Biomaterials by design: Harnessing data for future development

Abstract

Biomaterials is an interdisciplinary field of research to achieve desired biological responses from new materials, regardless of material type. There have been many exciting innovations in this discipline, but commercialization suffers from a lengthy discovery to product pipeline, with many failures along the way. Success can be greatly accelerated by harnessing machine learning techniques to comb through large amounts of data. There are many potential benefits of moving from an unstructured empirical approach to a development strategy that is entrenched in data. Here, we discuss the recent work on the use of machine learning in the discovery and design of biomaterials, including new polymeric, metallic, ceramics, and nanomaterials, and how machine learning can interface with emerging use cases of 3D printing. We discuss the steps for closer integration of machine learning to make this exciting possibility a reality.

Graphical abstract

Fig. 1

A-D) 8 amines, 8 aldehydes/ketones, 12 Fmoc-amino acids, and 3 isocyanides used for generating the library of hydrogels. E) Screening results of hydrogels (red, a gel formed; gray, solution state. F) Preparation of peptides library. Adapted from the literature [82].

Fig. 2

(A) ANN model was used to evaluate and calculate the water contact angle and fibrinogen adsorption on the surfaces of self-assembled monolayers (SAM). The red and blue arrows illustrate positive and negative weights after the training, respectively. (B1-2) Amounts of adsorbed fibrinogen onto SAMs predicted by the trained ANN plotted from single-lab data against corresponding experimental values. (C1-2) Results of analytical importance following the one-laboratory data set training. Standard deviation (N ​= ​2000) is the standard error bar. (d) Hypothetical SAMs adsorbed predicted quantity of fibrinogen. The amount of methylene units in their alkyl chain has been changed, preserving the same terminal groups, in (c) OH-, (d) CH3-, (e) NH2- and (f) COOH-terminated SAMs. Adapted from the literature [92].

Fig. 3

(A) Model evaluation and selection by estimating the test error for different models. A bootstrapping method was used, indicating that svr. r (radial basis function kernel) ML model has low test error and outperforms the others. (B) The results of the design of the experiment and EGO mediated iterative loops. The predicted hardness values were plotted versus the measured values for the alloys in the training data and experimental data. (C) The hardness of the newly synthesized alloys as a function of loop iteration number. The inset of (b) plots the predicted values as a function of iteration number, showing a similar tendency to the measured values. Adapted from the literature [107].

Fig. 4

The ML models are assessed to match the data supplied. a) The ML-predicted EFA, using random forest fit b) ML forest random fit projected EFA values with 108 chemical characteristics and eight CALPHAD-evaluated features against the known DFT EFA. c) Synthesized material microstructural analysis. Bar scale 100 μm [119].

Fig. 5

De novo approach to study the fracture problem with the crack patterns and crack length, based on a training set derived from molecular dynamics (MD) simulations. It can be used to replace the stochastic problem of fracture propagation in MD simulations through a machine-learning model that can predict the overall crack propagation pathways [120]. This obviates the need for time-consuming MD simulation, opening novel avenues for materials design solutions with atomistic-level degrees of freedom.

Fig. 6

(a) Schematic of high throughput screening to create self-assembling drug nanoparticles. (b) High-throughput testing of all 1440 combinations of 16 drugs and 90 excipients (inactive ingredients, generally recognized-as-safe food and drug additives and other FDA-approved compounds). (c) Molecular properties of drug-excipient pairs are encoded, and molecular simulations (molecular dynamics) determines interaction potentials. These are fed into a random forest machine learning model to indicate important features and predict potential to co-aggregate and form nanoparticles. (d) Out-of-bag performance analysis on different training dataset sizes with Matthews correlation coefficient showing convergence. (e) The machine learning model was used to model 2.1 million pairs of drugs and excipients for ability to co-aggregate and form nanoparticles, and the six named pairs were validated experimentally. The novel component that was not part of the initial high throughput screen in part A are underscored [53].

Fig. 7

(A) Stiff and soft building blocks are combined in patterns to form unit cells U1, U2 and U3. Pink refers to stiff building blocks, while black refers to soft building blocks. (B) The ratio between the modulus of stiff and soft blocks affects unit cell isotropy. (C) The microstructure of stiff and soft blocks is converted into unit cells U1 (blue), U2 (red) and U3 (yellow), which can be further converted into a data matrix with inputs of 1, 2, 3 corresponding to the type of unit cell. (D) Strength ratio is the strength of machine learning generated samples after 1000 loops (green), and after 1,000,000 loops (red), as compared to the highest training set strength. Toughness ratio is the toughness as compared to the highest training set toughness. (E) Stress-strain curves of machine learning generated 3D printed sample (ML-optimal and ML-minimum) [63].