Pilot configurations

Reference engagement patterns with explicit inputs, deliverables, and demonstration status. Pilot scope is configured to each lab's research question.

Pilot view
Reference patterns for academic engagements

Four reference pilot patterns, plus a path for configurations outside the reference set.

The patterns below are starting points demonstrated on this codebase, not a fixed menu. Each pilot is scoped to the lab's research question, data modality, and compute envelope; the engagement phasing and compute floor that follow apply to any configuration. Three patterns are tied to demonstrated runs; one is marked as a roadmap protocol pending cross-site execution.

A

Drug discovery pipeline

Demonstrated

Graph-convolutional bioactivity model with knowledge-graph integration applied to lab-specific compound libraries.

InputsSMILES with bioactivity labels, or curated subsets of ChEMBL or PubChem.
OutputsTrained GCN, validation report (5-fold CV, bootstrap CIs, calibration), versioned pipeline artifacts.
Evidence97.5% accuracy on ChEMBL binary bioactivity; 97.2% ± 0.6% under 5-fold CV.
B

Multi-modal clinical research

Demonstrated

Attention-based fusion across two or more modalities (clinical, omics, imaging, text) sharing patient- or sample-level identifiers.

InputsAt least two modalities with shared sample identifiers.
OutputsPer-modality baselines, attention-fusion model, leakage-free holdout evaluation, conformal prediction sets for uncertainty.
Evidence93.0% fusion accuracy on validation; omics 92.8% with vision PCA and spatial features on Visium breast tissue.
C

Autonomous experimentation

Demonstrated

Experiment-selection loop applied to lab datasets to explore model and configuration space under stopping rules.

InputsOne or more supervised datasets across any modality SNPTX supports.
OutputsCampaign results, intelligence-layer summaries (meta-analysis, adaptive defaults), DuckDB experiment catalog for post-hoc analysis.
Evidence1,100 experiments completed in 1.06 hours; 104 discoveries across 5 datasets and 6 model families.
D

Cross-institutional reproducibility

Roadmap

Proposed protocol for reproducing a reference pipeline across two compute environments, with hash-level artifact comparison.

InputsTwo environments with Python 3.11+ and GPU access.
OutputsReproducibility report comparing artifact hashes and metric drift across sites.
StatusWithin-environment reproducibility controls validated; cross-site execution pending.
E

Custom configuration

Most common path

Most lab engagements do not map cleanly onto a single reference pattern. A custom configuration scopes the pilot to the lab's research question, available modalities, and compute envelope, drawing on the same architecture, engagement phasing, and compute floor as the reference patterns above.

InputsAny supervised, generative, or analytical workflow expressible against SNPTX modality adapters and the execution spine.
OutputsDefined per scoping engagement: trained models, validation reports, comparative analysis, or autonomous campaign results, packaged with run manifests and artifact hashes.
EvidenceInherits platform-level evidence reported in Results and Validation; pilot-specific evidence is produced during the engagement.
ScopingConfiguration is fixed during the 1–2 week scoping phase; the same engagement model and compute floor apply.
Custom pilot scoping flow Lab inputs (research question, data modalities, compute envelope) flow into the SNPTX platform spine (modality adapters, execution spine, optional autonomy loop), producing a custom evidence package (trained model, validation report, run manifests). LAB SIGNALS SNPTX PLATFORM CUSTOM EVIDENCE PACKAGE Research question scope · endpoints Data modalities tabular · omics · imaging · text Compute envelope on-prem · cloud · HPC Modality adapters ingest · align · manifest Execution spine train · evaluate · report Autonomy loop (optional) experiment selection · stopping Mediated extensions contract-validated attach Trained model(s) artifact-versioned Validation report CV · CIs · calibration Run manifests hashes · configs · logs Comparative analysis baselines · ablations SCOPED IN 1–2 WEEKS · SAME ENGAGEMENT MODEL · SAME COMPUTE FLOOR

Operating envelope

Shared engagement phasing and the minimum compute floor that any of the four configurations assumes.

Engagement phases

Phased delivery

Common phasing across pilot types; durations are typical ranges for academic engagements.

PhaseDurationActivities
Scoping1–2 weeksDefine research question, select configuration, confirm compute compatibility.
Deployment2–4 weeksConfigure pipeline for lab data, write modality adapters, validate end-to-end run.
Evaluation2–4 weeksRun campaigns, compare against baselines, generate validation report.
ContinuationOngoingExtend to additional datasets or modalities; archive run manifests for reuse.
Compute floor

Minimum requirements

Baseline environment definition assumed by every pilot configuration; recommended values reflect demonstrated run conditions.

ResourceMinimumRecommended
CPU4 cores8+ cores
RAM16 GB32 GB
GPUNVIDIA, 8 GB VRAMA10G (23 GB) or better
Storage50 GB200 GB
Python3.11+3.11.14
OSUbuntu 22.04+Ubuntu 22.04 LTS

Scope and non-claims

Boundaries that apply uniformly to every pilot configuration. These are stated to keep the engagement model research-facing and pre-deployment.

Research framework

Not a medical device

SNPTX produces analytical artifacts intended for expert interpretation. It is not cleared for diagnostic, prognostic, or clinical decision use.

Model provenance

Trained on lab data

Pilots train models from declared inputs supplied by the lab. No pre-trained models are distributed for clinical deployment.

Data residency

In-place processing

Data remains on lab infrastructure. SNPTX is deployed into the lab's environment and does not host or transfer raw data.

Demonstration discipline

Configurations A, B, and C are tied to runs reproducible from this codebase under the stated compute floor. Configuration D is explicitly a protocol; cross-site results will be reported separately when the trial is executed. Configuration E inherits platform-level evidence and produces pilot-specific evidence during the engagement.