The number of degrees of freedom of molecular systems is exponentially proportional to its number of atoms. This means only small molecules can be computationally investigated with good accuracy within a reasonable timeframe and, hence, can be designed with little resources because the conformational space and, thus the potential energy landscape macromolecules is intractable for accurate in-silico design.
By thinking outside the box, we, at Oscerva MolEmerge, have devised an algorithm (no A.I., pure mathematics with a dash of heuristics) that makes the computational resource requirement virtually independent of the number of degrees of freedom. The result is that handling even the largest macromolecules is no challenge for current-day computers.
In order to build and edit molecules, computers benefit from useful representations. The more information these molecular representations can contain for an in-silico method, the easier, and closer, this in-silico method is able to fit reality.
although Machine Learning & A.I. systems are, indeed, exceptional, even they are no exception to this rule. Coincidentally, A.I. systems also benefit the greatest per extra information retained in the molecular representation.
Many have tried, in vain, to alleviate this problem. A popular (and more modern) approach is to employ a second A.I. system to infer a molecular representation, yet new problems arise when doing so:
a) Training requires time and resources and inherits all the problems that comes with dealing with any ML systems.
b) there's too much variation between training sessions and so the representations are potentially not deterministic and intrinsically not universal due to training biases.
c) neural networks remain black-boxes, so understanding the inferred representations is extremely difficult if not impossible.
We, at Oscerva MolEmerge, have invented a universal, deterministic and lossless in-house molecular representation which, itself, does not derive from Machine learning models but instead (again) from pure mathematics and is translation-and rotation-invariant. It's conceived specifically to enable Neural Networks to truly *see* molecules the same way humans do, i.e. with no information loss. This inevitably solves the many state-of-the-art problems specified and brings Machine Learning models ever closer to full control of molecular systems.
Due to above-mentioned obstacles, the world is yet to see a fully functional, artificial protein, designed completely from scratch, exhibiting a-priori assigned properties. Instead, even A.I. is currently only used for mixing and matching variations of domains with already known properties.
If we've created tools to overcome the specified problems, what's left to keep us from designing molecules and macro molecules, tailored to our needs, by building them completely from scratch, one atom at a time? Furthermore, what's keeping us from creating (macro)molecular subunits that self-assemble in a predefined way, as though the subunits are intelligent agents themselves?
M.Sc. / PhD candidate in the field of Medicinal Organic Chemistry with A.I. assisted small-molecular generation
Founder / CTO
M.Sc.; Experienced Full Stack software Developer, A.I. engineer
For updates and/or offers
contact@oscerva-me.com
+49 1704 173 384