# Biological AI and bioinformation services

The evolutionary signal that makes protein language models and structure predictors work — coevolution in multiple sequence alignments — and the failure modes: leakage, orphan families, conformational states, and the gap between predicting and designing.

Why a model trained on sequence alone learns structure, and where that inference stops.

Source: https://en.bioecon.ru/docs/synbio-enabling/bioinformatics-omics/biological-ai-bioinformation-services/
Updated: 2026-09-06



The central empirical fact behind biological machine learning is that evolution has already run the experiment. Residues that touch each other in a folded protein are constrained to change together: a mutation that would break a contact is tolerated only if a compensating mutation appears at the partner position. Across a deep family alignment, this leaves a statistical trace — correlated substitution patterns — from which contacts can be inferred without any physics. Structure prediction became accurate when models learned to extract that signal, and protein language models trained only to predict masked residues recover much the same information because the same constraints govern which sequences are plausible.

## What the signal depends on

Accuracy therefore tracks the depth and diversity of the evolutionary record available for a sequence. Families with many diverse homologues give sharp coevolutionary signals; orphan sequences, rapidly evolving regions, designed proteins and many viral surface proteins do not, and predictions there are correspondingly weaker. Confidence scores reported by predictors are informative about this and should be read as such: low-confidence regions are frequently intrinsically disordered, which is a real biological answer rather than a failure, but they are also where a model has least evidence.

## What a predicted structure is not

A prediction is typically a single, low-energy, apo-like conformation. Many proteins of interest exist as ensembles — transporters that alternate access, receptors with active and inactive states, allosteric enzymes — and a single model does not tell you which state you are looking at, nor how the protein moves between them. Nor does it place ligands, ions, post-translational modifications or membrane environment unless the method was built to. For drug discovery, side-chain positions in a binding pocket often need experimental confirmation before a docking result means anything, which is why structure prediction has increased demand for experimental structural biology rather than replacing it.

## Evaluation is where the field is most easily fooled

Because databases contain overlapping and highly similar sequences, a random train-test split leaks information and inflates performance. Meaningful benchmarks split by sequence identity or by time, holding out structures solved after the training cut-off — this is what makes blind community assessments such as CASP the reference point rather than self-reported figures. The same discipline applies to models predicting variant effects or binding: a model can score well by learning family membership rather than the property in question.

## Prediction and design are different problems

Generating a sequence that a model believes will fold is cheap; the model is judging plausibility under the distribution it learned, not thermodynamic stability, expressibility in a host, solubility, aggregation propensity or immunogenicity. Reported design successes are typically the surviving fraction of many candidates screened experimentally, and the success rate is the informative number. The honest description of the current state is that generative design compresses the search space substantially and does not remove the need to build and test.

Services built on this layer add computation, curated data and pipelines. The scientific content is the same in every case: the models are inference engines over an evolutionary record whose depth varies enormously between targets, and their outputs carry the uncertainty of that record.