operantclaude/gedi-soothsayer-models

GEDI-CNN & Soothsayer Models

Model checkpoints for automated neuron death detection and prediction from fluorescence microscopy.

Models

Core Pipeline

File	Description	Architecture	Metric
gedi_cnn_v7/best.pt	Live/dead classifier (GEDI channel)	ConvNeXt-Tiny (3ch→2cls)	96.8% val acc
siamese_msn_only/best.pt	Siamese hard-negative embedding model	ConvNeXt-Tiny (Siamese)	—
soothsayer_v2/embed_lstm_k5.pt	Death prediction (k=5 frames)	DualEncoderLSTM	AUC 0.79
soothsayer_v2/embed_lstm_k3.pt	Death prediction (k=3 frames)	DualEncoderLSTM	AUC 0.78
soothsayer_v2/embed_lstm_full.pt	Death prediction (full sequence)	DualEncoderLSTM	AUC 0.90

Model Grid (Variable-k LeJEPA)

Best models from 112-model grid search over encoder, projection head, and pretraining.

File	Encoder	DualProj	Pretrained	AUC (k=5)
soothsayer_grid/lejepa_variablek_dualproj_dinov3_pretrained.pt	DINOv3-reg	Yes	Yes	0.858
soothsayer_grid/lejepa_variablek_dualproj_pretrained.pt	DINOv2	Yes	Yes	0.845
soothsayer_grid/lejepa_variablek_dualproj_dinov3.pt	DINOv3-reg	Yes	No	0.842
soothsayer_grid/lejepa_variablek_dualproj.pt	DINOv2	Yes	No	0.832
soothsayer_grid/lejepa_variablek_pretrained.pt	DINOv2	No	Yes	0.815
soothsayer_grid/lejepa_variablek_gedicnn.pt	GEDI-CNN only	No	No	0.788

Pipeline

1. GEDI-CNN classifies neurons as alive/dead from a single fluorescence image (GFP + RFP + mask).
2. Siamese model produces hard-negative-aware embeddings (768-d) for temporal modeling.
3. Soothsayer (Variable-k LeJEPA with DualProjection) predicts future death from sequences of alive-frame embeddings.

Input Format

• GEDI-CNN: 3-channel 224×224 tensor (GFP, RFP, ones_mask), percentile-clipped and z-scored
• Soothsayer: Sequences of [GEDI-CNN embedding (768-d) + DINOv2/DINOv3 embedding (768-d) + scalars (4-d)]
  • Scalars: normalized GFP, CNN death probability, normalized timepoint, GFP delta

Inference

See soothsayer_inference.py for the complete inference API, designed for closed-loop integration with the Thinking Microscope.

Training Data

Trained on motor striatal neuron (MsN) longitudinal tracking data with GEDI (genetically encoded death indicator) ground truth labels. 12 experiments, ~168k tracks.

Code

Source code: github.com/operantclaude-hash/gedi-cnn-foundation