Turn a target protein into ranked, novel drug candidates — then visualize them anchored to the target.
What it is
Generate Novel Assets is a design run: you point the platform at a target protein, and a multi-engine design pipeline (ESM / Co-scientist) proposes a ranked list of novel candidate drugs against it. Instead of starting from a known compound, you start from the biology — a gene/protein in your hypothesis space — and let the engine design molecules to bind it.
The flow has three moving parts:
A target — a gene/protein voxel that carries a protein sequence (see the prerequisite below).
A design run — you choose how many candidates to return (top-K) plus optional constraints, then the engine runs through its declared stages while you watch a live stage tracker.
Top-K candidates — on completion, a ranked list of designed molecules with scores (binding, novelty, developability), which you can then add to your graph as drug voxels anchored to the original target.
AI output
Designed candidates are model proposals, not validated assets. Treat them as a starting shortlist to triage — the natural next step is to carry promising candidates into a BVCT to predict effect size. Prospective track record: data.bioinvestgpt.com.
Where it lives: Design runs start from the Hypothesis Generator's 3D knowledge graph. If you are not there yet, open it at bvct.bioinvestgpt.com/#/generator and load or build a hypothesis space first.
Prerequisite: the target must carry a protein sequence
The design pipeline needs an amino-acid sequence to design against. A target voxel qualifies when it carries either a protein_aaseq (a PrimeKG / UniProt protein sequence) or an esm_aaseq (an ESM Atlas sequence, inlined for MGYP-style accessions). If neither is present, there is nothing to design against — and the entry points behave accordingly.
Entry point
When the action appears
Details panel CTA
Only for a gene/protein node that already has protein_aaseq or esm_aaseq loaded. No sequence on file means no button.
Right-click a gene/protein voxel
Shown for any gene/protein voxel. PrimeKG sequences are lazy-loaded, so the menu item appears optimistically and the sequence is resolved when you click. If none can be resolved, the run reports that the target has no protein sequence on file.
Right-click a disease voxel
Opens a target picker that lists the disease's connected gene/protein voxels; only sequence-bearing targets are useful design starts.
If the Details panel shows no "Generate Novel Assets" button, the node has no sequence yet. Open the node's Details and confirm a sequence is listed, or use Expand (1 hop) on a nearby disease to pull in connected proteins from PrimeKG, then try again.
ESM Atlas targets are badged ESM Atlas in the picker and carry their sequence inline (via esm_aaseq), so they design directly without a separate sequence fetch.
Run a design
There are three ways to start a design run, all of which open the same Generate Novel Assets modal:
From a gene/protein voxel — right-click the voxel and choose Generate Novel Assets from the annotation menu.
From the Details panel — open a gene/protein node's details and click the amber Generate Novel Assets button (shown only when a sequence is on file).
From a disease — right-click a disease voxel and choose Generate Novel Assets for this disease. A target picker lists the connected druggable proteins ranked by relevance (disease drivers first, then known drug targets); pick one to proceed.
1
Open Generate Novel Assets
The modal opens with a target summary at the top — the target protein name and its sequence length in amino acids (aa). Confirm this is the protein you intend to design against.
You will need to be signed in: design runs are tracked on your account for cost and audit reasons. If you opened the modal anonymously (for example via a showcase deep link) or your session expired, the modal shows a Sign in prompt instead of the configuration form.
Tip: The header reads "Design candidate drugs against ( aa)" — the aa count is a quick sanity check that the right sequence was resolved.
2
Choose how many candidates (top-K)
Set Number of candidates (top-K) — how many designed molecules the run should return. The default is 5; you can request between 1 and 20. Higher top-K returns a broader shortlist but typically costs more and takes longer.
3
Pick an engine — or leave it on Auto
By default the Engine is set to Auto — the platform picks the best available design engine for this run, so you can skip straight to generating. Expand Advanced if you want to choose a specific engine; the dropdown lists each engine with its vendor, a deployment-type pill (Public API vs. Internal VPC), and an estimated cost range.
Tip: Stay on Auto unless you have a reason to pin a specific engine. Auto-picked runs are labelled "(auto-picked)" in the results; manually chosen ones read "(manual)".
4
Add optional constraints
Expand Constraints (optional) to steer the design. The current control is Modality — choose peptide, small molecule, or antibody. Leave the accordion collapsed to use the engine's defaults.
5
Acknowledge any safety checks
Depending on the engine you may see one or two acknowledgements before you can generate:
Sequence of concern — required by some engines if the target may be flagged for safety review. The acknowledgement is logged in the audit trail.
No PHI in this sequence — shown in Advanced mode only when you manually pick a public-API engine whose vendor is not covered by a BAA. It confirms the sequence is patient-free so it can be dispatched off-platform.
Tip: In Auto mode you will rarely see these — the platform filters non-compliant engines for you. They surface mainly when you pin a specific engine under Advanced.
6
Generate and watch the stage tracker
Click Generate Top-K Candidates. The modal switches to a live stage tracker that shows each engine-declared stage as it runs: pending stages are greyed, the running stage spins amber, completed stages turn emerald, and failures turn red. The header shows the engine name and version stamp.
If a run errors or stops on budget, a Retry button resumes the same inputs from the last completed stage — you do not lose finished work.
Tip: Design runs are server-side and can take several minutes. You can leave the modal open and watch progress; the tracker updates automatically.
Cost: design runs consume non-LLM compute budget. The selected engine's estimated cost range and your remaining daily budget are shown in Advanced mode, and the Generate button is disabled if a run would exceed your cap.
Review and visualize the Top-K candidates
When the run completes, the modal lists the ranked candidates. Each row shows:
Rank (#1, #2, …) — best candidate first.
The designed molecule — its amino-acid sequence or SMILES string.
Scores — binding affinity, novelty, and (for structure-predicting engines) a mean pLDDT confidence.
A developability chip — colour-coded green above 0.7, amber 0.4–0.7, and red below 0.4.
Predicted structure link — where available, an external link that opens the candidate's predicted 3D structure (minted on demand as a fresh signed URL).
Score
What it tells you
Binding
Predicted affinity of the designed molecule for the target. Higher is better.
Novelty
How distinct the candidate is from known chemistry/sequences — higher means more novel.
pLDDT
Mean confidence of the predicted structure (0–100). Only shown for engines that fold the candidate.
Developability
A rolled-up manufacturability/stability signal, shown as a green / amber / red chip.
7
Add the candidates to your graph
Click Add top-K to /generator/drugs. Each designed candidate becomes a new drug voxel (purple) in the 3D graph, anchored to the original target protein — the candidates appear as children of the protein with drug → target edges, so the design provenance stays visible. The view lands on the Drugs tab so you can inspect them immediately.
Tip: If you reset the scene between starting the run and adding candidates, the target protein is re-injected as a placeholder so the graph stays coherent and the anchoring still holds.
What to do next: The designed candidates are a shortlist, not a verdict. To estimate clinical value, carry a promising candidate into a BVCT — open the Validator at bvct.bioinvestgpt.com/#/validator and run a Drugs-stage decision to predict its effect size against a standard of care.
FAQ
The "Generate Novel Assets" button isn't showing.
The node has no protein sequence loaded. Open its Details to confirm a sequence is present, or Expand a connected disease (1 hop) to pull in sequence-bearing proteins from PrimeKG, then retry. The button is gated specifically on protein_aaseq or esm_aaseq.
I right-clicked a protein and saw the menu item, but the run says "no protein sequence on file."
The right-click menu appears optimistically for every gene/protein because PrimeKG sequences are lazy-loaded. If no sequence can be resolved on click, the run reports it — pick a different, annotated target or expand to load sequences.
Which engine should I use?
Leave it on Auto — the platform selects a suitable, compliant engine and shows you which one it picked in the results. Only switch to Advanced if you need a specific engine, modality behaviour, or deployment type.
Are the candidates real, validated drugs?
No. They are AI-designed proposals scored for binding, novelty, and developability. Use them to triage, then validate the most promising ones with a BVCT.
BioinvestGPT BVCT Platform User Guide — Designing Novel Assets (ESM / Co-scientist). BVCT outputs are model-based decision-support analyses, not investment advice. Prospective track record: data.bioinvestgpt.com.
