Mintomics - Integration of Transcriptomics to Proteomics using Transformer

mintomics is a multi-omics analysis pipeline designed to integrate transcriptomic and proteomic data using a transformer-based model. The objective is to elucidate the adaptive characteristics of the oviduct during natural fertilization, as described in Finnerty et al., eLife 2025.

The oviduct (fallopian tube) is the site of fertilization and preimplantation embryo development in mammals. This project investigates how the presence of gametes and embryos modulates oviductal gene and protein expression, using multi-omics data and advanced machine learning.

• Transcriptomics: Bulk RNA-seq data from mouse oviductal tissues at various preimplantation stages.
• Proteomics: Protein abundance data from oviductal fluid, comparing natural fertilization and superovulation.
• Machine Learning: A transformer-based model integrates transcriptomic and proteomic data to predict protein abundance from gene expression and identify key transcription factors.

For more details, see the eLife article.

Inputs are read from Dataset/.

• Gene expression (CPM): Dataset/Data_cpm/Data_control.csv, Data_0_5preg.csv, Data_1_5preg.csv, Data_2_5preg.csv
• Protein labels: Dataset/Labels_proc_log10_minmax/Labels_control.csv, Labels_0_5preg.csv, Labels_1_5preg.csv, Labels_2_5preg.csv
• Transcription-factor symbols: Dataset/Mouse_TFs1, one symbol per line. A gene in the expression table is marked with the TF input flag when its symbol is in this file.
• Gene-to-protein map: Dataset/genetoprotein.csv
• Differential inputs: Dataset/Collaboration_data.csv, Dataset/Labels_orig.csv, Dataset/Diff_data/, Dataset/Diff_labels/

Training and validation mask 15% of the mapped gene-protein pairs. Held-out inference uses the pairs unmasked.

The pinned environment is Python 3.11 with PyTorch 2.6.0 for CUDA 12.4. From the repository root:

python3.11 -m venv .venv
source .venv/bin/activate
pip install -r requirements.txt

requirements.txt includes the PyTorch CUDA 12.4 index and the packages used by training, inference, and the DESeq2 gene list (lightning, torchmetrics, pandas, pydeseq2, rnanorm, and their dependencies).

.venv/bin/python Training.py \
  --num_gpus 1 \
  --nodes 1 \
  --num_epochs 1 \
  --batch_size 1 \
  --abundance_threshold 0.6 \
  --save_dir exprscale_0.6

The seed is 42. Checkpoints are written under Trainings/<save_dir>/. Use the same --abundance_threshold at inference as was used for that checkpoint. Existing runs are Trainings/exprscale_0.6/ and Trainings/exprscale_0.8/.

Set WANDB_MODE=offline to keep Weights & Biases from contacting the network.

WANDB_MODE=offline .venv/bin/python Inference.py \
  --timepoint all \
  --abundance_threshold 0.6 \
  --save_dir exprscale_0.6 \
  --chkpt mintomics_epoch=00_valid_loss=0.089202.ckpt

For the 0.8 checkpoint:

WANDB_MODE=offline .venv/bin/python Inference.py \
  --timepoint 2.5 \
  --abundance_threshold 0.8 \
  --save_dir exprscale_0.8 \
  --chkpt mintomics_epoch=00_valid_loss=0.011100.ckpt

Inference and the differential-gene script write to intermediate_files/ at the repository root. They do not replace files in Dataset/.

Inference.py writes one set of files per stage. The stage tag is the timepoint with the dot replaced by an underscore (0_5, 1_5, 2_5).

• Tfs_allprot_<tag>.csv: for each mapped protein, the 25 genes with the highest attention. The table has 25 data rows. Gene symbols occupy the first half of the columns and the sigmoid-scaled attention scores occupy the second half, under the same protein headers.
• inference_confusion_matrix_<tag>.png: held-out binary confusion matrix at the chosen abundance threshold.
• inference_attention_<tag>.png: attention heatmap for the high-abundance proteins.

The top-25 ranking uses every gene in the expression table. It is not limited to Dataset/Mouse_TFs1.

From the repository root:

.venv/bin/python src/preprocess/Diff_Gene_proc.py

The script reads Dataset/Collaboration_data.csv, drops genes whose counts sum to less than 10, drops the 3.5 and pseudo samples, and runs DESeq2 for T0.5, T1.5, and T2.5 against the Finnerty control TC. Genes kept have adjusted p < 0.05 and absolute log2 fold change > 0.1, and they must map through Dataset/genetoprotein.csv to an accession in Dataset/Labels_orig.csv.

It writes one symbol per line, without a header:

• intermediate_files/Siggenebasedprotlist_TCT0.5.csv
• intermediate_files/Siggenebasedprotlist_TCT1.5.csv
• intermediate_files/Siggenebasedprotlist_TCT2.5.csv

Those three files are written before the later protein-ranking section. That section imports protrank, which is not on the default module path, so the process can exit after the lists have already been saved.

• Result_analysis.py reads intermediate_files/Tfs_allprot_{0_5,1_5,2_5}.csv and intermediate_files/Siggenebasedprotlist_TCT{0.5,1.5,2.5}.csv. For each stage it keeps the protein columns whose gene symbol is in that significant-gene list, sorts the ranked gene symbols in each column, and writes intermediate_files/AllTfs_diffcodinggene_{0_5,1_5,2_5}.csv. Current inference does not write a control table, so this script does not look for Tfs_allprot_control.csv.
• Training logs go to Weights & Biases. Use WANDB_MODE=offline for a local run.

• Finnerty RM, Carulli DJ, Hedge A, et al. (2025). Multi-omics analyses and machine learning prediction of oviductal responses in the presence of gametes and embryos. eLife 13:RP100705. https://elifesciences.org/articles/100705

This project is distributed under the terms of the Creative Commons Attribution License, as per the referenced publication.