A Fusion AI Framework for Accelerated Therapeutic Design: An In-Silico Research-Use-Only Doctrine

Abstract

CrisPRO.ai, an in-silico research-use-only (RUO) framework designed to accelerate drug discovery by fusing the capabilities of discriminative and generative Artificial Intelligence. Our platform orchestrates a generalist genome foundation model with a suite of specialist predictors and structural oracles to
⁠
⁠This paper outlines the doctrine for our framework, detailing the benchmarked evidence that forms its scientific backbone and the end-to-end workflow that transforms a biological hypothesis into a de-risked, therapeutic cure, accelearting R&D

Our results demonstrate that this fusion approach not only achieves SOTA performance across multiple benchmarks but also provides a comprehensive, transparent, and controllable system for in-silico drug discovery.
⁠
1.Introduction

The promise of precision medicine is fundamentally limited by our ability to interpret the functional consequences of genetic variation. CrisPRO.ai was conceived to address this challenge by creating an orchestration layer that combines a genome‑scale foundation model (Evo 2) with specialist predictors (e.g., AlphaMissense) and structure/epigenome oracles (e.g., AlphaFold 3, Enformer) [1], [2], [3]. The result is a system that can not only interpret the full spectrum of genetic variation but can also generatively design novel therapeutic constructs.

This paper details the scientific and operational doctrine of the CrisPRO.ai RUO framework, which is intended for research use only and not for diagnostic or therapeutic decision-making without independent validation. We present the evidence-backed benchmarks that validate our capabilities and outline the workflow that we employ to accelerate therapeutic research.

2. Evidence Backbone

The scientific integrity of the CrisPRO.ai platform is grounded in a rigorous set of benchmarks derived primarily from the capabilities of the Evo 2 genome foundation model [1].

2.1. Discriminative AI Capabilities

Our system's ability to interpret genetic variants is validated against multiple gold-standard datasets:

• Comprehensive Variant Coverage: While specialist models like AlphaMissense show superior performance on coding SNVs, Evo 2 establishes SOTA zero‑shot performance on coding non‑SNVs (indels) and non‑coding variants, providing critical coverage where other models fail [1, Fig. 1].
• Oncology-Specific Accuracy: For BRCA1/2 variants, a lightweight supervised classifier trained on Evo 2 embeddings achieves an AUROC of approximately 0.95, showcasing the power of our fusion approach for key clinical targets [1, Methods 4.3.16].
• Functional Genomics Correlation: Evo 2 likelihoods correlate strongly with experimental fitness from Deep Mutational Scanning (DMS) and are the only model scores to show the expected negative correlation with human mRNA decay rates [1, Fig. 2E, S3G].

Figure 1. Evo 2 enables accurate human clinical variant effect prediction.

This figure summarizes Evo 2's state-of-the-art zero-shot performance. (C) Zero-shot evaluation of variant pathogenicity within the noncoding regions (N = 34,761 SNVs; N = 3,894 non-SNVs) shows superior performance. (D) Zero-shot evaluation on splice-altering variants in SpliceVarDB, split by exonic (N = 1,181) and intronic (N = 3,769) scoring, also demonstrates SOTA performance. This proves its ability to interpret complex, non-coding regions where other models fail. (Adapted from Brixi, G., et al., 2024).

2.2. Generative AI Capabilities

Our platform's ability to design novel biological constructs is demonstrated by several key generative achievements:

• High-Fidelity Genome Generation: The system can generatively create coherent, full-length sequences for minimal prokaryotes (~580 kb) with a ~70% Pfam-hit rate, a dramatic improvement over the ~18% achieved by previous models [1, Fig. 5H].
• Predictable Epigenomic Design: In a demonstration of precise functional control, we show a predictable, log-linear relationship between computational budget and design fidelity for creating sequences with specific chromatin accessibility patterns [1, Fig. 6C].

3. The Fusion Approach

The competitive advantage of CrisPRO.ai lies in our fusion approach. We combine the generalist genome LM (Evo2) with specialist models to achieve SOTA across the entire R&D continuum.

• Discriminative Stack: Evo2 provides a baseline for all variants, with specialists like AlphaMissense layered in for coding SNVs.
• Generative Stack: Evo2 generates sequence proposals, which are then scored for function by oracles like Enformer/Borzoi and for structure by AlphaFold 3.

This approach gives us Breadth (covering all variant types), Depth (achieving SOTA on key targets like BRCA1), and Control (designing sequences with predictable functional properties).

Our platform operates as an orchestration layer. A generalist genome foundation model (Evo 2) provides broad, first-principles understanding. This is fused with specialist discriminative oracles (e.g., AlphaMissense) for high-accuracy scoring and generative oracles (e.g., Enformer, AlphaFold 3) for the de novo design and validation of therapeutic constructs.

4. End-to-End RUO Workflow

Our platform operationalizes these capabilities into a structured, seven-step in-silico workflow:

1. Problem Framing & Data Curation
2. Target Assessment (Discriminative)
3. Mechanistic Triage & Hypothesis
4. Design (Generative)
5. In-Silico Validation
6. Feedback & Calibration
7. Reporting & Provenance

5. Discussion and Governance

Our fusion framework provides a uniquely powerful tool for RUO therapeutic design. We position our system as an R&D assistant that augments human expertise, not a replacement for it. All outputs are flagged as RUO, and we enforce a mandatory wet-lab validation loop for any designed sequences. Safety is paramount; the system has intentionally reduced capabilities on human viral proteins, and all processes include audit trails and viral prompt guards.

6. Conclusion

The CrisPRO.ai RUO framework represents a significant step forward in the field of in-silico drug discovery. By grounding our platform in a rigorously benchmarked, evidence-backed doctrine, we have created a system that can reliably and transparently accelerate therapeutic research. This fusion of discriminative and generative AI transforms the process of drug development from a high-risk gamble into a predictable, high-certainty engineering discipline, building a clear, evidence-backed plan that makes the impossible, possible.