.67	9.0	22

TABLE III : Five lead candidates and their predicted properties. Every candidate clears the

\geq

0.70 confident band while presenting a large buried interface and favorable

\Delta

G, and the metrics disagree on the best lead (l94 by ipSAE, l72 by

\Delta

G).

\Delta

G in REU; dSASA in Å 2 ; SC

=

shape complementarity; HB

=

interface hydrogen bonds; Cont.

=

contacts within 5 Å; RMSD

=

BindCraft hotspot RMSD. Sequences in Supplementary S1. †

Panel of Figure 8 — Fig. 5 : Lead characterization. (a) Epitope coverage: filled circles mark the hotspots each lead engages (binder heavy atom within 5 Å, 3DUH chain-C numbering). No single lead touches all six, but the five together cover every one. (b) The five leads across ipSAE, interface PAE, epitope coverage, helix fraction and contact footprint, normalized to range; profiles are mutually distinct.

Panel of Figure 9 — Fig. 6 : Predicted structures of the five lead complexes (PyMOL). Each binder (blue) is shown on p19 (gray) with interface residues within 5 Å as sticks; p40 (tan) is shown for the two full-heterodimer leads. l92 docks on the p19 receptor face even with p40 present, the configuration most relevant to neutralizing intact IL-23.

Lead	ipSAE (p19)	ipTM (p40)
l94	0.804	0.12
l104	0.801	0.12
l72	0.785	0.11
l92	0.789	0.12
l83	0.763	0.12

TABLE IV : Counter-screen of the five leads against p40 alone. The p40 ipSAE comes from co-folding each binder with p40 alone and scoring with the identical metric. Every lead collapses to 0.00, far below its p19 value and the 0.60 threshold. Raw scores: Supplementary S3.

Protein Binder Design

Protein binder design creates proteins, antibodies, and peptides that bind selected molecular targets with desired specificity and function.

See all 23 papers

Natural-Language-Guided Generator-Agnostic Shortlisting for Protein Binder Design
Aug 21, 2026Gyubok Lee, Kiwoong Yoo, Jimin Seo +3LLM Reranking Bioinformatics
Kim Jaechul Graduate School of AI, Korea Advanced Institute of Science and Technology (KAIST), Daejeon, South Korea · LG AI Research, Seoul, South Korea · Department of Electrical and Computer Engineering, Seoul National University, Seoul, South Korea +1
Scaling Atomistic Protein Binder Design with Generative Pretraining and Test-Time Compute
Mar 30, 2026Kieran Didi, Zuobai Zhang, Guoqing Zhou +11Protein Binder Design Protein Design
NVIDIA · University of Oxford · Mila - Québec AI Institute +8
Joint Discrete and SE(3) Diffusion for Protein Interface Design
Oct 10, 2026Zhenqiao Song, Jianwen Xie, Lei LiProtein Binder Design Molecular Generation
Language Technologies Institute, Carnegie Mellon University · Lambda Inc.

Aug 21, 2026J/K move · Enter open · S save

Natural-Language-Guided Generator-Agnostic Shortlisting for Protein Binder Design

Gyubok Lee, Kiwoong Yoo, Jimin Seo, Jiyoun Kim+2

Kim Jaechul Graduate School of AI, Korea Advanced Institute of Science and Technology (KAIST), Daejeon, South Korea · LG AI Research, Seoul, South Korea · Department of Electrical and Computer Engineering, Seoul National University, Seoul, South Korea+1

LLM Reranking Bioinformatics Protein Binder Design

Open paper AskPDF arXiv

Modern de novo design workflows generate many candidate protein binders, but wet-lab validation capacity remains limited, making shortlisting a major bottleneck. We study whether LLMs can generate multi-metric ranking policies from precomputed structural-confidence and interface-quality proxy scores. Rather than proposing a new protein binder design pipeline, we focus on post-generation binder shortlisting: selecting the final top-K candidates from already generated binder pools using a shared panel of precomputed proxy scores. On the 10-target held-out split, averaging performance over five sampled global iterative gpt-4o policies reaches 0.589 Recall@10, modestly improving over the strongest single-feature fixed baseline, Protenix binder ipTM, which reaches 0.571 Recall@10. On the 3-target held-out subset comprising Nipah, RBX1, and TREM2, target-conditioned iterative gpt-5.4 policies reach the strongest LLM performance, with 0.519 Recall@10 and 0.583 NDCG@10. These results suggest that LLM-generated ranking policies can act as an interpretable post-generation decision layer for combining heterogeneous proxy metrics to prioritize binders from large candidate pools.

Abstract

Protein Binder Design

Explore similar work

Natural-Language-Guided Generator-Agnostic Shortlisting for Protein Binder Design

Scaling Atomistic Protein Binder Design with Generative Pretraining and Test-Time Compute

Joint Discrete and SE(3) Diffusion for Protein Interface Design