Arcis
An auditable design loop for protein binders.
Arcis coordinates sequence generation, fold validation, structure-conditioned variants, complex prediction, and ensemble scoring. Every candidate and decision remains traceable.
Computational evidence only. Independent wet-lab validation is the next major milestone.
End-to-end computational loop
Target inputs move through generation, validation, variants, complex prediction, scoring, and refinement.
Known-answer benchmark protocol
Barnase and barstar controls exercise the workflow, scorecard, candidate lineage, and stop decisions.
Experimental binding
The current evidence is computational. Independent wet-lab affinity data is the next major proof point.
The working system
From research intent to a ranked candidate.
The graph starts with explicit biological inputs, moves expensive model work onto asynchronous workers, and uses a structured scorecard to decide whether to stop or refine.
The loop is the product.
Individual models can change. The durable layer is the state, evidence, recovery behavior, and decision history around them.
- 01 / Research input
Define the design problem
Provide a target sequence or structure, chain definitions, epitope constraints, binder family, length, and liabilities.
Structured run config
- 02 / Architect
Propose a candidate
Gemini turns the research objective and constraints into an initial binder sequence or a conservative refinement.
Gemini
- 03 / Validator
Check fold confidence
The candidate is folded and screened before expensive complex prediction. Low-confidence sequences return for refinement.
ESMFold
- 04 / Decoder
Generate structural variants
Passing structures seed structure-conditioned sequence variants that are validated concurrently and filtered by constraints.
ProteinMPNN
- 05 / Builder and scorecard
Predict the complex
The selected binder and target are folded together, then assessed for affinity, contacts, buried surface, clashes, and confidence.
Chai-1 + PRODIGY
- 06 / Decision
Stop or refine
The loop records why a candidate passed or failed, preserves its lineage, and continues until the target or iteration limit is reached.
LangGraph state machine
Model-flexible
Architect, validator, decoder, builder, and critic are adapters rather than a single locked model stack.
Stateful and reviewable
Every candidate, score, failure, cache hit, lineage event, and stop reason remains attached to the run.
Built for asynchronous compute
Firestore and correlated Pub/Sub jobs let the graph suspend around long GPU work and resume safely.
Portable evidence
Runs export candidate tables, structures, scorecards, and machine-readable research history.
Current evidence
A working loop with controls and a paper trail.
The current POC includes a reproducible barnase and barstar protocol, separate control lanes, candidate rankings, structures, and structured failure and stop reasons.
Native barstar, a known barnase binder, is folded and scored.
A deterministic shuffled barstar sequence tests rejection behavior.
The known binder sequence and scaffold are withheld while the loop generates and ranks candidates.
What this proves today
The benchmark demonstrates workflow behavior and evidence capture. It does not establish a general hit rate or experimental binding efficacy.
Scorecard v1
A ranking heuristic assembled from available structural and sequence signals. It is not a substitute for experimental Kd.
- PRODIGY affinity
- Available
- Validator pLDDT
- Available
- Contacts, buried surface, clashes
- Available
- Sequence liabilities and diversity
- Available
- Native Chai iPTM and PAE
- Next
- Experimental Kd
- Not measured
RUN STATE
Persistent and exportable
GPU JOBS
Correlated and retryable
CANDIDATES
Ranked with lineage
EXPORTS
JSON, CSV, and structures
Validation roadmap
The next claim must be earned.
Arcis is ready to demonstrate orchestration, traceability, and computational iteration. The next work is to prove improvement across targets, then test prioritized candidates in the lab.
01Computational POC
A deployed, auditable loop with parallel ProteinMPNN variants, structured scorecards, controls, exports, and tested async recovery.
02Blind multi-target evaluation
Compare the closed loop against an equal-compute open-loop baseline across multiple targets, decoys, and random seeds.
03Independent experimental validation
Express a prioritized candidate set and measure binding, specificity, and developability with appropriate controls.
The system is inspectable now
Follow a candidate from prompt to scorecard.
The demo exposes active iterations, benchmark lanes, candidate rankings, score summaries, failure events, and stop decisions.