D2Deep

Cancer driver predictors have a habit of grading the gene rather than the mutation. Because annotated driver mutations pile up in a handful of intensively studied genes, a model trained on them can score well by learning that anything in TP53 or BRAF is probably a driver — which collapses precisely where clinicians need help, on non-hotspot positions. D2Deep, from the group of Wim Vranken at the Vrije Universiteit Brussel and the Interuniversity Institute of Bioinformatics in Brussels (IB2, a joint venture with the Université libre de Bruxelles), attacks that bias with a deliberately gene-agnostic feature set built from sequence alone.

The mechanism combines two signals that fail in different places. A general protein language model — here ProtT5-XL — learns the grammar shared by all protein sequences but knows nothing about what a particular family tolerates; a multiple sequence alignment carries exactly that family-specific constraint but says nothing about biochemical plausibility in general. D2Deep joins them by embedding every sequence in a query protein's alignment with ProtT5-XL and fitting a Gaussian mixture model to each alignment column. Wild-type and mutant residues are then scored by how far their log-likelihoods sit from that column's fitted distribution, and the per-position differences — concatenated across the whole sequence — become the features a small supervised classifier uses to call pathogenicity.

Two things fall out of that construction. Because a mutation shifts the language model's representation of the entire sequence, the feature vector registers effects far from the substituted residue, giving a sequence-only readout of epistasis; and because the mixture model is a statistical fit, its log-likelihood doubles as a confidence estimate for each prediction. The method was posted as a preprint in November 2023 and published in Briefings in Bioinformatics in November 2024; the same lab later generalized its feature core into the zero-shot, multi-task D2D framework.

#Key Features

• Sequence-only input: No experimental or predicted structure is required, so isoforms of the same gene get distinct predictions and disordered regions remain in scope.
• Per-prediction confidence: A grid-searched formula over the mixture model's average log-probability yields a 0–100% confidence, and restricting to confident calls lifts accuracy on every gene tested.
• Epistasis readout: The per-position feature profile highlights residues distant in sequence but close in the AlphaFold2 structure, with peaks that align with allosteric effects other tools derive from free-energy calculations.
• Gene-balanced training: Pathogenic and benign variants are balanced within each gene, so the classifier cannot fall back on gene identity as a shortcut.

#Technical Details

Alignments are built with MMseqs2 against UniRef100 and PDB70, and each aligned residue is embedded by ProtT5-XL into 1024 dimensions, max-pooled to 20. A scikit-learn Gaussian mixture is fitted per alignment column; the feature at each position is the difference between the wild-type and mutant distances to a 1% log-likelihood threshold. The downstream classifier is a two-layer perceptron — 2,200 inputs, 4,096 and 2,048 hidden units with batch normalization, ReLU and 0.3 dropout, roughly 17 million parameters — trained for 200 epochs with AdamW at a learning rate of 3e-5. Curation from COSMIC, the Cancer Genome Interpreter, ClinVar, humsavar, gnomAD and dbSNP yielded a balanced training set of 6,608 mutations (2,956 deleterious, 3,652 benign) across 1,012 genes. On the independent DRGN benchmark of 4,093 variants across 153 genes, D2Deep outperforms SIFT, PolyPhen-2, CADD, DEOGEN2, EVE and AlphaMissense on AUC, trailing only FATHMM-cancer — which itself degrades sharply on non-hotspot mutations, where D2Deep does not. Ablations against ProtT5 and ESM-1v features alone confirm that the evolutionary component, not the language model, supplies the gain. The ProtT5-XL dependency imposes a 2,200-residue ceiling on input sequences.

#Applications

D2Deep targets the interpretation step of clinical tumor sequencing: ranking somatic variants of uncertain significance, checking whether a novel substitution behaves like established drivers in the same protein, and flagging which calls are trustworthy enough to act on. Precomputed predictions and confidence scores for the 269 genes of the Compermed clinical NGS panel are browsable per gene and per position on the group's web server, and the released code scores arbitrary UniProt accessions on a 40 GB GPU. Its structure-free design also suits deep mutational scanning analyses and germline variant work outside oncology.

#Impact

D2Deep is one of the first large-scale tests of protein language models on cancer driver prediction, and its result is a useful negative as much as a positive: language-model embeddings alone underperform, and it is the protein-specific evolutionary statistics layered on top that make the approach competitive with predictors trained on pathway, domain and structural annotations. The confidence score is the more portable contribution, since most variant effect predictors return a bare number with no indication of when to distrust it. The model is supervised and single-task and its trained classifier weights are not distributed alongside the code, but the evolutionary-constraint features it introduced went on to carry the lab's broader D2D framework.