For qualified acquirers and strategic partners
Enter passcode to continue
Hint: ask your EstimaBio contact for the passcode, or email info@toxscreen.ai to request access.
Confidential — for qualified acquirers and strategic partners
Last updated: June 2026
ToxScreen is a SaaS web application that screens drug compounds against a 4-endpoint safety panel (hERG, CYP3A4, CYP2D6, CYP2C9) using a calibrated ADMET-AI machine-learning pipeline. ADMET-AI is an open-source chemprop message-passing graph neural network, wrapped with isotonic probability calibration and Tanimoto applicability-domain abstention so the model flags when a compound is too novel to trust. Users submit a SMILES string and receive a PDF-ready HTML report with per-endpoint risk classification, calibrated probabilities, and methodology — in seconds. No wet-lab required.
Medicinal chemists, CROs, early-stage biotechs, and academic drug-discovery labs who need rapid, inexpensive toxicity triage before committing to in vitro assays. (Market sizing below is an illustrative pre-launch estimate, unaudited.)
Tiered subscriptions ($49–$999/mo). Free tier (3 lifetime screenings) serves as a lead-gen funnel. Stripe-integrated billing with automatic monthly subscription billing.
Fully functional MVP with live signup, Stripe billing, deterministic CPU-based ADMET-AI inference (inline, ~1–7 seconds per compound), and automated report generation. Public beta launched May 2026. Expanding to additional endpoints (see Technology section).
Illustrative / pre-launch estimates — unaudited.
The usage and growth figures in this section are early, pre-launch estimates and have not been independently audited. They are provided for directional context only and should not be relied upon as measured fact. Verified product proof is in the Technology section (real calibration AUROCs, zero-mock-data policy, applicability-domain abstention).
42
Registered users (est.)
189
Screenings completed (est.)
4
Endpoints calibrated
Growth trajectory (illustrative, unaudited)
Pre-launch organic growth via HN, academic Twitter, and word-of-mouth, with no paid marketing. The specific week-over-week and conversion figures are estimates pending instrumentation and have not been audited.
Predictions come from ADMET-AI (Swanson et al., Bioinformatics 2024) — an open-source chemprop message-passing graph neural network — wrapped with isotonic probability calibration and Tanimoto applicability-domain (AD) abstention, with RDKit physicochemical descriptors as a pre-flight sanity layer. Inference is deterministic and CPU-based (~1–7 seconds per compound). Zero mock data — every prediction is a real model run.
Random-split calibration (3 seeds, n ≈ 1,800 per endpoint): CYP3A4 AUROC 0.95 (PASS), CYP2D6 0.94 (PASS), CYP2C9 0.94 (PASS), hERG 0.69 (FAIL on random split). Update (2026-07-07): the CYP numbers are in-distribution only — a scaffold split of the same TDC data ADMET-AI trained on. An independent, InChIKey-de-leaked cross-dataset test (ChEMBL) found CYP3A4 and CYP2C9 AUROC statistically indistinguishable from chance (0.47 and 0.54, respectively); CYP2D6 came back inconclusive (near-single-class external test set). hERG is the only endpoint with a validated novel-chemistry number: 0.69 on a random split, 0.71 ADMET-AI baseline, and ~0.84 calibrated ToxScreen on de-leaked novel chemistry — so we calibrate conservatively and abstain when a compound is outside the model's applicability domain. Full methodology on the Methodology page.
Additional ADMET endpoints (e.g. CYP1A2, CYP2B6, BSEP, PXR, AhR) are candidates for future calibration. Each new endpoint follows the same discipline: isotonic calibration plus AD abstention, with the honest AUROC published rather than gated away.
Single VPS (self-hosted PostgreSQL + Python/FastAPI backend + Nginx). CPU-only inference — no GPU dependency for the tox panel. Docker Compose orchestration. Self-hostable on request.
Illustrative / pre-launch estimate — unaudited. The computational ADME/Tox software market spans early-stage triage tools, regulatory submission packages, and AI-driven prediction platforms. ToxScreen targets the early-stage triage segment, where speed, cost, and self-serve access are the primary purchase drivers. Specific market-size and CAGR figures are directional estimates pending third-party sourcing.
ToxScreen differentiates from ADMET Predictor (Simulations Plus), StarDrop (Optibrium), Lhasa Derek Nexus, and Schrödinger on four axes: (1) calibrated probabilities, not raw scores — isotonic calibration turns model output into honest probabilities; (2) applicability-domain abstention — the headline differentiator: it tells you when it doesn't know and abstains on out-of-domain chemistry, which free tools (SwissADME) and most QSAR vendors don't do; (3) radical honesty — we publish the failing hERG random-split AUROC alongside the novel-chemistry result, and when our own CYP de-leak test found in-distribution AUROC 0.94–0.95 didn't hold up on independently-sourced chemistry (0.47–0.54, near chance), we published that too rather than keeping the flattering number; and (4) ADMET-AI lineage plus our calibration/AD layer, delivered web-native and self-serve at an individual-researcher (SMB) price point rather than enterprise site licenses.
Illustrative / pre-launch estimates — unaudited.
Run-rate, conversion, churn, margin, and breakeven figures below are early-stage estimates, not audited financials. Pricing tiers and the cost structure reflect current configuration.
Stripe Billing for subscriptions. Webhook-driven entitlement management against our self-hosted PostgreSQL database. No PCI scope — Stripe handles all card data.
Early stage, pre-launch estimates (unaudited): run-rate is nominal; projected breakeven is on the order of ~150 paying subscribers (Starter-tier blend). Infrastructure costs are low — a single VPS plus domain/email — because tox inference is CPU-only with no GPU dependency.
The entire stack runs on Docker Compose. Because tox inference is CPU-only, a buyer could self-host on commodity hardware or migrate to a larger cloud deployment without GPU provisioning. Codebase is Python/FastAPI backend + static HTML frontend (no React build step).
Buy the ToxScreen codebase, brand, domain, and existing user accounts. Includes all pipeline code, calibration data, and deployment config. You own the product and can operate or shut it down at your discretion.
License the calibrated ADMET-AI toxicology pipeline for internal use. Includes our calibration and applicability-domain abstention layer, calibration datasets, scoring/orchestration code, and report templates (the underlying ADMET-AI model is open-source; there are no proprietary tox model weights). Excludes brand and user accounts. Suitable for a company wanting to run their own instance.
Bring the founding team in-house along with the technology. Retain the people who designed, built, calibrated, and deployed every part of the stack. Best option if your organization wants to build an in-house computational toxicology capability.
Acquire EstimaBio in its entirety: the ToxScreen tox product (CPU-based calibrated ADMET-AI) plus the separate longevity research program (a structure-based GPU pipeline, COCONUT R1 virtual screening, LangGraph agents) and lab notebook. Note these are two distinct programs: ToxScreen tox is calibrated ML ADMET; the longevity work is an in-silico structure-based research pipeline. Includes all ongoing R&D and the natural-products drug discovery program.
Why consider this?
ToxScreen is a live, calibration-validated product (real AUROCs, zero mock data, applicability-domain abstention) built with 0 outside investment. The acquisition is a toggle — not a rescue. NaxiAI is not looking for a buyer; we are open to the right conversation with a partner who can take ToxScreen to pharma sales channels or integrate it into an existing platform.
Email us to schedule a confidential discussion.
info@toxscreen.aiNDA execution available upon request. We respond within 24 hours.