Deep2Full: Evaluating strategies for selecting the minimal mutational experiments for optimal computational predictions of deep mutational scan outcomes

Abstract

Performing a complete deep mutational scan with all single point mutations may not be practical, and may not even be required, especially if predictive computational models can be developed. Computational models are however naive to cellular response in the myriads of assay-conditions. In a realistic paradigm of assay context-aware predictive hybrid models that combine minimal experimental data from deep mutational scans with structure, sequence information and computational models, we define and evaluate different strategies for choosing this minimal set. We evaluated the trivial strategy of a systematic reduction in the number of mutational studies from 85% to 15%, along with several others about the choice of the types of mutations such as random versus site-directed with the same 15% data completeness. Interestingly, the predictive capabilities by training on a random set of mutations and using a systematic substitution of all amino acids to alanine, asparagine and histidine (ANH) were comparable. Another strategy we explored, augmenting the training data with measurements of the same mutants at multiple assay conditions, did not improve the prediction quality. For the six proteins we analyzed, the bin-wise error in prediction is optimal when 50-100 mutations per bin are used in training the computational model, suggesting that good prediction quality may be achieved with a library of 500-1000 mutations.

The “dark matter” of protein variants carries a distinct DNA signature and predicts damaging variant effects

“…In contrast to previous machine learning approaches utilising DMS data (e.g. [36, 37, 38]), here we introduce DNA mutational signatures as a feature to be evaluated for its predictive utility, and explicitly test the contribution of different feature combinations in classifying variant impact (Figure 5A). We note that only a small number ( n = 29 proteins from 83 experiments) of experimental DMS datasets are publicly available (as of August 2021), with even fewer proteins (9 proteins from 10 experiments) displaying a bimodal distribution of DMS scores amenable for a binary clasification task.…”

HARP: a database of structural impacts of systematic missense mutations in drug targets of Mycobacterium leprae

“…Computational saturation mutagenesis guides experimental approaches to study the impacts or help rationalise the consequences of known or emerging mutations [59] . Such approaches have been applied to other proteins like artificial (βα) 8 ‐barrel protein [19] or in deep mutation scans [60] . Experimental validation of mutation impacts in M. leprae are time and labour-intensive processes owing to the inability of bacillus to grow on an artificial culture media.…”

Facilitating Machine Learning‐Guided Protein Engineering with Smart Library Design and Massively Parallel Assays

“…[37,38] Instead of complex metrics derived from protein structure prediction software, more intuitively explanatory descriptors such as the number of contacts to other amino acid residues are excellent predictors. [39][40][41] Evolutionary descriptors including conservation score and PSSM weight are proven to be also extremely useful in predicting mutational effect [40,42,43] and site-wise tolerance to mutations. The use of multiple sequence alignments (MSA) to generate useful embeddings from transformer language models is under active research.…”

“…Thus far, massively parallel assays are almost built-in with ML predictions on phenotypic values. [11,15,39,49,56,[85][86][87][88] While methods. The computational selected variants harboring correct mutational combinations (variants in solid lines) constitute a portion of the entire library; combinatorial variants that were not selected for experimental validation are also present (variants in dash lines), since it is difficult to control precisely which oligos are cloned into each fragment.…”

“…Besides, ML prediction quality improves when positive selections with lower stress were conducted, as a broader range of functional scores were collected and more variants were surveyed. [39] Together, a reporter system independent of cell survival appears to be more beneficial for MLguided screens.…”