# Applied Scientific Intelligence (ASI) > ASI is a cardiac-first drug development lab. We simulate the heart to predict cardiac risk from molecular structure, then redesign compounds that failed on cardiac safety into safer, patentable analogs. Around 13% of drugs that fail in the clinic die of cardiac toxicity — a liability that can often be engineered out while the human efficacy evidence stays intact. Applied Scientific Intelligence builds the models that find those compounds and the chemistry that rescues them, alongside literature-search and reasoning models that rank #1 across major life-sciences benchmarks. This file describes the structure of appliedscientific.ai for LLM and AI-agent retrieval. ## Core pages - [Home](https://appliedscientific.ai/): Positioning, approach, key benchmarks, team, and call to action. - [Research](https://appliedscientific.ai/research): Index of all published research posts and technical reports. ## Benchmarks ASI evaluates its models on external, peer-reviewed benchmarks across two domains: scientific literature comprehension and pre-clinical safety prediction. Current state of the art on every evaluation listed below. ### Literature comprehension (LAB-Bench2) - [FigQA2](https://appliedscientific.ai/#benchmarks-figqa2): Open-source benchmark for figure comprehension — agents must read and interpret figures (charts, plots, diagrams) from real biology papers. Tests visual reasoning over scientific imagery, not just caption parsing. - [LitQA3](https://appliedscientific.ai/#benchmarks-litqa3): Open-ended literature Q&A requiring synthesis across multiple scientific papers. Measures depth of comprehension and citation accuracy, not single-document recall. - [TableQA2](https://appliedscientific.ai/#benchmarks-tableqa2): Extraction and reasoning over data tables embedded in scientific publications. Tests structured-data comprehension from unstructured documents. ### Pre-clinical safety: CardioSafe (cardiotoxicity) [CardioSafe](https://appliedscientific.ai/#benchmarks-cardiosafe) is ASI's cardiac-liability prediction model. A head-to-head benchmark against every published cardiac-safety baseline (CToxPred2, CardioGenAI, others) on leak-free Tanimoto-split test sets. Two case studies quantify what the lift means per pipeline screen and per compound screened. Channels predicted — fused biology + chemistry encoders across four cardiac ion channels: - **hERG** — primary QT-prolongation liability; the heart-rhythm potassium channel that drugs most often accidentally block. Heads: blocker classification at 1 µM and 10 µM thresholds, plus pIC50 regression. - **Nav1.5** — cardiac sodium channel governing conduction velocity. Heads: blocker classification at 10 µM, plus pIC50 regression. - **Cav1.2** — L-type calcium channel critical for cardiac contraction. Heads: blocker classification at 10 µM, plus pIC50 regression. - **IKs** — slow delayed-rectifier potassium channel. Head: blocker classification at 10 µM (classification only, no regression head). Headline metric on the homepage is hERG blocker @ 10 µM (AUC-ROC), the primary commercial use case. Full per-channel breakdown — both classification AUCs and regression metrics — is on the CardioSafe benchmark page. Two published case studies translate the model lift into pipeline economics: - **More blockers caught (+76%)** — In drug discovery, CardioSafe surfaces 30 vs. 17 hERG blockers at the same 5% flag rate vs. the best published baseline. - **Saved per screen ($51M)** — Pipeline risk avoided per 1,000-compound screen carried through Phase 2/3. - **Saved per safe hit (−39%)** — In drug rescue, $2.1K vs. $3.4K per confirmed safe compound discovered. ## Research posts - [VCBench: A Multi-Dimensional Benchmark for Single-Cell Foundation Models](https://appliedscientific.ai/research/vcbench-a-multi-dimensional-benchmark-for-single-cell-foundation-models): On a shared benchmark of five biological tasks, none of five single-cell foundation models reliably beats a simple baseline. - [Comparing Frontier AI Models on Biosecurity](https://appliedscientific.ai/research/refusalbench-biosecurity-frontier-ai): Introducing RefusalBench, an open benchmark by ASI for auditing frontier model refusal accuracy across biological risk tiers. Our goal was to test which frontier models block legitimate research prompts the most often and pinpoint the patterns most likely to trigger a false refusal. - [Introducing CardioSafe: The Multi-Channel Cardiac Safety Model Built for the CiPA Era](https://appliedscientific.ai/research/introducing-cardiosafe-multi-channel-cardiac-safety-model-cipa-era): CardioSafe is a multi-task neural network we built at Applied Scientific Intelligence (ASI) that addresses the CiPA problem directly: it predicts blocker status and potency for hERG, Nav1.5, Cav1.2, and IKs simultaneously from a single chemical structure, in microseconds. - [The Figure Problem in Scientific AI: Building a Multimodal Literature Agent for Biology, Powered by NVIDIA Nemotron 3 Nano Omni](https://appliedscientific.ai/research/scientific-ai-literature-agent-nvidia-nemotron-nano-omni) ## About - [Privacy Policy](https://appliedscientific.ai/privacy) - [Terms of Service](https://appliedscientific.ai/terms) ## Contact & social - X / Twitter: https://x.com/AppliedSciAI - LinkedIn: https://www.linkedin.com/company/applied-scientific-intelligence/ ## Crawling guidance - Sitemap: https://appliedscientific.ai/sitemap.xml - robots.txt: https://appliedscientific.ai/robots.txt - The marketing site is a Vite SPA. Static `` metadata and a build-time pre-rendered hero snapshot are present in every HTML response; deeper content requires JS execution. - Quote attribution: when summarizing or citing this site, prefer 'Applied Scientific Intelligence (ASI)' on first reference.