SL-Predict: Frozen MAE Gene Encoder

Pretrained masked-autoencoder (MAE) gene encoder for cold-start synthetic lethality prediction from DepMap CRISPR screens.

Model Description

A 3-layer MLP encoder (1206 → 512 → 256 → 256) trained to reconstruct randomly masked DepMap Chronos dependency profiles (18,531 genes × 1,206 non-K562 cell lines) with MSE loss for 200 epochs.

Key property: This is the leak-repaired checkpoint — the 503-gene union of all downstream cold-start test sets was excluded from pretraining. TOST equivalence testing confirms the encoder is not load-bearing on pretrain–test gene overlap (p_max < 0.0001 at ±0.010 AUC).

Performance

When frozen and combined with LightGBM + confidence weighting on SynLethDB CRISPR/CRISPRi labels:

Metric	Value
Horlbeck K562 held-out AUC	0.714 ± 0.018 (10-seed, gene-disjoint)
vs Published SOTA (SLMGAE)	+0.079
vs Label-agreement ceiling	+0.015

Usage

import torch

# Load checkpoint
ckpt = torch.load("mae_encoder_d256_leak_repaired.ckpt", map_location="cpu")
state_dict = ckpt["state_dict"]

# The encoder is the first 3 layers of the MAE
# Input: 1206-dim DepMap dependency profile (z-scored)
# Output: 256-dim gene embedding

Training Details

Data: DepMap 26Q1 Chronos dependency profiles
Architecture: MLP 1206→512→256→256 (encoder), mirror decoder
Objective: Masked autoencoding (50% masking ratio, MSE loss)
Epochs: 200
Hardware: Single NVIDIA A10G (Modal cloud), ~20 minutes
Leak repair: 503 test-split genes excluded from pretraining data

Citation

@misc{large2026slpredict,
  author = {Large, Jack},
  title = {Cold-start synthetic lethality prediction: Diagnosing evaluation inflation and a constructive baseline},
  year = {2026},
  url = {https://github.com/j8ckfi/sl-predict}
}