Journal: bioRxiv

Article Title: Meta Learning Improves Robustness and Performance in Machine Learning-Guided Protein Engineering

doi: 10.1101/2023.01.30.526201

Figure Lengend Snippet: Design and screening of yeast display antibody (scFv) libraries. (A) Workflow overview: Degenerate oligonucleotides are computationally designed based on previously published deep mutational scanning data (DMS) ( Mason et al ., 2021 ; Koenig et al ., 2015 ) and crystal structures (PDB: 1N8Z, 4ZFF, 4ZFG), transformed into yeast, and screened by FACS for binding to antigen. The sorted populations are then used for targeted deep sequencing of the antibody variable regions. (B) The crystal structure of 4D5 scFv (grey) in complex with HER2 (blue) (PDB: 1N8Z). The amino acid positions of 4D5 targeted for combinatorial mutagenesis are highlighted in pink. (C) The 4D5 scFv library is screened for binding to HER2 antigen by FACS; dot plots show the library is fractionated into HER2-binding populations (High-, Low-, or Non-binding) based on fluorescence intensity. Multiple rounds of FACS are performed (Fig. S1). (D) Protein sequence logo plots represent the mutagenesis regions of the 4D5 scFV and are derived from deep sequencing of the various HER2-binding populations. (E) The crystal structure of 5A12 scFv (grey) in complex with VEGF antigen (orange) (PDB: 4ZFF). The amino acid positions of 5A12 targeted for combinatorial mutagenesis are highlighted in pink (F) The 5A12 scFv library is screened for binding to VEGF antigen by FACS; dot plots show the library is fractionated into VEGF-binding populations (High- or Low/Non-binding) based on fluorescence intensity. Multiple rounds of FACS are performed (Fig. S2). (G) Protein sequence logo plots represent the mutagenesis regions of the 5A12 scFV and are derived from deep sequencing of the various VEGF-binding populations. (H) The crystal structure of 5A12 scFv (grey) in complex with Ang2 antigen (orange) (PDB: 4ZFG). The amino acid positions of 5A12 targeted for combinatorial mutagenesis are highlighted in pink (I) The 5A12 scFv library is screened for binding to VEGF antigen by FACS; dot plots show the library is fractionated into Ang2-binding populations (High- or Low/Non-binding) based on fluorescence intensity. Multiple rounds of FACS are performed (Fig. S3). (J) Protein sequence logo plots represent the mutagenesis regions of the 5A12 scFV and are derived from deep sequencing of the various Ang2-binding populations.

Article Snippet: The 5A12 library was first incubated with either 0.5 nM VEGF (Acro, VE5-H82Q0-200ug) or 12 nM Ang2 (Sino Biological, 10691-H08S-B) plus Streptavidin-AlexaFluor647 (Bioloegend, 405237) for 30 minutes at 4 C. Following the fist stain, cells were centrifuged at 8000 G for 30 seconds and washed once before incubation with 1ng/μL anti-FLAG-PE for 30 minutes at 4 C. Following staining, cells were washed twice and kept on ice and away from light until sorting. scFv expressing (FLAG+) cells in the 4D5 library were sorted by FACS (BD Aria Fusion) first into HER2 binding and non-binding fractions, then into three populations based on AlexaFluor647-conjugated HER2 Mean Fluorescent Intensity: High, Low, and Negative. scFv expressing (FLAG+) cells in the 5A12 library were sorted by FACS (BD Aria Fusion) into High binding (VEGF+ or Ang2+) or Low/Non-binding (VEGF- or Ang2-) populations.

Techniques: Transformation Assay, Binding Assay, Sequencing, Mutagenesis, Fluorescence, Derivative Assay

Journal: bioRxiv

Article Title: Meta Learning Improves Robustness and Performance in Machine Learning-Guided Protein Engineering

doi: 10.1101/2023.01.30.526201

Figure Lengend Snippet: Meta learning applied to supervised machine learning models trained to predict multi-antigen binding classification using yeast antibody (scFv) sequence data with largely single-antigen binding classification labels. (A) Schematic representation of machine learning task. The training set is constructed from deep sequencing of 5A12 libraries following the final round of the VEGF FACS screen (Fig. S2). Positive (High-binding) and negative (Low/Non-binding) VEGF labels are retained and sequences are arbitrarily assigned an Ang2 binding classification. Test and meta sets consist of 5A12 variants with binding labels for both targets and are batched by combining deep sequencing from both VEGF and Ang2 final round FACS screens (Fig. S2, S3). (B,C, and D) Meta learning and baseline prediction performance (Matthew’s Correlation Coefficient) as a function of the number of training samples. FT refers to Fine-Tune Baseline. Points correspond to mean performance and shaded regions to 95 % confidence intervals across 3 random seeds. Performance curves plotted for meta sets consisting of 32 (bright yellow), 96 (dark yellow), 288 (red), and 864 (purple) sequences.

Article Snippet: The 5A12 library was first incubated with either 0.5 nM VEGF (Acro, VE5-H82Q0-200ug) or 12 nM Ang2 (Sino Biological, 10691-H08S-B) plus Streptavidin-AlexaFluor647 (Bioloegend, 405237) for 30 minutes at 4 C. Following the fist stain, cells were centrifuged at 8000 G for 30 seconds and washed once before incubation with 1ng/μL anti-FLAG-PE for 30 minutes at 4 C. Following staining, cells were washed twice and kept on ice and away from light until sorting. scFv expressing (FLAG+) cells in the 4D5 library were sorted by FACS (BD Aria Fusion) first into HER2 binding and non-binding fractions, then into three populations based on AlexaFluor647-conjugated HER2 Mean Fluorescent Intensity: High, Low, and Negative. scFv expressing (FLAG+) cells in the 5A12 library were sorted by FACS (BD Aria Fusion) into High binding (VEGF+ or Ang2+) or Low/Non-binding (VEGF- or Ang2-) populations.

Techniques: Binding Assay, Sequencing, Construct