Generate and evaluate
Selected path“Propose drug candidates for my target.”
GenMol → Boltz-2 → Toxicity assessment
Generate molecules, screen target binding, then assess toxicity concern.
◉ Recorded demo · real saved resultsLoading the verified example study…
DISCOVERY, WITH HUMAN CONTEXT
Generate new molecules or evaluate your candidates.
Let the agent plan the discovery workflow.
See how toxicity concern changes the follow-up.
One research question.
Every decision traceable.
RECORDED LOCAL REVIEW
These are the saved, sanitized inputs from the original study.
RECORDED WORKFLOW
AGENT-LED DISCOVERY
Your prompt and candidate inputs → LLM agent → BioNeMo workflow → Toxicity assessment
The agent interprets your research goal and plans the supported workflow. The discovery path depends on whether you need new molecules or want to evaluate candidates you already have.
“Propose drug candidates for my target.”
GenMol → Boltz-2 → Toxicity assessment
Generate molecules, screen target binding, then assess toxicity concern.
“Evaluate these candidates against my target.”
Boltz-2 → Toxicity assessment
Use your candidate structures and target. Screen binding directly, then assess toxicity concern.
This study: liver toxicity (DILI)
FROM QUESTION TO EVIDENCE
Complete a study to explore binding predictions, human liver concern and the next experiment in one dashboard.
YOUR STUDY, IN CONTEXT
Shared across Overview, Candidate explorer and Discovery lab. Display order only; scientific ranks stay unchanged.
FROM DISCOVERY TO FOLLOW-UP
CANDIDATE EXPLORER
Select a molecule to inspect model context, mapped fingerprint features and recommended experiments.
CANDIDATE EVIDENCE
Binding likelihood and DILI concern are independent model outputs. Structural confidence is a separate Boltz-2 output. Select a candidate to inspect the evidence.
| Candidate | Discovery rank | Binder likelihood | Predicted affinity | Structural confidence | Human DILI | Follow-up |
|---|
No candidates match this view.
Training similarity describes applicability, not prediction confidence. A lower DILI score does not establish safety. No missing score is treated as zero.
TWO PATHS, ONE WORKSPACE
Each example is one complete recording. Scores, structures and agent notes stay with their original study.
METHODS & SETUP
A static, interactive replay of two real ABL1 runs from 20 September 2026. No model is running in this website. The original assistant was NVIDIA Nemotron, not Rosalind; a callable Rosalind API was unavailable.
The product vision extends toxicity assessment across organs, including the liver, heart and kidneys. The current demo illustrates liver toxicity.
In the current ABL1 implementation, supplying candidate structures selects screening; without candidates, the workflow prepares generation. After privacy approval, Nemotron interprets the question and validates the supported plan. Broader prompt-driven model routing is part of the product direction.
The target-only walkthrough presents a curated cohort of 18 successfully screened candidates from the original 20-candidate run. Both shortlisted candidates need liver validation. Original review inputs and agent notes describe the full source run. Original 20-candidate report · Original recording and generation ledger. Generated molecules are not a labelled evaluation cohort. The four supplied drugs overlap model fitting or selection and are not unseen evaluation cases.
Predicted binding is not measured efficacy. DILI scores are not dose-specific outcomes or an assurance of safety. Binding and DILI scores cannot be subtracted. Model evaluation belongs to its held-out reference set, not these demo candidates.
New prompt text and file contents never leave this page or enter browser storage. Only the selected example, reached step and playback position are remembered. Refresh restores paused playback. Reset clears that progress. There are no analytics or accounts.
The existing local app handles new inputs, real privacy filtering, approval and live provider calls. It accepts free-text targets, but execution currently supports the prepared ABL1 domain; other targets need preparation. Connecting a local privacy companion to hosted execution remains future work.
./scripts/toxoracle-web
Source downloads retain scientific evidence and model provenance. Raw provider responses are not published; content URNs preserve their hashes. All displayed poses use the recorded candidate’s verified coordinates.