The Rise of the Open-Source Frontier: Why We Cross-Validate Across AlphaFold 3, ESMFold2, and Boltz

In the rapidly evolving landscape of structural biology and AI-native drug discovery, the frontier is moving faster than ever. Almost weekly, new computational models are released, each claiming to outperform previous benchmarks on critical macromolecular structure prediction tasks. For a lean, mission-driven startup like AlgorithmicRx, this explosion of open-science tools is a massive equalizer. But it also introduces a critical scientific question: How do we separate model-specific artifacts from true biological signals?

At AlgorithmicRx, we are building an early-stage computational pipeline focused on Duchenne Muscular Dystrophy (DMD). Our strategy centers on designing "molecular rescue" candidates—motif-grafted peptide libraries engineered to bind the exposed pocket of β-dystroglycan, restoring a vital mechanical linkage that is broken when the dystrophin gene is mutated or truncated. Because we operate in silico prior to wet-lab validation, our designs are only as good as the structural models we use to evaluate them.

To build high-confidence, translatable designs, we have adopted a strict multi-model orchestration approach: we never rely on a single model's predictions. Instead, we systematically cross-validate every candidate binder across AlphaFold 3, ESMFold2, OpenFold3, and Boltz-1/Boltz-2. Here is why this open-source multi-model strategy is essential for the future of rare disease drug design.

The Peril of Single-Model Bias

In molecular machine learning, it is incredibly common for a model to overfit to its training set or produce high-confidence structures that are actually computational artifacts. This "single-model bias" is a major strategic risk. If an in silico pipeline relies solely on one predictor (even a state-of-the-art closed platform like AlphaFold 3), a researcher might spend precious resources pursuing a candidate binder that only "looks" good to that specific neural network.

This risk is magnified in rare and ultra-rare diseases where biological uncertainty is highest and data is sparse. For example, simulating how the Dystrophin-Associated Protein Complex (DAPC)—including dystrophin, β-dystroglycan, the sarcoglycan complex, and syntrophin (SNTA1)—fits together in healthy versus mutated disease states requires highly sensitive interface modeling. Relying on a single model's prediction of these intricate protein-protein interfaces can lead to false positives.

The Rising Tide of Open Source: Outperforming the Giants

For a long time, the highest-tier structure prediction was the exclusive domain of proprietary, centralized platforms. However, the open-source and open-weights communities have begun to match and, in some cases, surpass these closed models.

Independent benchmarks on challenging, out-of-distribution structures reveal that open-science models are yielding major accuracy improvements. For example:

• On FoldBench v1 (evaluating 172 interfaces across 113 complex assemblies), the published AlphaFold 3 baseline lands at a 47.9% DockQ success rate. By comparison, the open-weights model ESMFold2 achieves a 71.1% success rate on identical inputs.

• On the Elofsson/Fromm benchmark—a strict, out-of-distribution test with a September 2021 training cutoff designed to prevent data leakage—AlphaFold 3 comes in at 60.9% success, while ESMFold2 clears it at 65.5%.

These benchmark results are highly consequential for downstream de novo drug design. Most cutting-edge protocols use structure predictors as "oracles" to evaluate design quality and iteratively refine binders. Better, more accurate open-source evaluators translate directly into higher-fidelity candidates with significantly higher downstream wet-lab hit rates.

Our Cross-Validation Protocol: Measuring Structural Convergence

By combining both proprietary and open-weights models into a unified orchestration layer, AlgorithmicRx avoids the vulnerabilities of single-model dependency. Our computational pipeline acts as a rigorous filter, scoring designed peptides through a sequential multi-model loop:

1. Target Modeling & Design: We map the disease-state consequences of dystrophin truncations (e.g., Exon 50 deletions) and generate custom candidate rescue-peptide libraries targeting the exposed β-dystroglycan interface.

2. Multi-Model Folding: Every promising design is folded and evaluated against the target pocket across AlphaFold 3, OpenFold3, ESMFold2, and Boltz-2.

3. Structural Convergence Checks: We measure success not by one model's score, but by structural convergence—the proportion of interface predictions where independent models like OpenFold3, Boltz-2, and AlphaFold agree on the 3D binding pose within strict confidence thresholds.

4. Interface Quality Analysis: We analyze the predicted binding confidence (ipTM), interface reliability (pLDDT), predicted alignment error (PAE), and inter-chain contact hotspots across all models to compile a unified, evidence-graded consensus.

When we see high-confidence signals across multiple independent architectures (for instance, an ipTM ≈ 0.85 backed by high-confidence interface predictions in both AlphaFold 3 and ESMFold2), we know we have a robust structural hypothesis worth bringing forward.

At AlgorithmicRx, we believe the future of AI in medicine won't be dominated by a single closed model. Instead, it will be driven by the open-source community—a rising tide that lifts all boats, enabling small, patient-centered teams to tackle complex diseases like DMD with world-class scientific rigor.

Disclaimer: AlgorithmicRx is an early-stage research project. All results described are computational/in silico predictions and have not been experimentally or clinically validated. This is not a medical treatment and should not be interpreted as therapeutic advice.