ThatGuyBassam/AtlasGen-SLM

A variant pathogenicity classifier, trained on North African-representative genomic data, designed to reduce the VUS rate in Moroccan clinical genomics. # AtlasGen-SLM > A ~15M parameter genomic language model for variant pathogenicity classification in North African and Moroccan populations. --- ## The Problem Genomic AI tools are trained almost exclusively on Western European data. When they encounter a variant they've never seen — which happens constantly with North African genomes — they return "Variant of Uncertain Significance." Not because the variant is actually uncertain. Because the training data never included it. That lands as a VUS — a result that's technically inconclusive — for Moroccan and Maghrebi patients at a much higher rate than for European ones. Resolving a VUS means more testing, more time, more cost. In Morocco's healthcare system, those aren't abstract problems. --- ## What AtlasGen-SLM Does It's a variant pathogenicity classifier. Feed it a genomic variant, get back a probability: benign, pathogenic, or uncertain. The difference from existing tools is the training data — built around African, Maghrebi, and Moroccan genomes rather than European ones. It runs offline. The trained weights sit on disk. No API, no cloud, no per-query cost. A clinic with no stable internet can run it. --- ## How It Works ``` Input: variant (chr, pos, ref, alt) ↓ Extract 1024bp window from GRCh38 around the variant position Inject the alternate allele at the variant site ↓ Tokenize into overlapping 6-mers (vocab: 4,096 tokens) ↓ Pass through encoder-only transformer (~15M parameters) ↓ Output: Benign / Pathogenic / Uncertain + confidence score ``` Training runs in four sequential phases — global pretraining, African and Maghrebi fine-tuning, Moroccan specialization, then supervised pathogenicity classification. Details in `docs/architecture.md`. --- ## Model Specs | | | |---|---| | Architecture | Encoder-only transformer | | Parameters | ~15M | | Tokenization | 6-mer sliding window, vocab 4,096 | | Sequence input | 1024bp VCF-guided flanks | | Training hardware | 8GB VRAM | | Attention | FlashAtte …